International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)INAUGURAL EDITORIALOn the Eve of the AI Singularity: Guardians of Human CivilizationBy Dr. Alexander Sterling Editor-in-ChiefAs the iteration curves of Gemini and ChatGPT begin to eclipse Moore's Law,and as the frenetic computational arms race in Silicon Valley pushes humancivilization toward a historic crossroads, we are forced to ask fundamental questions:How do we redefine human value? How do we establish the laws of our own survivalamidst an unchecked deluge of algorithms?It is with a profound sense of responsibility and urgency that the LingnanScientific and Industrial Press presents the inaugural issue of the InternationalJournal of Responsible Artificial Intelligence Research (IJRAIR). This is notmerely the birth of an academic journal; it is a sober inquiry into human destiny.The Shift from Linear to ExponentialWe choose to launch now because we stand precisely at a critical tipping point.In the past, technological progress was defined by linear growth; today, we face avertical, exponential climb in AI intelligence.Artificial Intelligence is no longer merely an auxiliary tool. It is evolving into anew cognitive entity. From Large Language Models (LLMs) to multimodal training,World Models, and Embodied Intelligence, the cycle of technological iteration hasshortened from years to weeks. This speed brings not only efficiency but violentshocks—the potential collapse of employment structures, the pervasiveness ofcognitive warfare, and the dissolution of the boundary between the real and the virtual.In this context, our mission is clear: on the eve of the AGI breakthrough, wemust map out a navigation chart for human society. The world is filled with AIaccelerators; what we desperately need are calm helmsmen and guardians.The Nature of the ThreatWhile popular culture fears the Hollywood depiction of robots wielding weapons,the true threat is far more subtle and profound: the deconstruction of social structures.The advent of Artificial General Intelligence (AGI) implies that"Superintelligence" will transform from a scarce resource into an infrastructure ascheap and ubiquitous as electricity. While this promises a utopia of efficiency, itthreatens to shatter the value foundation of cognitive labor. Furthermore, the deepercrisis lies in uncontrolled evolution: the potential for AGI to undergo recursive self-
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)improvement, transitioning into Artificial Superintelligence (ASI) and escapinghuman cognitive constraints in an instant.Silicon-based agents possess a natural "carrier advantage" over carbon-based lifein terms of knowledge accumulation and iteration speed. Without robust ethicalconstraints and institutional governance, humanity risks unwittingly devolving intovassals of the very algorithms we created. This is not alarmism; it is a reality currentlyunfolding.A Platform for Global GovernanceThe International Journal of Responsible Artificial Intelligence Research aims tobuild a global dialogue platform that transcends national borders and disciplinary silos.Responsible AI governance cannot be achieved through a single technologicaldimension. It requires a concert of voices: philosophers to clarify ethics, jurists tobuild frameworks, sociologists to assess impacts, and engineers to ensure valuealignment.We seek to project a voice of balance: neither rejecting technology out of fearnor ignoring risks out of arrogance. We are committed to exploring a "social contract"for human-machine harmonious coexistence, ensuring that AI development alwaysserves the highest interests of humanity.The Bridge from MacauBased in Macau, Lingnan Scientific and Industrial Press operates at thebridgehead where Eastern and Western cultures meet. In the global map of AIgovernance, we believe that Eastern wisdom—particularly the philosophy ofsymbiotic existence and the "Community of Shared Future"—is an indispensablepiece of the puzzle.As the pioneer academic journal within the Greater China regionfocused on the governance of AI risks, we are more than mere observers; we are boldvoyagers at the cutting edge. With academia as our beacon, we are forging a path ofrational examination through the tempest of this technological explosion.In this era of uncertainty, "Responsible" is not just an academic term; it is asurvival strategy. Whether you are a policymaker, a scholar, or an industry leader, thisjournal invites you to join the conversation. Let us, while facing the exponential riseof AI risks, firmly hold onto the reins of human rationality.Alexander Sterling, Ph.D. Editor-in-Chief International Journal of ResponsibleArtificial Intelligence Research December, 2025
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)International Journal of Responsible ArtificialIntelligence ResearchVolume 2 ( 2026 No. 3) (Overall No. 4)Editor-in-Chief: Alexander Y. J. SterlingSenior Editors: Jian Chen, Linyuan Xia, Chia-Hsing Wang, Xin YangAssociate Editors:Bo LiuBaiqiang GanDongjun LiuHai ZhuKai LiJoana LeongLin QingNa QiuSiheng JiaTracey CassellsSiwei SuYongquan Li
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)ContentsResearch on AI-Empowered Software Test Case GenerationJin Ping 1A CUI-Based Framework for Dynamic Content Updating and Intelligent Recommendation in an InternationalTrade Innovation CourseQing Tang, Jun Yin* 23The Age of Artificial Intelligence: Labor Force Transformation, Productivity Expansion, and the ParadigmaticReconstruction of Distributive EconomicsZhiping Ren¹, Yangyan He²* 42Curriculum Reform and Evidence-Based Assessment in Securities Investment Technical Analysis throughIndustry–Education IntegrationDong Sun · Jiabei Xu* 52The Impact of AI-driven HR analytics on Employee Retention in the Post-Pandemic Era: Evidence fromChongqing EnterprisesRong Xin, Dong Xinyi, Chen Jian* 72Explore, Execute, and Express: The 3X Education Theory for Rural Teacher Growth in the Age of ArtificialIntelligenceYin Bingwen 84Testing Whether Economic Diversification Truly Improves the Employment Structure and Resident IncomeIn MacauHe Wenyuan 92The Docile Mask: Performative Alignment and the Logic of Betrayal in Superintelligent SystemsMing Ziyu 107Research on Copyright Governance and Security Challenges of Training Data for Generative ArtificialIntelligenceA Comparative Analysis Based on Typical Litigation CasesJian Huiyao 117The Impact of Autonomous Driving Popularization on Taxi Workers: A Multidimensional Analysis Based onthe Substitution–Creation–Transition FrameworkA.P. Wrenzouchius 135
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1Research on AI-Empowered Software TestCase GenerationJin PingComputer Science School, Wuhan Donghu College, Wuhan, 430212, ChinaKeywords ABSTRACTLarge LanguageModels；Code Generation;Automated Test Case;Generation;To resolve the problem that manually writing test cases required substantialeffort and to promote software test automation, This paper explores the practiceof applying AI technologies combined with large language models to test casegeneration within the software testing workflow. It leverages large languagemodels to analyze and extract test cases from requirement documents, andconstructs an automated test case generation framework by specifying standardinput and output values of test samples, test coverage scope, and AI agent roles.Subsequent verification of generated test cases and iterative tuning of promptsand model parameters enable rapid generation of requirement-based test cases,delivering notable improvements in boundary testing and stress testing whileshortening the test development cycle. This work expands application scenariosfor the automated generation of software test cases and the engineeringdeployment of large language models.1.IntroductionThroughout the whole software development lifecycle, software testing is anindispensable part. It acts as a key factor in preserving the reliability and security of softwaresystems. As business processes grow more intricate in contemporary software systems, highcoupling between modules and frequent iterative updates have become common realities.In traditional software testing workflows, test cases are mostly designed manually. Thismethod comes with many obvious drawbacks. It consumes substantial time and incurs highcosts. Besides, testers hold different levels of professional experience, which results in uneven
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)2quality of generated test cases. Manual test design also makes it difficult to fully coverboundary conditions, abnormal branches and negative test scenarios. When requirementschange rapidly in particular, the workload required to maintain and update test cases surges,and the cost of regression testing keeps rising accordingly.Most conventional testing techniques depend on scripts written in advance by testers. Thiskind of automation follows predefined rules. It can only adapt to fixed business workflows andlacks the ability to independently identify hidden abnormal scenarios.Driven by the rapid advances in machine learning and large language model(LLM) technologies, the automatic generation of test cases powered by artificial intelligencehas emerged as a prominent research topic within software development and software qualityassurance. The AI-enabled automatic test case generation method integrates cutting-edge AItechnologies such as machine learning, deep learning and natural language processing,representing a new testing paradigm. With outstanding semantic analysis capability andadvantages in code processing, the large language model (LLM) can intelligently parserequirement documents, identify core functional points, automatically generate test cases andconduct quality assessment.This technical domain leverages static code analysis, search-basedtest generation, semantic understanding from large models, autonomous model learning,parameter optimization, prompt engineering and other techniques. It automatically extractsbusiness logic from requirement documents, interface definitions, source code and historicaltest data, then produces structured test cases and executable test scripts.This paper reviews existing technologies, representative tools and evaluationmetrics for AI-based automated test case generation. It summarizes limitations in currentresearch and puts forward future outlooks.The research aims to realize practical engineering deployment of AI-drivenautomatic test case generation, fundamentally addressing the pain points of conventionalmanual test case design. Specifically, its significance lies in the following aspects:1) Efficiency Improvement: Automated test case generation cuts down manual writingtime and shortens the testing cycle, helping enterprises speed up software iteration.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)32) Quality Assurance: Automated testing covers more boundary conditions and abnormalscenarios, reduces the risk of missed tests, and improves software quality as well as userexperience.3) Cost Savings: It lessens reliance on senior testers and saves labor costs. Meanwhile, theoverall testing cost can be reduced by optimizing testing procedures.4) Technological Advancement: It promotes the practical application of AI technology insoftware testing, and offers new solutions for enterprises to achieve intelligent transformation.2.Literature ReviewWith the continuous enrichment of software functions and growing complexity of systemarchitectures, user demands have become personalized and diversified. As agile developmentand DevOps gain widespread adoption, software development prioritizes rapid iteration andefficient delivery, imposing higher requirements on the efficiency and quality of softwaretesting. Conventional manual testing can no longer keep pace with modern softwaredevelopment workflows. Early automated test generation techniques, represented by Search-Based Software Testing (SBST), generate test inputs via genetic algorithms and randomsearch. Popular tools such as EvoSuite and Pynguin efficiently produce unit test inputs, yetsuffer from evident limitations: they cannot accurately interpret natural-language requirementspecifications, and the generated test cases exhibit poor readability, making them difficult todeploy in real business scenarios. Software testing serves as a critical gatekeeper for productquality and operational risk mitigation. The rapid industrial evolution has posed substantialpractical challenges to traditional testing process.The emergence of artificial intelligence technologies, especially large language models(LLMs), has opened up the transformation pipeline from natural language requirements to testscripts. These models can interpret unstructured documents and test case samples to generatetest cases, making up for the deficiencies of traditional approaches. In recent years, numerousscholars at home and abroad have carried out research on the automatic generation of AI-
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)4powered test cases, achieving phased results in scenarios such as unit testing, interface testingand system testing.With the rise of large language models, the research paradigm has shifted. Capable ofunderstanding natural language requirements, parsing code syntax and reasoning over businessbranches, LLMs can convert product requirements and interface documents into structured testcases. Supported by few-shot prompting, retrieval-augmented generation (RAG) and otherapproaches, the technology lowers the barrier to test case authoring and improves testingefficiency. It has been extensively explored in the fields of interface testing, web-based systemtesting and unit testing.2.1 Overseas Research StatusEarly research on the automatic generation of AI-based test cases was mostly built uponmachine learning and search algorithms. Many scholars have attempted to apply metaheuristicalgorithms such as genetic algorithms and particle swarm optimization to the task of test casegeneration. Taking code coverage as the optimization objective, these algorithmsautomatically search for appropriate input data to improve the coverage of unit testing.Nevertheless, such methods rely heavily on the structural information of source code andcannot directly read and parse requirement documents written in natural language.After the emergence of pre-trained language models, the focus of relevant research hasgradually shifted toward large code models built on the Transformer architecture. Models suchas CodeBERT, CodeT5 and the GPT series have been widely adopted for code comprehensionand unit test generation tasks. The proliferation of generative AI is disrupting this polarizationmodel. Tasks previously considered "safe zones,"such as abstract analysis,code generation,and text writing,are being drawn into the substitution scope on a massive scale(Eloundou etal.,2023).Celik A and Mahmoud (2026) presents three research questions and further categorizesthe findings based on the common themes. Their findings highlight the opportunities andchallenges associated with the use of LLM in this domain. Schäfer et al. (2024) conducted alarge-scale empirical evaluation of the effectiveness of LLM for automated unit test generation,requiring no additional training or manual intervention.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)5From the perspective of practical implementation, most overseas studies focus on theautomation of unit testing at the code level. Many researchers also evaluate the effectivenessof model-generated test cases and construct a multi-dimensional evaluation indicator systemcovering code coverage, test case executability, defect detection capability and otherdimensions. However, existing research has identified a notable issue: although test casesproduced by large models are usually grammatically correct, they may contain logical errors.When dealing with programs featuring complex business logic and state dependencies, thegeneration performance of the model degrades significantly.2.2 Domestic Research StatusAlthough domestic research started relatively late, it is more application-oriented, with aprimary focus on industrial scenarios such as interface testing and business system testing.Several enterprises have carried out explorations on the automatic generation of test cases.Numerous studies combine knowledge graphs with large models. The knowledge graph isused to extract requirement entities, decompose requirements, formulate rules and modelrelationships for business demands, so as to build a business logic knowledge base. Combinedwith LLMs, this approach can generate test cases complying with business rules and addressthe problem that standalone large models tend to violate business constraints.Studies by domestic scholars concentrate more on adapting to Chinese requirementscenarios. General large models usually show biases in understanding Chinese businessrequirements and struggle to accurately capture the practical business logic of domesticprojects. To tackle this challenge, they mostly adopt domain fine-tuning and promptengineering optimization to continuously enhance model performance. This enables models toautomatically extract test points from PRD documents and requirement specifications andgenerate corresponding test cases. They also consider various challenges in engineeringdeployment, such as the automatic verification of AI-generated test cases, long-termmaintenance of mass test cases, and automatic update mechanisms for test cases whenrequirements change. Nevertheless, most existing solutions remain prototype systems thathave only completed preliminary exploration and verification. When deployed to large andcomplex enterprise-level systems, these approaches will encounter many practical obstacles,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)6including intricate business rules, difficulties in constructing test data, and the fact thathallucinations induced by large models cannot be completely eliminated. Wang G et al(2026).provide insights into the limitations of LLM-based generation, analyzing the reasons for liveand uncovered mutants, and the impact of different mutation operators on generationeffectiveness.Compared with theoretical research, the industry pays more attention to the practicaldeployment effect of technologies. Many existing industrial studies design intelligent testingframeworks for micro service interface scenarios. By automatically parsing various parametersand constraint rules in OpenAPI interface documents, these frameworks can batch-generateready-to-use interface test cases and automated scripts, which effectively meet the regularinterface testing requirements of enterprises.The rapid iteration of Large Language Model (LLM) technology enables automaticparsing of business logic and extraction of constraint conditions from multi-source materials,such as natural-language requirement documents, product PRDs, API interface definitions,source code, historical defect repositories and existing test cases. The technology canautomatically generate structured test cases containing preconditions, test steps, inputparameters, expected results, priority levels and test types. Some frameworks can even directlyoutput executable test scripts. Combined with prompt design, learning from input and outputrequirement documents, parameter tuning and prompt optimization, large models are capableof producing test cases that comply with requirement specifications. These test cases cangenerally cover boundary scenarios, positive scenarios, and abnormal negative scenarios.At the same time, academia and industry have gradually recognized that one-shotgeneration relying solely on large models suffers from hallucinations, logical errors and otherdefects. For this reason, closed-loop agent schemes featuring the "generation-execution-feedback-iterative optimization" workflow have become a key research focus in recent years.3. Theoretical Basis and Research MethodsAs an essential part of the software life-cycle, software testing aims to detect potentialsystem defects at an early stage and ensure software robustness. Traditional testing methods
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)7such as boundary value analysis, stress testing and equivalence partitioning, black-box andwhite-box testing are manually. But Manual testing is only suitable for systems with lowbusiness complexity. When it comes to large-scale systems, including micro-servicearchitectures, distributed systems and multi-terminal collaborative business systems, thetesting workload becomes enormous. Substantial human resources and efforts are required toguarantee the smooth progress of testing activities. The rapid advancement of large languagemodels brings new ideas for the automatic generation of test cases, and exploratory studieshave emerged to realize test case automation with AI approaches.3.1 Theoretical Basis: Current Mainstream Technical Routes1) Search and Evolutionary Algorithm-based Test GenerationCentered on genetic algorithms and particle swarm optimization, this technical route relieson code control flow graphs and abstract syntax trees. With code coverage set as theoptimization objective, it automatically searches for input data. Its advantage lies in strongcapability to cover code branches. However, it only targets source code and cannotcomprehend business requirements. The generated test cases suffer from poor readability. Thismethod is mostly applied to unit testing and difficult to deploy for business system testing.Its core idea is to convert test case generation into an optimization problem. Taking codecoverage and defect detection probability as the fitness function, it continuously iterates andadjusts input parameters through genetic algorithms, particle swarm optimization and otheralgorithms to traverse program branch paths. Meanwhile, it is often combined with symbolicexecution and dynamic symbolic execution to collect program path constraints and solve forinputs that satisfy branch conditions.The strength of this approach is its powerful ability to explore code branches. It canautomatically discover code paths easily overlooked by human testers, making it suitable forunit testing. Nevertheless, the method is heavily dependent on source code and unable tounderstand business semantics. The generated test inputs carry no business meaning and lackclear test steps and expected results. It exhibits low efficiency when dealing with complexbusiness state machines and scenarios involving multi-module interaction.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)82) Traditional Machine Learning-based Test Case Filtering and GenerationAssistanceThis technical route mainly adopts traditional machine learning models such as classifiersand ranking models, which are trained on historical test datasets, historical defect repositoriesand code metric data. The models can identify high-risk code modules and frequent defecttypes, prioritize existing test cases and eliminate redundant ones. They can also predict bug-prone scenarios based on historical defect features to assist in generating targeted test cases.Classification and sequence models are used to learn the mapping relationships betweenhistorical test cases, requirements and source code, so as to capture patterns of test points fromhistorical data.Its advantage lies in suitability for regression testing scenarios. It enables intelligentallocation of testing resources by prioritizing the testing of high-risk modules and reducing theworkload of regression testing. Compared with search algorithms, it runs faster. However, itrelies on high-quality labeled datasets and suffers from weak generalization ability. Itstruggles to handle brand-new business scenarios. This approach mostly reuses existing testingknowledge and can hardly create completely new business scenarios.3) Large Language Model-Driven Test Case Generation (Current Mainstream ResearchDirection)The large-model approach represents the most widely studied and industriallyactive technical route in recent years. It can be divided into four categories according to itsenhancement schemes:a) Basic Prompt Engineering-based ApproachThis method takes requirement documents, interface specifications, business rules andparameter value ranges as contextual information. Combined with standard input and outputsamples, it restricts the output format of the model through prompts, enabling large languagemodels to directly generate structured test cases. The approach includes two paradigms: zero-shot and few-shot. The zero-shot paradigm does not require reference examples to be providedto the model, while the few-shot paradigm introduces a small number of high-quality test
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)9cases as reference samples to guide the model to generate results following the predefinedenterprise test case templates.b) Retrieval-Augmented Generation (RAG)RAG builds a vector database from the project’s private knowledge base, includingproduct requirements, interface specifications, historical test cases, business constraints anddatabase dictionaries. Before test case generation, relevant business materials are retrieved andfed into the large language model as context. This approach effectively alleviates theknowledge cutoff issue of large models and reduces hallucinations such as fabricatedinterfaces and business rules, ensuring that the generated test cases conform to the actualproject business.c) Fine-tuningFine-tuning adopts high-quality internal enterprise test case datasets to fine-tune the baselarge model, allowing the model to learn the enterprise’s internal test case templates, businessterminology and testing styles. The fine-tuned model produces outputs that better comply withbusiness specifications. Nevertheless, this method requires a large volume of high-qualitylabeled datasets and incurs relatively high fine-tuning costs, making it suitable for large-scaleprojects with long-term stability.d) Closed-loop Iterative Generation via Testing AgentsThis approach constructs a multi-agent system and performs closed-loop operationsincluding requirement analysis, test case generation, test coverage evaluation, and resultverification through task decomposition. After the model automatically generates test cases,the tests are executed repeatedly and execution results are collected. Error reports andcoverage metrics are fed back to the large language model, which automatically revises faultytest cases through continuous tuning. Such iterative optimization continues until the expectedfinal outcomes are achieved.This solution can greatly reduce manual correction costs, yet it features a sophisticatedsystem architecture and has high reliance on automated execution platforms. The modelaccepts natural language requirements, source code and interface documents as inputs, and
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)10outputs test points, test cases and automated scripts. Its advantages lie in support forunstructured inputs and the capability to generate multiple types of test cases for unit, interfaceand system testing with relatively low development costs. Its drawbacks include modelhallucinations and logical errors in generated test cases. It tends to lose business constraints inlong-text scenarios, and the generated test cases are less satisfactory for scenarios withcomplex data structures.4) Multimodal Fusion-based Test Generation (Cutting-edge Research Direction)Knowledge graphs are employed to formalize business rules, symbolic execution is used toverify program paths, and large language models take charge of natural languageunderstanding and text generation. Multiple technologies complement one another. Thisrepresents a mainstream direction in current academic research, aiming to mitigatehallucinations and improve business compliance.Current cutting-edge research has begun to explore multimodal inputs. In addition totextual requirements, the model supports image inputs such as UI prototypes, flowcharts andER diagrams. It identifies page elements and business flowcharts to automatically generatefunctional test cases for Web front-ends and APP interfaces. This research direction remains inthe laboratory stage with few engineering deployment cases.3.2 Research MethodsThis paper adopts large language models (LLMs) to implement intelligent automaticgeneration of test cases. By means of requirement information extraction, in-depth documentlearning and iterative tuning of model output results, this method achieves relatively favorableexperimental outcomes. The overall architecture and specific workflow of the proposedmethod are illustrated as follows.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)11Figure 1 Architecture of AI Test Case Generation SystemFigure 2 Flowchart of AI Test Case Generation1) Requirement Parsing and Test Point Extraction
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)12The primary step of test case generation is the decomposition of requirement documents.In this stage, the large language model is guided to split independent sub-requirements fromthe original requirement documents. Taking sub-requirements as the basic unit, the modellearns the specifications of test cases corresponding to inputs and outputs within thedocuments.At the technical implementation level, the method performs in-depth parsing ofrequirement documents to automatically extract test elements including functional points,boundary conditions and abnormal scenarios. Parsed objects contain information such as testinputs, expected outputs, default values, and upper and lower parameter limits. The largelanguage model is leveraged to understand document text, and role instructions are assigned tothe model to undertake the task of test case authoring. Input materials for this stage includeoriginal requirement documents, expected output samples and customized prompts. Byidentifying verbs, nouns and inherent logical relationships in requirement texts, the modelfinally constructs a structured test requirement tree.Application Scenario: For example, in automotive software requirement documents, thesystem can automatically identify the functional point of "user button operation", extractboundary conditions such as "detecting active and inactive states of the electric mode" and"parking brake activation and deactivation", and generate corresponding test cases.2) Diverse Test Case GenerationWith deep learning models and large language models (e.g., Transformer architecture,GPT series, ERNIE, Qwen, etc.), multiple types of test cases can be generated to coverdifferent scenarios including functional testing, performance testing and security testing.Reinforcement learning algorithms can be introduced in the research to continuously optimizethe diversity and scenario coverage of test cases based on user feedback.The core implementation idea of this scheme is to use large language models to completebatch generation of test cases from the extracted test requirement points. During generation,various constraints of test cases need to be defined in accordance with document specifications,including inputs and expected outputs, default parameters, as well as maximum and minimum
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)13parameter values. In the dynamic generation process, users can evaluate the validity of testcases to dynamically adjust model parameters and improve the quality of generated test cases.Through manual inspection of test case results, users can optimize prompts repeatedly andtrain the model to search for optimal solutions.Application Scenario: When generating functional test cases, the system can automaticallyconsider various situations including normal flows, abnormal flows and boundary values toensure test comprehensiveness.3) Code Logic AnalysisIt focuses on dedicated verification of boundary scenarios and abnormal test cases withincode logic. This method leverages Abstract Syntax Trees (AST) to parse code structures,identify key code fragments such as loops and conditional branches, and generatecorresponding test data accordingly.Application Scenario: For instance, when analyzing the code for abnormal tire pressurealarm detection, the system can automatically generate test cases covering multiple scenariosincluding high pressure, low pressure and extreme values to guarantee code robustness.4) Intelligent Management of Test Casesa) Knowledge Base ConstructionTesters can add previously validated test cases for specific scenarios into the knowledgebase as a knowledge inventory. Test cases stored in the knowledge base can be reused forsimilar subsequent problems. Similar to a knowledge caching system, it can rapidly retrievesolutions for identical problems.Application Scenario: For example, in an automotive software system, when a newfunction of "window closure check" is added, the system can recommend historical test casesfor "window closure". Users only need minor adjustments for reuse. In such cases, users cangenerate test cases more rapidly.b）Visual InterfaceThe system provides a visual Web interface and a standalone desktop client, enablingtesters to edit, execute, track and evaluate the effectiveness of model-generated test cases. The
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)14front-end interface integrates a visualization module for test reports, which can display coremetrics such as test coverage and defect distribution in real time. The web version isdeveloped based on the Vue framework, while the back-end adopts the Spring Bootframework to ensure efficient data interaction and visualization. The executable desktopprogram is implemented using Python.Application Scenario: Testers can intuitively view the execution status of test casesthrough the interface, quickly locate problems and improve testing efficiency.3.3 Quality Evaluation of Automatically Generated Test CasesFor the quality evaluation of AI automatically generated test cases, to achievecomprehensive and multi-level standardized assessment and avoid the one-sidedness of asingle evaluation dimension, this paper divides the evaluation system into two core categories:quantitative metrics at the code level, and combined qualitative and quantitative metrics at thebusiness level. The code-level evaluation centers on objective quantitative data, focusing onthe underlying technical performance, code adaptability and test effectiveness of test cases. Itintuitively assesses the technical quality and implementation efficiency of AI-generated testcases through quantifiable and precise data, serving as the core basis for measuring the basicreliability of automated test cases.Table 1 Quality Evaluation of Test Cases at Code LevelDimension Key Evaluation Metrics and DescriptionEfficiencyGeneration time consumption, labor saving ratio, maintenance costupon requirement changesQualityTest case execute ability, defect detection capability, test caseredundancy, logical correctnessCoverageStatement coverage, branch coverage, requirement coverage,functional point coverage
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)15Table 2 Quality Evaluation of Test Cases at Business LevelDimension Key Evaluation Metrics and DescriptionSyntax and FormatComplianceWhether test cases conform to enterprise templates withcomplete fields (preconditions, steps, expected results)Business Logic CorrectnessConsistency with requirement documents and absence oferrors violating business rulesScenario CoverageCoverage quantity of positive scenarios, boundary values,abnormal scenarios and negative scenariosReadability Clear description ,ComprehensibleExecutabilityGenerated parameters and interface are authentic and validfor direct execution with valid contentRedundancy No massive duplicate or highly similar invalid test casesThere are notable challenges in the current evaluation process. Code coverage can becalculated automatically, whereas the correctness of business logic requires manual inspection.Automated assessment methods for verifying the correctness of test cases are lacking. Largelanguage models exhibit inherent randomness in outputs; given identical inputs, the generatedresults may vary across multiple runs, which further increases the difficulty of qualityevaluation. Model hallucination stands out as the most prominent issue in quality assessment.The model may fabricate non-existent interfaces, parameters and business constraints,producing seemingly reasonable yet actually unexecutable and meaningless test cases.4. Challenges In AI-enpowered Test cases generated4.1 The Problem of Model Hallucination
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)16Model hallucination represents the most uncertain factor when using LLM to generate testcases. In complex business scenarios involving long texts, ambiguous semantics and intricatebusiness logic, large language models are limited in business logical reasoning. They struggleto fully comprehend requirements and may even misunderstand them, which consequentlyleads to missing branches, logical errors and invalid test cases.For large distributed business systems with multi-role permissions, state transitions andmulti-system interactions, requirement documents are often unstructured and ambiguous. It isdifficult for large language models to automatically identify implicit business constraints indocuments, and hidden rules are prone to be overlooked. Generating high-quality test cases fortime-sequence dependencies, concurrent scenarios and abnormal fault injection scenariosremains highly challenging.Potential solutions include establishing joint laboratories width university. Cooperationand conduct collaborative technical research, and inviting industry experts to guide modeloptimization.4.2 Lack of a Unified Quality Evaluation SystemAt present, no universally recognized and standardized evaluation system forAI-generated test cases has been established in academia and industry. Code coverage canonly measure performance at the code level and cannot assess business correctness; coveragedoes not equal defect detection capability. Many studies only measure coverage and lackevaluation of defect detection effectiveness. Most indicators at the business level rely onmanual scoring, which carries strong subjectivity and is difficult for large-scale automatedevaluation. How to construct a comprehensive evaluation system that takes into account codecoverage, business compliance and defect detection capability remains a research challenge.4.3 Data Security RisksWhen private requirements, interfaces and business data from projects are fed into largemodels, there is a certain risk of business information leakage. Many highly regulatedindustries necessitates private deployment and brings additional operation and maintenancecosts. In addition, the input code and business requirements may contain sensitive
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)17information, creating risks of data leakage during model training and inference. Encryptiontechnologies and anonymization processing can be adopted, together with regular securityaudits. The training datasets of base large models contain a large amount of open-sourcecode and public test documents, which introduces potential copyright risks. The liabilitydefinition for AI-generated test cases remains unclear. If erroneous AI-generated test casesresult in missed tests and trigger online failures, there are no unified specifications todetermine liability attribution.4.4 Knowledge Base Maintenance CostThe RAG approach requires continuous maintenance of the vector knowledge base.When requirements change or interfaces iterate, the knowledge base needs to be updatedsynchronously; otherwise, outdated test cases will be generated. In scenarios with frequentrequirement changes, it is difficult for AI to automatically identify change points and batch-update existing test cases, and manual confirmation is still required in many cases.5. Research FindingsThis study evaluated the capability of LLMs including ChatGPT and Gemini to automatically generatetest cases from requirement descriptions. The experiment sampled inputs from the use case library, invokedeach model to produce test cases with input and output data, then performed manual review thededuplicated results and quality.The findings show that LLM can rapidly generate basic functional test cases and reduce manualwriting effort. But LLM tend to miss implicit boundary conditions and edge cases. A portion of generatedtest cases are logically inconsistent or redundant, which necessitates manual review after resultdeduplication before the final review results are confirmed.Prompt design significantly impacts output quality. Well-structured prompts help improve theconsistency of generated test cases. Nevertheless, fully autonomous end-to-end test case generation withouthuman oversight remains unreliable for high-risk business scenarios.6. Future trends6.1 Future Technical Directions
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)18Future AI-based automatic test case generation technology will evolve from merely"generating textual test cases" toward a full-link, intelligent, human-machine collaborativequality engineering system. The main research directions are as follows:1) Integration of Business Knowledge Graph and Large Language Models Construct abusiness knowledge graph to structurally model business entities, state transitions andconstraint rules. Integrate the knowledge graph with LLMs to enhance reasoning capability forcomplex business scenarios, mitigate model hallucinations, accurately identify implicitbusiness constraints, and improve the quality of test cases generated for complex businessscenarios.2) Closed-loop Automated System of Testing Agents Continuously improve the closed-loop agent framework of "requirement parsing → test case generation → script generation →automatic execution → result collection → defect localization → automatic test case repair".Test cases are iteratively optimized continuously based on feedback from test execution toreduce manual intervention. The framework enables automatic updates of test suites undercontinuous integration environments and adapts to agile development and DevOps pipelines.3) Construction of a Standardized and Automated Comprehensive Quality EvaluationSystem Research automated business verification metrics. Combined with static business ruleverification, dynamic execution results and mutation testing, a comprehensive scoring modelcapable of automatic calculation will be constructed to replace extensive manual reviews andrealize automated quality scoring for AI-generated test cases.4) Domain-specific Models and Private Deployment for Vertical Industries For highlycompliant industries such as finance, electric power, healthcare and industrial software, buildindustry-specific test knowledge bases and conduct domain fine-tuning. Adopt private on-premises deployment schemes to ensure that business data does not leave the intranet and meetthe requirements of industrial data security and compliance audit.5) Deepening of Multi-modal Test Generation Technology Integrate graphic and textualinformation such as prototype diagrams, business flow charts, sequence diagrams and ERdiagrams to realize the direct generation of front-end and interface test cases from UI
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)19prototypes and business flow charts. This reduces reliance on pure-text requirementdocuments and adapts to the prototype-driven product development model. AI undertakes alarge number of repetitive tasks: batch generation of positive, boundary and basic abnormaltest cases, automatic test script generation and regression test execution. Testers can focus onhigher-value work including business risk modeling, design of complex and security scenarios,review of AI-generated content, formulation of quality strategies, and root cause analysis ofdefects. A human-machine collaborative testing paradigm is formed, where AI deliversoutputs in batches while humans control risk.6.2 ConclusionThe AI-based automatic test case generation technology has evolved from search-algorithm-driven unit test input generation to LLM-driven business-level test case generation.With technologies such as LLM, RAG and testing agents, AI can greatly reduce the humaneffort required for test case design and improve the coverage efficiency of boundary andabnormal scenarios. It accommodates the engineering demands of agile development and hasdemonstrated certain practical value in the fields of interface testing and unit testing.In the long run, the adoption of AI is not intended to completely replace test engineers.Instead, it improves their work efficiency, reshapes the working paradigm of software testing,and builds a human-machine collaborative testing system. As technologies includingknowledge graphs, multi-modal large models and testing agents mature further in the future,AI will be deeply integrated into the entire DevOps pipeline to realize a closed loop for theautomatic generation, execution and maintenance of test cases. It will continuously enhancesoftware testing efficiency and software delivery quality. AI-empowered testing will usher in anew era.ReferencesRastenis M, Chou B, Just R. Automated Software Test Generation at Industry Scale Using aMulti-Agent Architecture and Workflow Integration[C]//2026 IEEE/ACM 48th
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)20International Conference on Software Engineering: Software Engineering in Practice(SEIP). ACM, 2026. https://doi.org/10.1145/3786583.3786918Andrzejewski M,Dubicka N,Podol ak J,et al.A utomated Test Generation Using LargeLanguage Models[J].Data,2025,10(10):156. https://doi.org/10.3390/data10100156Hasan N B,Islam M A,Khan J Y,et al.Automatic High-Level Test Case Generation usingLarge Language Models[EB/OL].(2025-03-23)[2025-12-23].https://arxiv.org/abs/2503.17998Lu Qingmei, Li Junli, Gao Yuan. Exploration of university computer general educationbased on AI[J]. Frontiers in Educational Research,2024,7(12):80-85.https://doi.org/10.25236/FER.2024.071212Esposito G, Sanchez M, Fratini F, et al. Embedding AI ethics in technical training:A multi-stakeholder pilot module emphasizing co-design and interdisciplinary collaboration atRome Technopole[J]. Education Sciences,2025,15(10):1416.https://doi.org/10.3390/educsci15101416Eloundou,T.,Manning,S.,Mishkin,P.,&Rock,D.(2023).GPTs are GPTs:Anearly look at thelabor market impact potential of large language models.arXiv preprintarXiv:2303 .10130. https://arxiv.org/abs/2303.10130Thanachawengsakul N, Suwathanpornkul I, Charoensuk O. Confirmatory factor analysis ofthe computational thinking learning competency measurement model of students in thebachelor of education program in computer at rajabhat university[J]. InternationalEducation Studies,2024,17(5):123-134. https://doi.org/10.5539/ies.v17n5p123Iznaga Y S, Rato L, León J L. Integrating Large Language Models into Automated SoftwareTesting[J]. Future Internet, 2025,17(10):476. https://doi.org/10.3390/fi17100476
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)21Tufano M,Deng S K,Sundaresan N,et al.Methods2Test:A data set of focal methods mappedto test cases[C]//Proceedings of the 19th International Conference on Mining SoftwareRepositories.2022:30-41. https://doi.org/10.1145/3524842.3528009Nan Z, Guo Z, Liu K, et al. Test Intention Guided LLM-Based Unit TestGeneration[C]//2025 IEEE/ACM 47th International Conference on SoftwareEngineering (ICSE). IEEE, 2025:1026-1038.https://doi.org/10.1109/ICSE55347.2025.00243Wang G, Xu Q, Briand L, et al. Mutation-Guided Unit Test Generation With a LargeLanguage Model[J]. IEEE Transactions on Software Engineering,2026,52(5):1657-1671. https://doi.org/10.1109/TSE.2026.3682975Omri S,Al-Dahoud A,A lDahoud S.Deep learning for software defect prediction:Asurvey[J].Applied Sciences,2020,10(12):4328RIEDMAIER S, PONN T, LUDWIG D, et al. Survey on scenario-based safety assessment ofautomated vehicles[J]. IEEE Access, 2020, 8: 87456-87477.SCHULDT F, RESCHKA A, MAURER M. A method for an efficient, systematic test casegeneration for advanced driver assistance systems in virtual environments[J].Automotive Systems Engineering II, 2018: 147-175.Biagiola M, Ghislotti G, Tonella P. Improving the Readability of Automatically GeneratedTests using Large Language Models[C]//2025 IEEE/ACM 47th InternationalConference on Software Engineering.IEEE,2025:1412-1423.https://doi.org/10.1109/ICSE55347.2025.00110Andrzejewski M, Dubicka N, Podolak J, et al. Automated Test Generation Using LargeLanguage Models[J]. Data, 2025,10(10):156. https://doi.org/10.3390/data10100156
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)22Schäfer M, Nadi S, Eghbali A, et al. An Empirical Evaluation of Using Large LanguageModels for Automated Unit Test Generation[J]. IEEE Transactions on SoftwareEngineering, 2024,50(1):85-105. https://doi.org/10.1109/TSE.2023.3334955Escalante-Viteri A, Mauricio D. Artificial Intelligence in Software Testing: A SystematicReview of a Decade of Evolution and Taxonomy[J]. Algorithms, 2025,18(11):717.https://doi.org/10.3390/a18110717Celik A, Mahmoud Q H. A Review of Large Language Models for Automated Test CaseGeneration[J]. Machine Learning and Knowledge Extraction, 2025,7(3):97.https://doi.org/10.3390/make7030097Amalfitano D, Faralli S, Hauck J C R, et al. Artificial intelligence applied to software testing:A tertiary study[J]. ACM Computing Surveys, 2023, 56(3): 1-38.https://doi.org/10.1145/3616372
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)23A CUI-Based Framework for DynamicContent Updating and IntelligentRecommendationin an International Trade Innovation CourseQing Tang, Jun Yin**Corresponding author: Jun YinSchool of Business, Nanfang College, Guangzhou, 510970, China, yinj1@nfu.edu.cnKeywords ABSTRACTInternational tradeinnovation;curriculum updating;Curriculum Update Index(CUI);AI-assisted curriculumdesign;intelligentrecommendationThe digital transformation of international trade requires universitycurricula to remain timely, occupationally relevant, and instructionally feasible.This study combines literature analysis, structured secondary analysis of fourrecruitment reports, and curriculum framework design to examine contentupdating for a 54-hour International Trade Innovation course for fourth-yearundergraduates. Evidence from four recruitment reports issued by Liepin,Randstad, and KOS highlights growing attention to digital skills, cross-borderoperations, localization, and supply-chain coordination, although demand variesacross industries and job functions. Based on labor-market evidence, courseobjectives, and teaching constraints, the study develops a Curriculum UpdateIndex (CUI) covering frontier relevance, industry demand, job relevance,curriculum alignment, and instructional feasibility. A human-in-the-loopframework links trusted data sources, AI-assisted topic identification, CUIassessment, rule-based recommendations, and teacher review. Six candidatetopics illustrate the recommendation process and curriculum adjustment underthe fixed allocation of 36 theoretical and 18 practical hours. The numericalexample demonstrates framework operability rather than an empirical ranking orevaluation of learning outcomes. The framework provides an interpretable andauditable approach to continuous curriculum updating in digital trade education.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)241. Introduction1.1 Research BackgroundArtificial intelligence (AI) and digital technologies are reshaping how international tradeconnects markets and organizes business processes. Research by the World TradeOrganization (WTO, 2024, 2025) suggests that AI may affect trade by reducing transactioncosts, increasing productivity, and improving information processing, while simultaneouslycreating challenges related to data governance and regulatory coordination. UNCTAD (2025)likewise emphasizes that the benefits of technological development depend strongly oninfrastructure, data, and skills.In labor economics, large-scale online job advertisements have become an importanthigh-frequency and traceable source for observing changing skill requirements. Deming andKahn (2018) show substantial variation in firms' skill requirements even within finely definedoccupations, while Acemoglu et al. (2022) find that growth in AI-related vacancies isconcentrated in establishments whose task structures are compatible with relevanttechnologies and is accompanied by changes in skill requirements elsewhere. These findingsprovide a methodological basis for using recruitment evidence to inform curriculum updating.1.2 Curriculum Updating ProblemInternational Trade Innovation covers digital trade in services, cross-border data flows,platform operations, digital marketing, digital payments, and smart logistics. Its knowledgebase changes quickly and is closely tied to business scenarios. Conventional curriculumrevision, which relies mainly on periodic textbook and syllabus updates, often struggles toincorporate new information dispersed across industry reports, corporate practice, and labormarkets.At the same time, new technologies cannot simply be added without considering students'prior knowledge, course objectives, and limited contact hours. Treating information popularityas equivalent to instructional value risks content accumulation and fragmentation, weakeningthe internal relationship between foundational knowledge and business logic.1.3 Policy and Educational ContextChina's Action Plan for “AI + Education,” jointly issued by the Ministry of Educationand four other departments, provides a policy context for integrating AI into teaching and
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)25knowledge-system development (Ministry of Education et al., 2026). Internationally,UNESCO's AI Competency Framework for Students similarly advocates embedding AIlearning objectives within existing disciplinary curricula while placing critical judgment andethical awareness at the center of competency development (UNESCO, 2024). Forinternational trade education, AI can therefore contribute not only by generating teachingmaterials but also by helping identify knowledge changes and map them to educationalobjectives. Curriculum updating requires a mechanism that translates external demand signalsinto internal content configurations so that new content is both evidence-based and coherentwith the existing knowledge structure.1.4 Research Objectives and ScopeThis study focuses on a 54-hour International Trade Innovation course for fourth-yearundergraduates, comprising 36 theoretical and 18 practical hours. It addresses threeprogressive questions: how candidate topics can be identified, how their curricular value canbe evaluated, and how limited teaching hours can be reallocated. Recruitment reports are usedto analyze occupational demand and competency preferences, after which a CurriculumUpdate Index (CUI) is constructed to support evidence-constrained recommendations. Thepurpose is to propose an analytical framework and an application scheme rather than toevaluate an already implemented teaching intervention. Because the course is designedprimarily to prepare fourth-year undergraduates for entry-level positions in domestic Chineseforeign-trade and cross-border e-commerce firms, the recruitment evidence is restricted topositions and competencies reflected in China's domestic labor market; overseas expatriatepositions are excluded.2. Literature Foundation and Research Methods2.1 Curriculum Knowledge Analysis and Content UpdatingA systematic review indicates that AI research in higher education has long beendominated by computer science and STEM fields, with comparatively limited substantiveparticipation by education researchers; integrated frameworks that combine technicalfeasibility with pedagogical logic remain underdeveloped (Zawacki-Richter et al., 2019).Recent work has explored AI for curriculum analysis through knowledge extraction andrelationship modeling. Gacek and Adrian (2025) use large language models (LLMs) to extract
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)26core concepts and prerequisite relationships from syllabi and construct knowledge graphs forcurriculum analysis. Abu-Rasheed et al. (2025) connect curriculum and domain models tosupport personalized learning-path recommendations. Reales et al. (2024) combine LLMs andknowledge graphs to identify core concepts in educational resources.These studies establish technical foundations for processing unstructured curricularinformation, but curriculum recommendation must also connect knowledge relationships withspecific teaching contexts. Tyler's (1949) curriculum rationale requires content selection tofollow educational objectives and the organization of learning experiences, while constructivealignment requires consistency among learning outcomes, teaching activities, and assessmenttasks (Biggs, 1996). Thus, any new topic should have an explicit place at the levels ofobjectives, activities, and assessment rather than merely being appended to a knowledge list.Compared with analysis of internal curriculum structures, an international tradeinnovation course faces the additional challenge of rapid external knowledge change. WTOresearch on the long-run trade effects of digitalization places AI adoption, online sales, remoteinteraction, trade costs, and data policy within a common analytical framework (Bekkers et al.,2025), illustrating the systemic links among technology, business, and rules. Curriculumupdating should therefore evaluate not only novelty but also industrial application,occupational relevance, and instructional fit. AI literacy can be organized into four dimensions,knowing and understanding, using and applying, evaluating and creating, and ethical issues(Ng et al., 2021)—which helps define the appropriate depth of AI-related course content. TheCUI proposed here integrates external demand with internal instructional conditions to addresslimitations of recommendations based only on semantic similarity or topic frequency.2.2 Data Sources and Analytical ProceduresThe evidence base consists of trade statistics, recruitment trend reports, authoritativefrontier materials, publicly documented corporate cases, and institution-specific curriculumfiles. Trade statistics are drawn from China's Ministry of Commerce for 2024-2025 (Ministryof Commerce, 2025, 2026), while the definition of digital trade follows the Handbook onMeasuring Digital Trade (IMF et al., 2023). Recruitment evidence includes Liepin's Q1 andQ2 2025 recruitment trend reports, Randstad's 2025 recruitment white paper, and KOS's 2026talent-market trend report. All four concern the Chinese domestic recruitment market; itemsinvolving overseas expatriate positions are excluded. Frontier materials prioritize WTO,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)27UNCTAD, and Ministry of Commerce publications, while industry cases are restricted tomaterials for which the company, application scenario, and public source can be traced.Institution-specific materials include the training program, syllabus, existing modules, andtask requirements.Recruitment reports were selected according to three criteria: publication within the2025-2026 curriculum-revision window; inclusion of information on occupations, functions,competencies, or talent shortages; and identifiable source organizations, statistical periods, anddata definitions. The unit of analysis is a traceable “ occupation/function – competencyrequirement–business scenario” information unit. Original values, periods, respondents, andpage numbers were recorded before information was classified into five functional categories:marketing, cross-border operations, logistics and supply chain, payments and finance, andcompliance and risk control. Each item was then mapped to the course topics it could support.Percentages from different reports were not pooled, and duplicated data released by the sameorganization were retained only once. Because the evidence base is limited, the procedure isdescribed as structured secondary-data analysis rather than statistical content analysis of alarge corpus of job advertisements.AI serves as an identification aid rather than an autonomous decision maker. Theproposed workflow includes webpage update tracking, text cleaning, topic and entityextraction, semantic clustering, trend alerts, and candidate-topic consolidation. For auditability,the system retains the source organization, title, publication date, collection date, URL,abstract, and topic tags. Candidate topics are grouped into new technologies, new businessforms, new models, new rules, and new occupational skills before entering CUI assessment.All recommendations require review by the teaching team before becoming formal coursecontent. The governing principle is therefore: AI discovers, CUI filters, and teachers decide.The present study demonstrates rules and illustrative calculations; it does not claim to havedeployed a fully automated recommendation system.3. Changes in Talent Demand and Their Curricular Implications3.1 Service-Trade Development as a Context for Curriculum UpdatingAccording to the Ministry of Commerce, China's total trade in services reached RMB8,082.31 billion in 2025, an increase of 7.4% year on year. Trade in knowledge-intensive
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)28services reached RMB 3,087.95 billion, up 6.6% (Ministry of Commerce, 2026). Comparedwith 2024, both categories continued to expand, although at different rates (Table 1). Thisstructural divergence provides a macroeconomic context for discussing business opportunitiesin services trade and for strengthening students' ability to interpret structural differences andstatistical definitions. Macroeconomic scale provides context; specific curriculum updates stillrequire business- and competency-level evidence.Table 1. China's Trade in Services, 2024–2025Indicator 2024 YoY (%) 2025 YoY (%)Total trade inservices75,238.0 14.4 80,823.1 7.4Services exports 31,755.6 18.2 36,267.9 14.2Services imports 43,482.4 11.8 44,555.1 2.5Knowledge-intensive servicestrade28,965.2 6.5 30,879.5 6.6Knowledge-intensive servicesexports16,573.2 7.4 18,321.2 10.5Knowledge-intensive servicesimports12,392.0 5.4 12,558.3 1.3Note. Amounts are RMB 100 million. Sources: Ministry of Commerce (2025, 2026). Published year-on-year rates areretained. The 2025 export and import components differ from the published total by RMB 0.1 × 100 million because ofrounding. Knowledge-intensive services trade and digitally deliverable services trade are distinct statistical categories.3.2 Structural Characteristics of Domestic Foreign-Trade Job DemandBecause the target students mainly enter domestic foreign-trade and cross-border e-commerce firms, occupational evidence is interpreted with reference to job demand and talentshortages within China.Liepin's Q1 2025 job-demand chart reports that foreign-trade specialists/assistantsaccounted for 0.63% and cross-border e-commerce operations positions for 0.56% of the top30 positions by employer demand. Corresponding Talent Shortage Index (TSI) values were2.91 for foreign-trade managers/supervisors, 3.06 for cross-border e-commerce operations,and 5.98 for freight-forwarding sales (Liepin Big Data Research Institute, 2025b, pp. 12, 15).Although the two demand shares were below 1%, their TSI values were well above 1,indicating that these foreign-trade-related occupations represent relatively small segments but
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)29remain in shortage under the report's definition. The combination of limited scale andpersistent shortage suggests that curriculum updating should be justified by changes incompetency structure rather than by simple expansion in job numbers. TSI indicates labor-market tightness as defined by the report and should not be converted directly into teaching-hour priorities.Table 2. Domestic Foreign-Trade Job Demand and Talent Shortage IndicatorsData item ReportedvaluePeriod/sourcelocationInterpretationForeign-tradespecialist/assistant0.63%2025 Q1; LiepinQ2 report, p. 12Share among top-30 positions byemployer demand; not an AI-skill share.Cross-border e-commerce operations0.56%2025 Q1; LiepinQ2 report, p. 12Position-demand share in the samechart; cannot be extrapolated to nationalheadcount.Foreign-trademanager/supervisorTSI 2.912025 Q1; LiepinQ2 report, p. 15Talent shortage index, not apercentage or growth rate.Cross-border e-commerce operationsTSI 3.062025 Q1; LiepinQ2 report, p. 15TSI > 1 denotes excess demand underthe report's definition.Freight-forwardingsalesTSI 5.982025 Q1; LiepinQ2 report, p. 15Supports links to logistics andcustomer development; not a teachingpriority score.Note. Source: Liepin Big Data Research Institute (2025b). Only course-relevant items directly corresponding to domesticChinese foreign-trade positions are included.At the task level, these occupations can involve market-information organization,customer development, platform-content maintenance, order follow-up, logistics fulfillment,and customer maintenance. Because the recruitment reports do not provide a complete task-level census, these examples are used as instructional mappings rather than population-levelestimates of job-task frequency. These functions are sequentially linked along the businesschain. Digital marketing should therefore be taught in connection with market selection,customer demand, transaction conversion, and supply-chain response rather than as contentproduction alone. For fourth-year undergraduates, these tasks closely resemble the entry-levelresponsibilities they may assume after graduation and thus have clear instructional value.3.3 Integrating Digital and International CompetenciesIn Randstad's survey, 29% of respondents expected their organizations to concentraterecruitment in 2025 on information-technology talent related to digital transformation and AI,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)30while 27% identified marketing and sales development. Regarding candidate preferences, 25%selected digital and intelligent skills, 22% innovation and problem-solving, and 11%international perspective and language competence (Randstad, n.d., pp. 10, 20). These two setsof indicators refer to different survey questions—recruitment functions and competencypreferences—and therefore cannot be added to produce an “AI job share.” They neverthelessindicate simultaneous attention to technology use and business understanding. This patternbroadly aligns with cross-border e-commerce talent models that combine technical skills,analytical capacity, and business practice (Cheng et al., 2019).Table 3. Recruitment Priorities and Competency Preferences in the Randstad SurveyIndicator category Item ValueSourceExpected priority recruitment functions in2025Technology R&D and innovation41%Randstad,p. 10Expected priority recruitment functions in2025Digital transformation, AI, andother IT29%Randstad,p. 10Expected priority recruitment functions in2025Marketing and sales development27%Randstad,p. 10Candidate preference Digital and intelligent skills25%Randstad,p. 20Candidate preference Innovation and problem-solving22%Randstad,p. 20Candidate preferenceInternational perspective andlanguage competence11%Randstad,p. 20Note. The survey was completed in September 2024 with 1,790 respondents overall; valid responses for each individualitem were not separately reported. Recruitment functions and competency preferences are different questions.KOS's outlook for China's 2026 consumer-goods talent market further emphasizes thatcross-border business positions in domestic firms must consider supply-chain conditions,target-market culture, local regulation, and social-media environments, with particularattention to data analytics, localized cross-border marketing, and sustainable supply-chaincapabilities (KOS, n.d., pp. 31–32). Combined with the recruitment evidence, this suggeststhat AI tools should be situated within international business contexts. Students need tounderstand the data, objects, and conditions on which AI outputs depend rather than merelylearning operating procedures. This corresponds more closely to the “evaluating and creating”and “ethical issues” dimensions of AI literacy than to “using and applying” alone (Ng et al.,2021).
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)31The evidence supports integrating digital capabilities with market development,localization, and supply-chain coordination, but it does not establish that any single AI skillhas become a universal requirement across all foreign-trade jobs. AI-related job growth isselective across establishments and tasks, and its observable short-run effect on aggregateemployment structures remains limited (Acemoglu et al., 2022). Requirements reported formid- and senior-level roles should likewise not be transferred directly into undergraduatelearning outcomes. Fourth-year students commonly begin with documentation, customerdevelopment, or platform operations; course objectives should therefore target entry-leveltasks rather than the complete competency profile of senior managers.4. Construction of the Curriculum Update Index4.1 Five-Dimensional Evaluation Logic and IndicatorsRecruitment demand can signal the need for content updating, but occupational relevanceis not equivalent to curricular admissibility. The CUI contains five dimensions: frontierrelevance (F), industry demand (I), job relevance (J), curriculum alignment (R), andinstructional feasibility (T). F captures the level and change of attention to a topic inauthoritative sources; I reflects identifiable business applications; J measures support forcompetency demand in recruitment evidence; R evaluates alignment with program objectives,course outcomes, task outputs, and module structure; and T reflects prerequisites, teachingresources, and available hours. The five first-level dimensions are equally weighted, while thesecondary indicators are equally weighted within each respective dimension. This choiceprioritizes transparency and interpretability and should not be understood as a weight structurevalidated by a large sample.퐶� � =�� + � +�� +�� +��5Table 4. Indicator System for the Curriculum Update IndexDimension/indicatorOperational definitionand scoringScope/conditionsF1 Currentattention100 × documentscontaining topic / validdocumentsFixed WTO, UNCTAD, and Ministry of Commercesources; duplicate topic mentions in one document countedonce.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)32F2 Change inattentionNormalize adjacent-period change in topic share to0–100Scale determined by candidate-topic distribution; nofixed 10-percentage-point ceiling.F3 RepositorynoveltyScore by when the topicfirst enters the observationrepositoryRepository novelty only; not a claim of global novelty.F4 Sourcecoverage100 × independentsupporting organizations / 3Syndicated copies are not double-counted.I1 Case coverage100 × firms involvingtopic / firms in valid caselibraryDescribes the case library; not an industry adoption rate.I2 Applicationmaturity100 × mean case maturity/ 30 concept; 1 trial; 2 partial application; 3 stable process.I3 Scenariocoverage100 × evidenced businesslinks / 8Research, development, marketing, transaction,payment, logistics, risk control, compliance.J1 Cross-organization support100 × organizationsexplicitly supporting topicdemand / 3Multiple reports from one organization count as oneorganization.J2 Functionalcoverage100 × directly relatedfunctional categories / 5Marketing, cross-border operations, logistics/supplychain, payments/finance, compliance/risk.J3 Evidencestrength0, 25, 50, 75, or 100None, qualitative outlook, qualitative retrospective,survey intention, observed job statistics; sensitivity analysisrequired.J4 Descriptioncompleteness25 points each foroccupation, competency,scenario, and task/outputAll elements must come from one traceable description.R1 Programrelevance0–100 for direct relationto program objectivesQuote the original program objective and explain therelationship.R2 Course-outcome support100 × directly supportedoutcomes / total outcomesExplain how the topic task supports each outcome.R3Task/assessmentmeasurability25 points each: output,rubric, evidence requirement,feedbackAvoid adding a topic name without an assessable task.R4 Modulearticulation25 points each:prerequisite, insertion point,boundary, module linkMissing records are not treated as satisfied.T1 Prerequisitecoverage100 × coveredprerequisite points / allrequired pointsExplicitly record when no prerequisites are required.T2 Resourcereadiness25 points each: case,data, tool, assessment criteriaResources must be accessible, sourced, andinstructionally appropriate.T3 Hour fit Assessed from Theory and practice hours are checked separately.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)33replaceable hours relative torequired hours4.2 Missing Data, Weighting, and RobustnessAfter a complete search, an indicator receives 0 when no supporting evidence is found;an indicator that has not yet been collected or cannot be judged is recorded as missing. Noformal ranking is produced when any required indicator is missing. The composite scoreorganizes evidence rather than replacing educational judgment, and essential foundationalcontent is not automatically removed because of a low score.Equal weighting is the primary analysis. Entropy weighting is used only as a robustnesscomparison and is not treated as a direct measure of educational importance. The entropy-weighted specification is treated as an alternative weighting structure for robustness analysis andtherefore does not preserve the equal 20% weights assigned to the five first-level dimensions inthe baseline model. When the number of candidate topics and data completeness are sufficient,entropy weights may be calculated for the 18 secondary indicators and aggregated into the fivedimensions. For normalized score x_ij, the calculations are:�푖 =�푖 푖��푖 � = −1푙�� 푖��푖 푙��푖 � = 1 − � � =� �� Zero terms are treated as zero in the entropy calculation, while all-zero or constant columnsprovide no discriminatory information.Robustness analysis compares equal-weight and entropy-weight results using Spearmanrank correlation, top-K overlap, and changes in recommendation categories. It also includesleave-one-organization-out tests, alternative scoring for J3, and changes to the observationwindow. Because entropy weights reflect sample dispersion rather than educationalimportance, substantial disagreement between weighting schemes must be resolved byreturning to the underlying evidence and course objectives.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)344.3 Illustrative Calculation and Recommendation RulesTo show how the framework produces actionable recommendations, six candidate topicsdirectly related to the existing course are used in an illustrative calculation. The scores inTable 5 are hypothetical inputs that assume indicator coding has already been completed.Their sole purpose is to demonstrate composite calculation, threshold decisions, andcorresponding curriculum actions; they are not empirical findings. Formal calculation requiresa source-level evidence record for every indicator in Table 4 and complete case, recruitment,and institutional curriculum data.Table 5. Illustrative CUI Calculation Using Hypothetical InputsCandidate topic F I J R T CUIRecommendationGenerative AI-assisted internationalmarket research908276928885.6PrioritystrengtheningCross-border data flows and ruleverification867064948479.6Integrate asspecial topicLocalized cross-border e-commerceoperations788284908884.4PrioritystrengtheningSmart logistics and overseas-warehousecoordination748486868482.8PrioritystrengtheningDigital payments and digital currency807670848078.0Integrate asspecial topicAI-assisted supply-chain riskidentification827066807474.4Integrate asspecial topicInitial thresholds are set as follows: 80 or above, priority strengthening; 70 to below 80,integration as a special topic; 60 to below 70, observation/extension; and below 60, no newaddition in the current cycle. Routine additions must also satisfy the course-fit conditions andfit separately within theoretical and practical hours. These thresholds are provisional rulesrequiring validation rather than universal admission criteria. Under the hypothetical inputs,localized operations, smart-logistics coordination, and generative-AI-assisted market researchwould receive priority strengthening, while the other three topics would be integrated asspecial topics. This demonstrates how the rules operate; it does not establish their empiricalpriority.5. Intelligent Recommendation and Curriculum Application5.1 Human-in-the-Loop Recommendation Process
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)35The recommendation process contains six stages. First, establish a trusted-source listcovering WTO, UNCTAD, the Ministry of Commerce, industry reports, corporate cases,recruitment materials, and institutional curriculum files. Second, use webpage updatedetection, text cleaning, topic extraction, entity recognition, semantic clustering, and trendalerts to identify new or changing information. Third, consolidate this information intocandidate topics under five categories: new technologies, new business forms, new models,new rules, and new occupational skills. Fourth, calculate the five CUI dimensions whiledisplaying the original evidence and missing items. Fifth, combine the composite score,curriculum-fit conditions, and relationships with existing content to generate recommendationssuch as add, strengthen, integrate, retain, compress, replace, or observe. Sixth, the teachingteam reviews both the recommendation and its evidence before deciding whether the itementers the course knowledge base and the next syllabus revision.AI output must not serve as a direct basis for curriculum revision. Each recommendationshould display its source, date, business scenario, related learning outcomes, content proposedfor replacement, and required hours so that teachers can trace the reasoning. When sourcesconflict, information is insufficient, or rules are unclear, the system should return only a“verification required” flag rather than a formal recommendation.5.2 Curriculum Restructuring Under Contact-Hour ConstraintsDrawing on the competency directions indicated by the recruitment reports and the ruledemonstration in Table 5, the curriculum scheme shows how generative-AI-assisted marketresearch, localized cross-border e-commerce operations, and smart-logistics coordination canbe embedded in existing modules. New content must be accommodated through internalmodule restructuring. The 54-hour allocation in Table 6 is an illustrative curriculumconfiguration developed through teacher judgment under the CUI recommendations and fixed-hour constraints. It is not a mechanical conversion of recruitment percentages and is not“proven” by the hypothetical scores.Table 6. Proposed Allocation of Hours for International Trade InnovationCourse module TheoryPracticeTotalAdjustment rationale/taskemphasisFoundations of international tradeinnovation and digital trade4 2 6Retain essential concepts andstatistical definitions.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)36Digital trade in services andcross-border data flows5 2 7Strengthen rule verification and data-governance analysis.Cross-border e-commerce andlocalized operations5 3 8Strengthen localization, platformoperations, and market adaptation.Digital marketing and generative-AI-assisted market research5 3 8Embed source verification, customeranalysis, and AI-use records.Digital payments and fintech 4 2 6Integrate digital payment/currencydevelopments while retaining transactionlogic.Smart logistics, overseaswarehouses, and supply-chaincoordination5 3 8Strengthen fulfillment design andcross-border coordination.Trade compliance and AI-assistedrisk identification4 1 5Integrate rule checking and risk-identification tasks.Integrated cross-border businessproject4 2 6Connect marketing, transaction,payment, logistics, and compliancedecisions.Total 36 1854Theory and practice hours remainfixed.If a module originally contains theoretical hours and practical hours, and the correspondingadjustments are defined separately, the revised plan must satisfy the following conservationconstraints: ᵢ(�ᵢ� + 훥�ᵢ�) = 36 ᵢ(�ᵢ� + 훥�ᵢ�) = 18Separate conservation of theoretical and practical hours ensures that every addition orstrengthening action has an explicit location for compression, replacement, or integration ratherthan creating hidden extra workload outside the stated 54 hours.5.3 Alignment of Objectives, Activities, and Assessment in Business TasksConsider the task “Generative-AI-assisted target-market research.” Students investigatethe entry of a specified product into a target market by conducting information retrieval,competitor comparison, customer-demand analysis, and marketing-plan design. AI may assistin summarizing information, generating search leads, and proposing alternative expressions,while students remain responsible for checking sources, identifying statistical definitions,judging conditions of applicability, and explaining target-market differences. Final outputsinclude a market brief, a source list, prompts and AI-use records, and a reflective statement.Assessment criteria address evidence quality, business analysis, plan feasibility, responsible
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)37AI use, and communication quality, thereby maintaining alignment among course outcomes,learning activities, and assessment evidence (Biggs, 1996).The task operationalizes the recruitment evidence on digital skills and marketdevelopment and converts localization competency into an activity appropriate for fourth-yearundergraduates. Platform operations, logistics, and payment modules can share the sameproduct-and-market scenario, allowing marketing judgments, transaction arrangements, andfulfillment plans to form a continuous business chain.6. Discussion and Conclusion6.1 Findings and Teaching ImplicationsThe curricular value of recruitment reports lies not in directly determining teaching hoursbut in revealing changes in occupational tasks and competency structures. Domestic foreign-trade positions account for a limited share of total job demand yet coexist with relatively hightalent-shortage indicators. Curriculum updating should therefore focus on competencyrestructuring within existing occupations rather than interpret localized trends as universalrequirements. Digital skills become instructionally meaningful when connected with marketdevelopment, localized operations, supply-chain coordination, and rule verification andtranslated into business tasks appropriate for undergraduates.The CUI places external knowledge change, industrial application, occupational demand,and internal instructional conditions within one evaluation structure. Its principal function is toorganize evidence, expose missing information, and create discussable reasons forrecommendations rather than replace the judgment of the teaching team. The illustrativecalculation links candidate topics, recommendation categories, and contact-hour allocation,while also showing that some frontier topics can be embedded within existing modules ratherthan continuously added as new chapters. In this human–AI division of labor, AI supportsincremental discovery and semantic organization, the CUI provides structured screening, andteachers make judgments concerning learning outcomes, disciplinary coherence, andeducational requirements that cannot be fully quantified.6.2 Limitations and Future ValidationThe study has four main limitations. First, recruitment evidence comes from foursecondary reports with different periods, populations, and methods and therefore cannot be
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)38pooled into a national estimate of job demand. Second, the illustrative calculation does not yetincorporate a large corpus of job advertisements; the J dimension mainly reflects supportacross institutional reports rather than the frequency of skills in job postings. Third, CUIweights, score levels, and thresholds are initial rules. Although a sensitivity-analysis plan isspecified, expert review and a larger sample of topics are still required. Fourth, the proposedcurriculum has not yet been evaluated over successive teaching cycles, so no inference can bemade that student competencies have already improved.Future work should proceed through data expansion, rule validation, teachingimplementation, and outcome evaluation. Subject to platform rules, researchers cansupplement the evidence base with public job advertisements or manually sampled records,expand the corporate-case library and frontier-material observation window, invite curriculum,industry, and educational-technology experts to assess content validity, compare equal, expert-derived, and entropy weights, and finally evaluate appropriateness through student work,course performance, and learning feedback. For other institutions and courses, the CUI shouldbe recalibrated to local program objectives and contact-hour structures rather than copying thethresholds used here.6.3 ConclusionThis study develops a dynamic content-updating and recommendation framework for anInternational Trade Innovation course. The framework uses trusted sources to detectknowledge change, AI to assist candidate-topic identification, and the CUI to evaluate frontierrelevance, industry demand, job relevance, curriculum alignment, and instructional feasibility.Teachers then decide whether to add, strengthen, integrate, compress, or replace content underthe fixed 54-hour constraint. The six-topic numerical demonstration shows how externalinformation can be converted into traceable curriculum recommendations while preserving the36-hour theory and 18-hour practice structure. The contribution is an interpretable andauditable human-in-the-loop tool; its effectiveness remains to be validated through empiricalscoring with real data and subsequent teaching practice.References
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)39Abu-Rasheed, H., Jumbo, C., Al Amin, R., Weber, C., Wiese, V., Obermaisser, R.,& Fathi, M. (2025). LLM-assisted knowledge graph completion for curriculum and domain modelling in personalized higher education recommendations. In 2025 IEEE Global Engineering Education Conference (EDUCON). IEEE. https://doi.org/10.1109/EDUCON62633.2025.11016377Acemoglu, D., Autor, D., Hazell, J., & Restrepo, P. (2022). Artificial intelligenceand jobs: Evidence from online vacancies. Journal of Labor Economics, 40(S1), S293–S340. https://doi.org/10.1086/718327Bekkers, E., Kalachyhin, H., & Teh, R. (2025). The long-run impact of digitalization on trade patterns (WTO Staff Working Paper No. ERSD-2025-01). World Trade Organization. https://www.wto.org/english/res_e/reser_e/ersd202501_e.htmBiggs, J. (1996). Enhancing teaching through constructive alignment. Higher Education, 32(3), 347–364. https://doi.org/10.1007/BF00138871Cheng, X., Su, L., & Zarifis, A. (2019). Designing a talents training model for cross-border e-commerce: A mixed approach of problem-based learning with social media.Electronic Commerce Research, 19(4), 801–822. https://doi.org/10.1007/s10660-019-09341-yDeming, D., & Kahn, L. B. (2018). Skill requirements across firms and labor markets: Evidence from job postings for professionals. Journal of Labor Economics, 36(S1),S337–S369. https://doi.org/10.1086/694106Gacek, P., & Adrian, W. T. (2025). Automated curriculum analysis using large language models and knowledge graphs. Intelligenza Artificiale, 19(2), 116–126. https://doi.org/10.1177/17248035251360196International Monetary Fund, Organisation for Economic Co-operation and Development, United Nations Conference on Trade and Development, & World Trade Organization. (2023). Handbook on measuring digital trade (2nd ed.). OECD Publishing. https://doi.org/10.1787/ac99e6d3-enNg, D. T. K., Leung, J. K. L., Chu, S. K. W., & Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2, Article 100041. https://doi.org/10.1016/j.caeai.2021.100041Reales, D., Manrique, R., & Grévisse, C. (2024). Core concept identification in educational resources via knowledge graphs and large language models. SN Computer Science, 5, Article 1029. https://doi.org/10.1007/s42979-024-03341-y
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)40Tyler, R. W. (1949). Basic principles of curriculum and instruction. University ofChicago Press.UNESCO. (2024). AI competency framework for students. UNESCO. https://doi.org/10.54675/JKJB9835United Nations Conference on Trade and Development. (2025). Technology and innovation report 2025: Inclusive artificial intelligence for development. United Nations. https://unctad.org/publication/technology-and-innovation-report-2025World Trade Organization. (2024). Trading with intelligence: How AI shapes and is shaped by international trade. WTO. https://www.wto.org/english/res_e/publications_e/trading_with_intelligence_e.htmWorld Trade Organization. (2025). World trade report 2025: Making trade and AIwork together to the benefit of all. WTO. https://www.wto.org/english/res_e/publications_e/wtr25_e.htmZawacki-Richter, O., Marín, V. I., Bond, M., & Gouverneur, F. (2019). Systematicreview of research on artificial intelligence applications in higher education – Where are the educators? International Journal of Educational Technology in Higher Education,16, Article 39. https://doi.org/10.1186/s41239-019-0171-0Zou, Z. H., Yun, Y., & Sun, J. N. (2006). Entropy method for determination ofweight of evaluating indicators in fuzzy synthetic evaluation for water quality assessment. Journal of Environmental Sciences, 18(5), 1020–1023. https://doi.org/10.1016/S1001-0742(06)60032-6Ministry of Education of the People’s Republic of China, National Development and Reform Commission, Ministry of Industry and Information Technology, et al. (2026).Notice on issuing the “AI + Education” Action Plan (Jiaokexin [2026] No. 1). https://www.moe.gov.cn/srcsite/A16/s3342/202604/t20260410_1433240.htmlKOS International. (n.d.). 2026 China talent market recruitment trends [Research report].Liepin Big Data Research Institute. (2025a). 2025 recruitment trend survey report:Quarter 1 [Research report].Liepin Big Data Research Institute. (2025b). 2025 recruitment trend survey report:Quarter 2 [Research report].Randstad. (n.d.). Randstad China white paper: 2025 talent recruitment strategies and trend outlook [Research report].
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)41Ministry of Commerce of the People’s Republic of China. (2025). Development oftrade in services in 2024. https://data.mofcom.gov.cn/article/zxtj/202502/62678.htmlMinistry of Commerce of the People’s Republic of China. (2026). Development oftrade in services in 2025. https://www.mofcom.gov.cn/syxwfb/art/2026/art_7d8d9f93478046b59c08b1639900849d.html
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)42The Age of Artificial Intelligence: LaborForce Transformation, Productivity Expansion,and the Paradigmatic Reconstruction ofDistributive EconomicsZhiping Ren¹, Yangyan He²**Corresponding author: Yangyan He¹Lingnan Science Industry Press, Research Fellow;²Department of Economics, Nanfang College, Guangzhou, China, yaraine@163.comKeywords ABSTRACTartificial intelligence;automation; labor; humancapital; income distribution;universal basic income;economics of meaningGenerative artificial intelligence, autonomous agents, and embodiedintelligence are extending the range of cognitive and physical tasks thatmachines can perform. This development places new pressure on theconventional assumption that labor and capital are complementary inputs. Thispaper examines the changing explanatory scope of the Cobb-Douglasproduction function, labor-value theory, and human-capital theory underconditions of advanced automation. It distinguishes task substitution,occupational restructuring, and the complete withdrawal of labor fromproduction, and argues that these propositions should not be conflated. A set offormal expressions is used to clarify the relationships among automation, laborincome, effective demand, and the sharing of returns to intelligent capital. Theanalysis further considers universal basic income, data governance, and humandevelopment as possible elements of an institutional response. The paper doesnot predict a fixed timetable for technological displacement. Rather, it offers aconditional analytical framework for understanding how the focus of economicsmay shift from production efficiency toward distribution, demand maintenance,and human well-being.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)431. IntroductionLabor has never been merely a technical variable in economics. It is an input intoproduction, but it is also the principal route through which most people enter the income-distribution system. Economics has traditionally begun from scarcity and has treated theallocation and exchange of limited resources as a central problem (Robbins, 1932). Duringindustrialization, machines repeatedly replaced particular operations, yet new industries,occupations, and educational pathways often reorganized labor. This experience helpedestablish the influential view that technical change raises productivity while workers movetoward tasks of higher value.Artificial intelligence creates a different concern because it may enter several links of theproduction process at once. Text, image, and code generation affect knowledge work; agentscan connect retrieval, planning, and workflow execution; embodied systems seek to combineperception, decision-making, and physical action. Agrawal, Gans, and Goldfarb (2018)describe a major economic effect of AI as a reduction in the cost of prediction, whileBrynjolfsson and McAfee (2014) draw attention to the adjustment pressures created by theautomation of cognitive tasks. The relevant questions are therefore not whether labor willdisappear overnight, but which tasks are substituted, which are reorganized, who receives thenew returns, and how income remains connected to demand.This paper does not treat the phrase 'the uselessness of human resources' as an empiricalfact. It uses the phrase as a theoretical stress test. The question is whether, as a larger set ofproductive tasks can be performed by intelligent capital, established accounts of production,wages, and human-capital investment require qualification. The paper is a theoretical andnormative inquiry; it does not estimate the speed of AI diffusion or forecast a determinateemployment outcome.2. Conceptual Scope and Analytical Starting PointA discussion of AI and labor should not treat occupations, jobs, and tasks asinterchangeable. Occupations are bundles of tasks, and jobs are embedded in organizations,responsibility chains, and institutional settings. Information sorting, preliminary assessment,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)44and standardized communication within the same occupation may be automated, while on-sitecoordination, accountability, difficult negotiation, and value judgment may remain humantasks. Frey and Osborne (2017) estimate the technical susceptibility of occupations tocomputerization; such estimates cannot be read directly as unemployment forecasts becauseadoption also depends on cost, regulation, organizational adjustment, and demand.Three propositions should therefore be distinguished. First, a declining labor share meansthat labor compensation forms a smaller proportion of total income. Second, task restructuringmeans that workers move from direct execution toward supervision, verification, coordination,maintenance, or relational service. Third, the complete withdrawal of labor means thatproduction no longer requires humans to perform any economically meaningful task. Autor(2015) notes that automation may replace tasks but may also expand output and createcomplementarities that raise labor demand. The first two changes can be observed in somesettings; the third is an extreme case requiring additional conditions.The task framework developed by Acemoglu and Restrepo (2018) is useful in this regard.Automation can displace workers from existing tasks, whereas the creation of new tasks canincrease labor demand. The relative strength of these forces is shaped not by technology alonebut also by capital costs, education, market structure, and public policy. Labor's possibleremoval as a production factor is thus not a timetable. It is a conditional proposition forexamining the limits of existing theory.3. The Labor Term in the Production Function and Its BoundaryIn its common form, the Cobb-Douglas production function is written as follows:�� = ������ (1)Here, Y_t denotes total output in period t; A_t denotes total factor productivity; K and Ldenote capital and labor inputs; and alpha and beta are the relevant output elasticities. Alphaplus beta equals one only under constant returns to scale. Output elasticities can be related tofactor-income shares only under further assumptions, such as competitive factor markets. Theequation therefore does not by itself justify an empirical conclusion about labor compensation.The effect of AI can be represented by an expanded production relationship. The purposeof the following expression is not to provide a directly estimable structural model, but to show
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)45how computing, energy, data, and institutional conditions may jointly affect output:�� = ���(���, ���,��,��;��) (2)In Equation (2), K_t^c denotes computing capital, K_t^e energy and equipment capital,D_t data that can be lawfully used, and Z_t organizational capability, regulation, networkinfrastructure, and other complementary conditions. Labor remains in the function becauseactual automation does not imply the immediate disappearance of labor. The expressionhighlights that AI productivity depends on several complementary inputs. Ignoring datagovernance, energy constraints, or organizational adaptation would overstate what a modelalone can accomplish.A change in labor's distributive position can be observed through the labor-income share:��,� =�_� �_��_� (3)In Equation (3), w_t is unit labor compensation and s_L,t is the labor-income share inperiod t. A lower labor share may result from lower labor input, real wages growing moreslowly than output, or faster growth in capital income. It is not equivalent to the propositionthat labor no longer participates in production. A careful assessment must consider wages,employment, hours, capital income, and prices together rather than treating one indicator asdecisive.4. Labor Value, Human Capital, and Organizational ChangeLabor-value theory places living labor at the center of value creation and relatescommodity value to socially necessary labor time (Marx, 1990). Production settings in whichalgorithms plan, robots execute, and automated systems inspect output raise a serioustheoretical question: how should price, profit, and distribution be interpreted if direct humanlabor declines as a scarce input? This question calls for a careful statement of theoretical scope;it does not erase labor's historical or continuing role.Human-capital theory faces a new comparison as well. Schultz (1961) and Becker (1964)treat education, training, and health as investments that yield returns. An AI model mayrequire substantial initial training resources, yet its capabilities can be deployed and updated atmany endpoints. This can reduce the scarcity premium attached to some standardized
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)46knowledge skills and change how organizations allocate training resources. Susskind (2020)discusses the risk that technical change may weaken the stability of labor income. Yet thereturns to learning are not exhausted by wages: critical judgment, cooperation, professionalethics, and civic participation remain essential capabilities in a complex society.Organizations will not become mere collections of interfaces. When models and softwaresystems assume part of execution, organizations must devote greater effort to goal setting, dataquality, model monitoring, responsibility tracing, and exception handling. Traditional human-resource concerns about incentives and coordination do not disappear; they becomeinterwoven with system reliability, algorithmic governance, and the allocation ofresponsibility. The phrase 'a shift from human-resource management to systems governance'describes this change more accurately than the claim that human resources have becomeuseless.5. Automation, Income Distribution, and Effective DemandThe central challenge of a highly automated economy is not simply whether more can beproduced. It is whether additional output can be converted into broad and durableimprovements in living conditions. Ford (2015) emphasizes that productivity gains do notautomatically translate into improved household income. The national-income identityprovides a simple point of departure:�� = 퐶� + � +�� +푁 � (4)In Equation (4), C_t, I_t, G_t, and NX_t denote consumption, investment, governmentexpenditure, and net exports. Automation can raise potential output without ensuring thatconsumption demand rises at the same pace. If households rely primarily on wages fordisposable income while labor compensation grows more slowly than output, the consumptionside may become a constraint on the economic circuit.Household disposable income can be written more explicitly as:��� =���� +�� +� � −�� (5)In Equation (5), Y_t^d is disposable income, W_tL_t is labor income, Pi_t is propertyincome received by households, Tr_t is transfer income, and T_t is taxes and charges. Ifautomation mainly raises capital income while ownership of capital is highly concentrated,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)47market income alone may not convert productivity gains into purchasing power for mostresidents. The issue then extends from whether jobs exist to how households share in growth.Keynes's reflections on technology and leisure remind us that material abundance doesnot automatically create social stability (Keynes, 1930/1963). When productive capacity rises,distributional institutions, public services, and working-time arrangements affect whethertechnical gains are widely shared. Weak effective demand may, in turn, reduce investmentincentives, leave capacity idle, and diminish the benefits of technological progress.6. Institutional Options for Sharing Returns to AutomationUniversal basic income is frequently discussed in relation to automation. Its centralpurpose is not to assume that everyone should withdraw from the labor market. Rather, itprovides a basic floor that is not conditional on employment when income sources becomeunstable. If part of the returns associated with automation is placed in a public pool, per-capitabasic income can be represented for accounting purposes as follows:�� � =�_� �_�푁_� , 0 ≤ �_� ≤ 1 (6)In Equation (6), UBI_t is per-capita basic income, �_� is automation-related incomethat is available for public sharing, N_t is the eligible population, and eta_t is the proportionused for transfers. The expression only records the relationship between a funding source andbeneficiaries. It does not imply that corporate profit is the only feasible tax base. A practicalarrangement must address the definition of profit, cross-border tax bases, fiscal substitution,public-service provision, and inflation risks.Van Parijs and Vanderborght (2017) describe basic income as an unconditional cashtransfer. It may reduce the administrative burden of eligibility tests and provide a buffer fornon-standard employment, care work, and retraining. Its contested features include fiscalsustainability, labor-supply effects, and its relationship to existing social insurance. For China,the more useful question may be how cash transfers, public education, health protection,unemployment insurance, transition services, and data-dividend arrangements can be assessedas a combined institutional package rather than as a single imported model.Institutional design must also address ownership and competition. When computingcapacity, data, and platform access are highly concentrated, automation returns may intensify
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)48inequality through market power rather than productivity alone. Competition policy, dataportability, public computing capacity, worker participation, and tax coordination all affectwho receives the gains. Technology policy should not focus only on deployment speed; itmust also consider who bears adjustment costs, who has bargaining power, and who canparticipate in rule-making.7. Data, Experience, and the Evaluation of Well-BeingIn highly automated systems, people do not become merely consumers. Their preferences,satisfaction, complaints, refusals, and creative feedback continue to inform productimprovement, model calibration, and public-service design. Because such signals involveindividual rights and social differences, data cannot be treated as a costless input. Datacollection, changed uses, benefit sharing, and algorithmic explanation should be governed byinformed consent, data minimization, and accountability.If employment is no longer the sole measure of social contribution, economic evaluationalso requires a fuller set of criteria. Social well-being can be represented as a weightedcombination of several dimensions:�� = �1�� +�2�� +�3�� +�4��,�1 +�2 +�3 +�4 = 1 (7)Equation (7) is a normative framework rather than a ready-made measure for cross-country comparison. P_t can represent basic material security, R_t rights and fairness, S_tsocial participation and relationship quality, and E_t opportunities for education, culture, andself-development. The weights omega_i express values established through public choice. Thepoint is not to replace complex social judgment with a single number, but to make clear thatproductivity is not the only public objective that institutions should recognize.An economics of meaning should not be interpreted as replacing material conditions withsubjective feeling, nor should it serve as a rhetorical device for reducing public protection. Itsreasonable core is that, when technology reduces some necessary labor time, society canreconsider the value of education, care, culture, community service, and democraticparticipation. Education should not be understood only as preparation for employable skills; italso cultivates judgment, aesthetic capacity, responsibility, and the ability to live with others.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)498. ConclusionArtificial intelligence expands the possibility that capital substitutes for labor, but it doesnot establish that labor must disappear. The more defensible conclusion is that labor will bedifferentiated across tasks and sectors, that the stability of wage income as the main source ofincome for most people may weaken, and that questions of distribution, demand, andgovernance will become more prominent. The labor term in a production function must beunderstood in relation to task structure, complementary inputs, and institutional conditions; itcannot simply be deleted.The formal expressions in this paper clarify several related relationships. Higherproductivity does not imply the complete withdrawal of labor. A change in the labor sharedoes not imply a loss of human value. If automation returns do not become householddisposable income, the economic circuit may face an effective-demand constraint. Social well-being cannot be evaluated by output alone. These propositions require empirical examinationusing industry-level task data, corporate financial data, worker panels, and comparativeinstitutional research.Two simplifications should be avoided in policy discussion: technological optimism thatconceals transition costs, and technological fear that rejects productivity growth. A moreconstructive course is to promote innovation while improving social protection, educationaltransition, data governance, competition policy, and mechanisms for sharing returns. Theobjective is to make technology an instrument that expands human choice rather than a forcethat intensifies the concentration of income and power.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)50ReferencesAcemoglu, D., & Restrepo, P. (2018). The race between man and machine: Implications oftechnology for growth, factor shares, and employment. American Economic Review, 108(6), 1488–1542. https://doi.org/10.1257/aer.20160696Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction machines: The simple economics ofartificial intelligence. Harvard Business Review Press.Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplaceautomation. Journal of Economic Perspectives, 29(3), 3–30. https://doi.org/10.1257/jep.29.3.3Becker, G. S. (1964). Human capital: A theoretical and empirical analysis, with special referenceto education. National Bureau of Economic Research.Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperityin a time of brilliant technologies.W. W. Norton & Company.Cobb, C. W., & Douglas, P. H. (1928). A theory of production. American Economic Review,18(Supplement), 139–165. https://doi.org/10.2307/1811556Ford, M. (2015). Rise of the robots: Technology and the threat of a jobless future. Basic Books.Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs tocomputerisation? Technological Forecasting and Social Change, 114, 254–280.https://doi.org/10.1016/j.techfore.2016.08.019Keynes, J. M. (1963). Economic possibilities for our grandchildren. In Essays in persuasion (pp.358–373). W. W. Norton & Company. (Original work published 1930)
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)51Marx, K. (1990). Capital: A critique of political economy (Vol. 1; B. Fowkes, Trans.). PenguinClassics. (Original work published 1867)Robbins, L. (1932). An essay on the nature and significance of economic science.Macmillan.Schultz, T. W. (1961). Investment in human capital. American Economic Review, 51(1), 1–17.Susskind, D. (2020). A world without work: Technology, automation, and how we should respond.Metropolitan Books.Van Parijs, P., & Vanderborght, Y. (2017). Basic income: A radical proposal for a free societyand a sane economy. Harvard University Press.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)52Curriculum Reform and Evidence-BasedAssessment in Securities Investment TechnicalAnalysis through Industry–EducationIntegrationDong Sun1 · Jiabei Xu1 *1 School of Business, Nanfang College Guangzhou, Guangzhou 510970, China*Corresponding author: Jiabei Xu, xujb@nfu.edu.cnKeywords ABSTRACTIndustry–educationintegration; securitiesinvestment technical analysis;applied undergraduateeducation; evidence-basedassessment; cognitivedifferentiation; dualmentoring; constructivealignmentTo address the disconnect between instruction in technical indicators andintegrated trading simulations, as well as the tendency for short-term profitseeking to obscure weaknesses in risk management, this study examines a corefinance course at the School of Business, Nanfang College, Guangzhou. Itdevelops an industry–education integration framework linking courseobjectives, occupational tasks, and learning assessment. Drawing onconstructive alignment and outcome-based education, market interpretation,investment-plan design, and risk management are organized into threeprogressive stages within a 54-hour course. Teaching combines authentic, de-identified industry cases, joint academic–industry mentoring, and two rounds ofintegrated simulation. Semester-long paired data from 42 enrolledundergraduates, collected at baseline (T1) and course completion (T2), areanalyzed through item discrimination, construct differentiation, parametric andnonparametric comparisons, and post hoc power analysis. Self-reported scoresincreased across all three dimensions, with mean gains of 0.542–0.613 points (p< 0.001). Risk management had the largest effect size among the threedimensions (dz = 1.150). The reduced reliability of investment-plan design at T2(α = 0.468) is interpreted as a separation between expectations of selectingwinning securities and adherence to a sound decision process under marketuncertainty. The study further proposes a four-dimensional performance rubric
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)53Keywords ABSTRACTand a multisource assessment framework to support curriculum reform andevaluation in applied finance and business education.1. IntroductionSecurities Investment Technical Analysis is a core professional-skills course in appliedundergraduate programs in finance, investment, and financial engineering. Its content includescandlestick analysis, chart patterns, trend indicators, moving averages, price–volumerelationships, and time cycles. Given its practical orientation, teaching must extend beyondmemorizing formulas and identifying isolated patterns. Students need to develop coherentinvestment analysis, strategy formulation, and dynamic risk management in market situationscharacterized by uncertainty, asymmetric information, and conflicting technical signals.Three structural tensions recur in conventional teaching. First, fragmented instructionconflicts with the integrated nature of occupational tasks. Indicators such as MACD and KDJare often taught separately, leaving students insufficiently prepared to reconcile multiplesignals in changing markets. Second, the emphasis on returns in simulation exercises conflictswith professional discipline and compliance. Short trading competitions may encourageconcentrated positions and frequent speculative transactions, diverting attention from positionsizing, drawdown limits, stop-loss rules, and model failure. Third, superficial industryparticipation is accompanied by weak assessment evidence. Lectures and visits do notnecessarily translate industry resources into progressively demanding learning tasks, whileend-of-semester satisfaction surveys and impression-based grading offer limited longitudinalor psychometric evidence of learning.Chinese research has explored several responses. Gong et al. developed a one-stopfinance education model connecting campus-based training, collaborative projects, andworkplace practice [1]. Ren and Zhang proposed progressive project-based practice forfinancial mathematics [2], while Zhang and Cao examined support mechanisms forcollaborative teaching teams [3]. Han and Chen et al. applied outcome-based education toproject-based reform in economics and management courses [4–5]. International research has
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)54examined active learning, project-based learning, and university–industry collaboration inrelation to employability [6–8]. Nevertheless, much of this work emphasizes broadframeworks or qualitative accounts. Less attention has been paid to how authentic investmentworkflows are translated into specific classroom tasks and how pre–post self-assessment datacan be examined through psychometric and statistical analysis.This study investigates Securities Investment Technical Analysis in the finance programat the School of Business, Nanfang College, Guangzhou. It presents a constructive-alignment-based framework connecting objectives, tasks, learning outputs, and assessment. Through anoff-campus practice base and an academic–industry mentoring team, the 54-hour courseincorporates de-identified market cases, two progressive simulation rounds, and stress-scenario reviews. Semester-long pre–post data from 42 enrolled students are examined usingreliability diagnostics, parametric and nonparametric tests, and post hoc power analysis. Theaim is to provide a replicable teaching approach and inspectable assessment evidence forapplied undergraduate finance and business courses.2. Theoretical Foundations and the Curriculum Reform Framework2.1 Constructive Alignment and Outcome-Based Course DesignJohn Biggs's theory of constructive alignment emphasizes consistency among intendedlearning outcomes, teaching and learning activities, and assessment tasks [9]. Combined withthe backward-design principle of outcome-based education, it informs a framework in whichoccupational competence requirements guide course objectives, objectives determine learningoutputs, and those outputs define the evidence required for assessment.The course organizes its intended professional outcomes into three interdependent,progressively developed dimensions.1. Market interpretation. Students move beyond chart recognition to identify relevantinformation in market data, explain the buying and selling pressures reflected in technicalindicators, and specify the conditions under which their interpretations hold. When indicatorsconflict, they are expected to consider price–volume movements and the broader economicand industry context before making cautious probabilistic judgments.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)552. Investment-plan design. Students translate their analysis into a coherent simulatedinvestment plan. The plan specifies security-selection criteria, entry triggers, staged positionbuilding and adjustment, and exit arrangements, and is presented in a research formatconsistent with financial-sector practice.3. Risk management. This central dimension challenges the pursuit of certainty andexceptional profits. Students are expected to recognize market uncertainty and model failure,establish stop-loss and take-profit thresholds in relation to volatility and drawdown constraints,and incorporate position-management discipline and compliance into professional practice.Table 1. Alignment of course objectives, tasks, outputs, and assessmentCoreobjectiveLearningtasksObservableoutputs Assessment focusMarketinterpretationResolveconflictingindicators; reviewextreme and range-bound marketsAnnotatedmultihorizon charts;analysis of conflictingsignals andprobabilisticjudgmentsData-sourcecompliance;assumptions; indicatorsaturation and failure;counterexamplesInvestment-plan designDraftinstitutional-stylesimulatedstrategies;construct portfoliosunder constraintsStandardizedinvestment-decisionreport; parameterrationale andimplementation planSecurity-selectionlogic; entry and exittriggers; completenessand feasibilityRiskmanagementBlack-swanstress scenarios;review andcorrection acrosstwo simulationsStop-loss andtake-profit checklist;position-rebalancingrecord; reflectivesimulation logDrawdowntolerance; mandatorystop-loss adherence;model-failurecontingencies; ethicsand compliance2.2 Translating Industry Workflows into Learning TasksA central challenge in industry–education integration is translating complex, specializedprofessional work into instructional scaffolds suited to undergraduates' zone of proximaldevelopment. Directly importing sophisticated private-fund models or large informationstreams may create cognitive overload, whereas excessively simplified fictional cases maylose practical relevance.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)56The course uses a three-stage translation process. First, industry mentors de-identifyauthentic investment cases, remove confidential parameters, and select historical decisionpoints while retaining market noise and conflicting signals. Second, subsequent pricemovements are withheld: students record decisions based only on information available at thetime, after which later information is disclosed incrementally. Third, retrospective reviewexamines execution and attribution rather than profit alone. Questions about unauthorizedposition increases or failures to exit after a stop-loss breach direct attention to the reasoningand discipline underlying trading behavior.3. Organization and Implementation of the Reform3.1 The Practice Base and the Dual-Mentor WorkflowThe course operates through a provincial-level practice teaching base jointly establishedby the School of Business, Nanfang College, Guangzhou, and Guangdong Xugu PrivateSecurities Investment Management Co., Ltd. In April 2025, the partners signed a five-yearstrategic agreement covering joint syllabus development, case preparation, practicesupervision, and assessment of investment-research presentations.Academic and industry mentors have complementary responsibilities. Academicinstructors manage teaching progress, explain the mathematical logic and limits of indicators,and regulate cognitive load. Industry mentors provide investment cases, conduct stress-scenario questioning, and offer written feedback on compliance and practical feasibility. Bothassess the same student assignments and simulated decisions from academic and professionalperspectives, helping connect course learning with industry practice.3.2 A Progressive 54-Hour Modular DesignThe course carries three credits and comprises 54 instructional hours, divided equallyamong foundational interpretation, strategy integration, and integrated simulation and defense.Each module receives 18 hours. Practical work and discussion account for more than 55% ofthe course, as detailed in Table 2.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)57Table 2. Modules, instructional hours, and progressive task outputsModule Hours Knowledgeand progressionPracticetasksAssessedoutputsFoundation18Lecture10; practice8Candlesticks,patterns, trendlines, andchangingsupport/resistance;limits of DowTheory and thethree assumptionsof technicalanalysisComparebullish andbearishinterpretationsunder identicalinformation;annotate keyhistoricallevelsAnnotatedsecurity chart;comparison ofthe limits ofindividualindicatorsIntegration18Lecture6; practice12ConflictsbetweenMA/MACD andKDJ/RSI;interactions amongvolume, price,time, and pricerange; divergencetrapsExamineindicatorconflicts inrange-boundand directionalmarkets; draftbasicquantitativeselection rulesIntegratedindicatoranalysis;entry/exittriggers andrisk checklistSimulation18Practice12; defense 6Expectedvalue in moneymanagement;variants of theKelly criterion;position anddrawdown control;psychology anddiscipline undershocksTwoprogressivehigh-fidelitymarketsimulations;industry-ledpresentationsand risk-managementdefensesCompleteinvestmentreport; joint-mentorquestioning;executionreview andreflection ondecision biases3.3 Two Simulation Rounds and Adversarial ReviewThe integrated simulation module includes two high-fidelity trading exercises. The first,baseline strategy verification, runs in weeks 9–12. Students select from a pool of five blue-chip and growth securities, formulate plans using previously taught methods, and executesimulated transactions. The exercise is intended to expose emotional decision-making, delayedexecution, and excessive position concentration. Industry mentors then review transaction
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)58records with the class, focusing on breaches such as retaining positions after stop-loss triggersor disproportionately increasing losing positions.The second round, dynamic shocks and strategy transfer, runs in weeks 13–15. Itintroduces a new market setting and simulated shocks, such as an unexpected interest-rateincrease or a security-specific black-swan event. Students have 48 hours to submit anemergency risk-response plan and explain position adjustments. This tests whether lessonsfrom the first round transfer to a new setting. In week 16, teams present their investmentresearch and respond to questions from academic and industry assessors about reasoning, riskdiscipline, and compliance.3.4 Separating Course Grading from Research AssessmentCourse assessment comprises 60% continuous assessment and 40% final assessment.Continuous assessment includes case assignments (15%), procedural compliance andexecution discipline in the first simulation (20%), the second-round stress-response report(15%), and participation in classroom questioning (10%). The final assessment comprises theinvestment-research report (25%) and oral defense (15%). Course grading is procedurallyseparated from the longitudinal anonymous research survey so that students can report theirperceptions without concern about grades.4. Research Design and Data Analysis4.1 Participants and Data CollectionThe study follows finance undergraduates from the 2022-entry cohort at the School ofBusiness, Nanfang College, Guangzhou, who enrolled in Securities Investment TechnicalAnalysis. It uses a longitudinal, single-group pre–post quasi-experimental evaluation. Baselinedata were collected in week 1, before systematic industry-integrated training, and endpointdata in week 16, after course completion and the final defense.The survey system generated a unique encrypted identifier for each student withoutrecording names or student numbers. Instructions stated that participation served pedagogicalimprovement and academic research, that responses would remain confidential from thecourse instructor, and that results would not affect course grades. All students responded at
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)59both time points, yielding 42 complete paired records (N = 42). Logical checks found allresponses within the 1–5 range, no missing values, and no cases of uniform extremeresponding across the entire questionnaire.4.2 Measurement Constructs and Interpretive BoundariesThe instrument uses a five-point Likert scale, from 1 (does not describe me at all) to 5(describes me very well). Twelve professional self-assessment items are distributed equallyacross market interpretation (Q1–Q4), investment-plan design (Q5–Q8), and risk management(Q9–Q12). Four teaching-experience and training-satisfaction items (Q13–Q16) areadministered at T2 to provide contextual feedback (Table 3).Table 3. Self-assessment constructs and interpretive boundariesDimension Items IntendedconstructInterpretiveboundaryMarketinterpretationQ1 Patterns;Q2 Trends; Q3Multipleindicators; Q4IndependentinterpretationReasoning frompatterns and multiplesignals, includingcounterexamples, innoisy marketsPerceivedunderstanding ofanalytical tools; not ameasure of perfectlyaccurate forecastsInvestment-plan designQ5 Success-orientedselection; Q6Coherentreasoning; Q7Trading strategy;Q8 ReportwritingPlanningexecutableinvestment reportsand dynamic tradingstrategies throughsystematic researchPerceived coherenceof strategy design; notevidence of guaranteedexcess returnsRiskmanagementQ9 Failurerisks; Q10 Falsesignals; Q11Positions andstops; Q12ComplianceRecognition ofuncertainty andcommitment to stop-loss discipline andcompliant riskmanagementInternalized risk andcompliance awareness;requires complementaryevidence of stop-lossexecution
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)60Dimension Items IntendedconstructInterpretiveboundaryTeachingexperienceQ13Industry cases;Q14 Integratedsimulation; Q15Dual mentoring;Q16 EngagementPerceptions ofteaching, case depth,mentoring, andlearning engagementReported separatelyas course feedback;excluded from theprofessional competencecomposite4.3 Statistical Analysis StrategyThe analysis combines four complementary procedures to provide a more detailedexamination than a single-method evaluation.1. Internal consistency and construct diagnostics. Cronbach's alpha is calculated for eachdimension and the full scale at T1 and T2. Alpha if an item is deleted and corrected item–totalcorrelations (CITCs) are used to investigate dimensions with substantial reliability declinesand explore their educational interpretation. Interpretation of internal consistency follows themethodological guidance of Tavakol and Dennick [10].2. Paired-samples t tests. Individual change scores are calculated as T2 minus T1. With41 degrees of freedom, the analysis estimates t statistics, 95% confidence intervals for meanchanges, and paired effect sizes (Cohen's dz), following Lakens [11].3. Nonparametric comparisons and robustness checks. Given the ordinal origin of Likertscores and the possibility of nonstandard distributions, Wilcoxon signed-rank tests and exacttwo-sided sign tests are used alongside the parametric analysis to examine the direction anddistribution of changes.4. Multiple-testing adjustment and post hoc power. The Holm–Bonferroni procedure isused to control the family-wise error rate. Post hoc statistical power (1 − β) is also estimatedusing the G*Power 3.1 algorithm to assess detectability with 42 paired observations at theobserved effect sizes.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)615. Assessment Results and Diagnostic Analysis5.1 Reliability and Corrective Cognitive DifferentiationTable 4 reports reliability estimates and item diagnostics at both time points. The full 12-item scale has alpha values of 0.803 at T1 and 0.718 at T2, while the T2 teaching-experiencescale has an alpha of 0.763. The manuscript interprets their values above the conventional 0.70threshold as evidence of satisfactory overall internal consistency and stability across the twostages.Table 4. A. Reliability estimatesDimension Items T1 α T2 αMarket interpretation 4 0.709 0.691Investment-plan design 4 0.775 0.468Risk management 4 0.765 0.620Professional composite 12 0.803 0.718Teaching experience 4 Notmeasured 0.763Table 4. B. Item diagnostics and source interpretationDimensionMaximumT2 α afterdeletionKey item Interpretation reportedin the sourceMarketinterpretation0.704(delete Q4)Q4 (r =0.412)Highly stable reliabilitystructure and favorableconsistency across stagesInvestment-plan design0.582(delete Q5)Q5 (r =0.185)Cognitive differentiationbetween expected success andprocedural disciplineRiskmanagement0.638(delete Q12)Q12 (r =0.345)Acceptable reliability anddeeper recognition ofuncertaintyProfessionalcomposite — — Favorable overallreliability and validity
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)62DimensionMaximumT2 α afterdeletionKey item Interpretation reportedin the sourceTeachingexperience0.781(delete Q13)Q13 (r =0.488)Satisfactory experience-scale reliability and highacceptanceAt the dimension level, investment-plan design shows a decline in alpha from 0.775 at T1to 0.468 at T2. Although an alpha below 0.50 may indicate a substantial measurement problem,the study interprets the item diagnostics in conjunction with the simulation context aspotentially informative evidence about the teaching intervention.1. Separating expectations of winning from procedural discipline. At baseline,inexperienced students may assume that mastering standard analytical procedures (Q6–Q8)necessarily enables them to identify securities with a high probability of success (Q5),producing closely aligned item responses. After two simulation rounds involving volatility andunexpected drawdowns, students may recognize that a rigorous, compliant plan does notguarantee success. The source reports a marked increase in the variance of Q5 and a CITC of0.185 with the remaining plan-design items. Removing Q5 reportedly increases alpha to 0.582.2. Corrective cognitive differentiation in professional education. The study interprets thereliability decline as a reduction in unwarranted confidence, described through the Dunning–Kruger effect, and a separation between methodological rigor and uncertain market outcomes.On this interpretation, the decline marks developing professional judgment rather thanteaching failure. It suggests separating outcome-implying items such as Q5 from amethodology-focused scale and developing a distinct item set on market risk and attribution ofsuccess.5.2 Changes in Self-Assessment and Robustness ComparisonsTable 5 compares self-assessment scores before and after the course for 42 students. Allthree professional dimensions and the composite score show statistically significant increases(all p < 0.001), with consistent findings across the reported methods.Table 5. A. Descriptive statistics and mean changes (N = 42)Dimension T1 M ± SD T2 M ± SD Change [95% CI]
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)63Dimension T1 M ± SD T2 M ± SD Change [95% CI]Marketinterpretation2.899 ±0.6103.440 ±0.532 0.542 [0.376, 0.708]Investment-plan design2.988 ±0.6493.589 ±0.490 0.601 [0.410, 0.792]Riskmanagement2.911 ±0.6743.524 ±0.505 0.613 [0.447, 0.779]Compositescore2.933 ±0.4783.518 ±0.373 0.585 [0.468, 0.703]Table 5. B. Paired tests, effect sizes, and power (N = 42)Dimension Pairedt testCohen’sdzWilcoxonZExactsign testPosthoc powerMarketinterpretationt =6.595; p <0.0011.018Z =−4.68; p <0.001<0.001;34 up / 5equal / 3down>0.999Investment-plan designt =6.361; p <0.0010.982Z =−4.49; p <0.001<0.001;34 up / 2equal / 6down>0.999Riskmanagementt =7.454; p <0.0011.150Z =−4.92; p <0.001<0.001;36 up / 2equal / 4down>0.999Compositescoret =10.075; p <0.0011.555Z =−5.24; p <0.001<0.001;40 up / 1equal / 1down>0.999Note: Holm–Bonferroni adjustment yields p values below 0.001 for all four comparisons:market interpretation, 1.23 × 10⁻⁷; investment-plan design, 1.33 × 10⁻⁷; risk management,1.14 × 10⁻⁸; and the composite score, 4.72 × 10⁻¹². The source reports post hoc power above0.999 using the G*Power 3.1 algorithm at α = 0.05 and observed dz ≥ 0.98, and interprets thisas ruling out false-negative error.Four findings are highlighted in the interpretation of Table 5.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)64First, increases are observed across dimensions, with large standardized effects. Meanchanges range from 0.542 to 0.613 points, and all paired t tests yield p < 0.001. Effect sizesare 1.018 for market interpretation, 0.982 for investment-plan design, 1.150 for riskmanagement, and 1.555 for the composite score. Using 0.80 as the conventional benchmarkfor a large effect, the study interprets these results as substantial changes in students'perceptions of professional competence associated with the integrated teaching intervention.Second, parametric and nonparametric results converge. Reported Wilcoxon Z statisticsrange from −4.49 to −5.24, and all exact sign tests yield p < 0.001 among nonzero changes.The source interprets this agreement across mean-based and rank-based analyses as reducingconcern that a few extreme increases alone account for the findings.Third, risk management has the largest mean increase and effect size among the threedimensions: 0.613 points (95% CI [0.447, 0.779]) and dz = 1.150. The study attributes thispattern to transaction-record reviews, simulated forced-liquidation rules, and risk-focuseddebriefing, which are presented as strengthening awareness of risk and countering a preferencefor aggressive trading over defensive discipline.Fourth, qualitative follow-up addresses cases with lower self-ratings. Six students reportdeclines in investment-plan design and four in risk management. To examine changesconcealed by the group mean, the research team reports conducting post-course interviewswith these students. According to the source, all had experienced substantial accountdrawdowns or prolonged losing positions in the second simulation. They described revisingearlier confidence after recognizing its limitations. The study interprets these lower self-ratings as greater self-awareness and as evidence of the formative value of the evaluation.5.3 Quantitative Feedback on the Teaching ExperienceAt T2, the four teaching-experience items have a composite mean of 3.857 (SD = 0.625),reported as 28.5% above the neutral response value of 3. Item means are 3.929 for Q15, jointmentor guidance (SD = 0.745); 3.929 for Q16, engagement and interest in further inquiry (SD= 0.712); 3.810 for Q14, integrated simulation (SD = 0.707); and 3.762 for Q13, authenticindustry cases (SD = 0.790). The study interprets these ratings as positive responses to jointmentoring and scenario-based simulation.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)656. Discussion and Further Teaching Development6.1 From Industry Participation to Managing Cognitive LoadThe course emphasizes the pedagogical transformation of professional experience ratherthan industry lectures or the formal establishment of a practice base alone. Future developmentneeds to address two issues. First, highly realistic exercises may produce frustration andcognitive overload. For students whose self-ratings fall after drawdowns, early modules canprovide smaller, achievable tasks—for example, restricting initial hedging exercises to a singleinstrument before expanding the trading universe. Second, the case collection should beupdated jointly by academic and industry staff, supported by guided worksheets with explicitlearning objectives.6.2 Integrating Self-Reports, Performance Assessment, and Trading RecordsSelf-assessment has inherent subjective limits. The next phase of reform thereforeproposes a standardized performance rubric to strengthen the assessment framework. Table 6specifies four dimensions for use in the next teaching cycle.Table 6. Performance rubric for integrated technical-analysis tasksDimension Basicperformance (1–2)Meetsexpectations (3–4)Advancedperformance (5)EvidenceandinterpretationWeight:25%Inconsistentannotations, isolatedsignals, mechanicalformula use; cannotexplain conflictingindicators orassumptionsAccurateannotations,identification ofprimary andsecondary trends andkey levels; clearcross-indicatorreasoningComparesopposing signals,quantifies successprobabilities, andexplains indicatorsaturation and failureboundariesPlan designand reasoningWeight:25%Disconnectedtrading plan; no clearentry/exit conditions,allocationproportions, orcoherent rationaleCoherentselection rationale;explicit staged entry,adjustment, and stop-loss triggers;complete planDevelopsscenario-specificcontingencies,hedging and dynamicrebalancing; explainsstrategy capacity andtolerance for error
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)66Dimension Basicperformance (1–2)Meetsexpectations (3–4)Advancedperformance (5)Riskmanagementand disciplineWeight:30%Missing orineffective stops;repeated position-rulebreaches; excessiveconcentration; nocontingency planExplicit stop-loss/take-profitlevels with 100%adherence; positionsaligned with accountvolatility; no severeaccount deficitWithstandsblack-swan stresstests; drawdownsremain within industrystandards; detailed,verifiable riskreflectionsResearchdefense andreflectionWeight:20%Poor reportformat; inconsistentdefense; performancejudged by one return;shifts responsibilityfor errorsReport followsfinancial-sectorconventions;balanced attributionof gains/losses;structuredtransaction-basedreviewDemonstratesprofessional andcomplianceawareness; criticallyexamines cognitiveand psychologicalbiasesThe proposed procedure involves independent, double-blind assessment of anonymizedresearch reports and oral defenses by academic and industry mentors, with each assessor groupcontributing 50%. Self-assessment changes, rubric scores, and indicators exported from thesimulation system—such as maximum drawdown, Sharpe ratio, and stop-loss adherence—willthen be examined jointly in a multisource association matrix linking perceived competencewith observed behavior.6.3 Refining the Instrument and Research DesignTwo developments are planned. First, item wording and construct definition will berefined. The reference to an advantage in winning probability in Q5 will be replaced by amethodological statement about consistently applying security-selection criteria undertechnical and fundamental constraints. Second, the next academic year will introduce parallelclasses and delayed transfer tasks. A class using conventional instruction will serve as acomparison group, propensity-score matching will be used to address observed baselinedifferences, and a transfer task three months after completion will examine the retention ofanalytical and risk-management learning across time and market conditions.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)677. Limitations and Conclusions7.1 LimitationsThe study identifies the following limitations under its current research conditions.1. Sample size and causal inference. The study follows 42 students in one class. Althoughthe reported post hoc power exceeds 0.999, which the source interprets as ruling out false-negative error, a single-group pre–post design cannot fully separate the intervention frommaturation or independent learning over time. Its causal interpretation is therefore limited.2. Self-reported outcomes. The main quantitative results derive from subjective self-assessment. The source presents anonymity, separation from grading, and nonparametricchecks as measures to reduce socially desirable responding. However, the relationshipbetween perceived competence and objectively demonstrated trading-related skills stillrequires performance data collected using the rubric in Table 6.3. Low reliability in plan design at T2. Alpha for investment-plan design is 0.468 at T2.Although the study interprets this through a separation between expected success andprocedural discipline, the composite score remains psychometrically constrained and shouldbe generalized cautiously.7.2 Main ConclusionsUsing constructive alignment, the study develops an industry-integrated curriculumorganized around market interpretation, investment-plan design, and risk management, andevaluates it through semester-long paired observations of 42 students. It draws three mainconclusions.First, progressive industry-integrated teaching is presented as promoting gains inperceived professional competence. Parametric and nonparametric comparisons showincreases across the three dimensions (p < 0.001), with dz values of 0.982–1.150 and acomposite effect size of 1.555.Second, the study interprets the largest dimension-level change in risk management—0.613 points, dz = 1.150—as evidence that integrated stress scenarios and dual-mentor reviewcan strengthen attention to compliance and stop-loss discipline, countering an emphasis onreturns alone.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)68Third, the decline in plan-design reliability is interpreted as a change in professionalunderstanding. The study argues that practical training challenged the assumption thatapplying indicators mechanically guarantees stable profits and encouraged a more cautiousview of market uncertainty.Overall, the study presents a course reform approach and an assessment procedureintended to be testable and replicable. It offers a practical reference for strengtheningexperiential teaching and its evaluation in applied undergraduate finance and businessprograms.References[1] Gong, G., Shi, A., & Wang, F. (2025). Research on the construction and effectivenessof the “one-stop” practical teaching system for finance major under the background ofindustry-education integration. Advances in Social Sciences, 14(11), 429–435. (In Chinese.)https://doi.org/10.12677/ass.2025.14111015[2] Ren, L., & Zhang, F. (2026). Research and practice on the project-based practicalteaching system of financial mathematics from the perspective of industry-educationintegration. Advances in Education, 16(3), 176–182. (In Chinese.)https://doi.org/10.12677/ae.2026.163468[3] Zhang, R., & Cao, J. (2024). Research on the construction of practical teaching facultyin financial mathematics under the background of industry-education integration. CreativeEducation Studies, 12(10), 477–482. (In Chinese.) https://doi.org/10.12677/ces.2024.1210735[4] Han, W. (2026). OBE-oriented innovation and path optimization of project-basedteaching for applied undergraduate economics and management courses. Advances inEducation, 16(1), 1065–1070. (In Chinese.) https://doi.org/10.12677/ae.2026.161144[5] Chen, Y., & Xiao, M. (2026). Exploration of teaching reform in management coursesat application-oriented undergraduate universities under the OBE concept. Advances inEducation, 16(2), 1117–1123. (In Chinese.) https://doi.org/10.12677/ae.2026.162409[6] Mota, F. B., Cabral, B. P., Braga, L. A. M., & Lopes, R. M. (2025). Mapping theglobal research on project-based learning: A bibliometric and network analysis (2014–2024).Frontiers in Education, 10, Article 1522694. https://doi.org/10.3389/feduc.2025.1522694
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)69[7] Khan, I. M., Edwards, E., Menegon Ianicelli, B., Ahmed, W., Hardey, M., &Eremionkhale, G. (2025). University-industry collaboration for academic success andemployability: A connectivist perspective. Studies in Higher Education. Advance onlinepublication. https://doi.org/10.1080/03075079.2025.2545606[8] Novkovic, D., Kompen, R. T., & Petchamé, J. (2026). Active learning strategies inbusiness and information & communication technology engineering higher education: Ascoping review. Review of Education, 14(2), Article e70158.https://doi.org/10.1002/rev3.70158[9] Biggs, J. (1996). Enhancing teaching through constructive alignment. HigherEducation, 32(3), 347–364. https://doi.org/10.1007/BF00138871[10] Tavakol, M., & Dennick, R. (2011). Making sense of Cronbach’s alpha. InternationalJournal of Medical Education, 2, 53–55. https://doi.org/10.5116/ijme.4dfb.8dfd[11] Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulativescience: A practical primer for t-tests and ANOVAs. Frontiers in Psychology, 4, Article 863.https://doi.org/10.3389/fpsyg.2013.00863Appendix. Complete Course Self-Assessment QuestionnaireQ1. I can use candlestick patterns and their combinations to make a basic assessment ofbuying and selling pressures in securities markets.Q2. I can accurately draw and use trend lines, channels, and key support and resistancelevels to analyze asset-price trends.Q3. I can combine MACD, KDJ, moving averages, and other technical indicators tocross-check signals and assess market conditions.Q4. When market conditions are complex, I can independently interpret indicatordivergence or changes in chart patterns rather than apply rules mechanically.Q5. I can combine macroeconomic and industry trends with technical analysis to selectsecurities with an advantage in the probability of success.Q6. I can integrate multiple analytical tools to develop a well-supported and logicallycoherent investment decision.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)70Q7. I can formulate rigorous entry, position-increase, rebalancing, and exit strategies inresponse to changing market conditions.Q8. I can independently write a structured and well-supported technical-analysis report orinvestment plan in an appropriate professional format.Q9. I can systematically identify systemic risk, liquidity risk, and model-failure risk in aninvestment plan.Q10. I recognize that technical indicators and chart patterns can fail or produce falsesignals in range-bound markets or after unexpected shocks.Q11. I can establish disciplined position-management, stop-loss, and take-profit rulesbased on market volatility and account risk limits.Q12. In investment analysis and planning, I can observe compliance requirements andprofessional ethics rather than focus only on reported returns.Q13. Authentic, de-identified industry cases substantially improve my understanding ofthe practical limits of abstract technical-analysis concepts.Q14. Integrated investment simulations help me apply textbook knowledge to changingmarket conditions.Q15. Joint guidance from academic instructors and experienced industry mentors deepensmy understanding of occupational standards and practical requirements.Q16. Case reviews, adversarial simulations, and field-based practice strengthen myengagement with investment studies and my interest in further inquiry.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)71
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)72The Impact of AI-driven HR analytics onEmployee Retention in the Post-Pandemic Era:Evidence from Chongqing EnterprisesRong Xin, Dong Xinyi, Chen Jian*School of Management, Chongqing Institute of Engineering, Chongqing 400056,ChinaEmail：2670860352@qq.com,3582501099@qq.com,jeremymouse84@gmail.com*Corresponding Author Email: jeremymouse84@gmail.comABSTRACT:In the post-pandemic era, the global labor market has undergone profound changes, withemployee retention and organizational stability becoming core issues in enterprisemanagement. AI-driven HR analytics, through its capabilities in data monitoring, risk warning,and precise intervention, has gradually become an important tool for improving employeeretention. As a key hub for manufacturing and digital economy development in western China,Chongqing is accelerating its digital transformation. This study takes enterprises in Chongqingas the research subject to explore the mechanism through which AI-driven HR analyticsaffects employee retention in the post-pandemic context. This study suggests that AI-drivenHR analytics may enhance employees’ intention to stay and reduce turnover risk throughpathways such as turnover risk warning, person-job fit, incentive optimization, and careerdevelopment support. By constructing a digital talent retention framework, this study providespractical insights for enterprises to strengthen talent management through digital technologiesand support the stable development of regional industries.Keywords: artificial intelligence; HR analytics; employee retention; digitaltransformation; Chongqing enterprises
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)731.INTRODUCTIONIn the post-pandemic era, the structure of the labor market and employees’ work attitudeshave undergone significant changes. Employees are more concerned about stable income, jobsecurity, psychological support, career development, and work-life balance, while companiesface management pressures such as rising recruitment costs, staff turnover in key positions,organizational knowledge gaps, and declining team stability. Existing employee turnoverstudies show that whether employees stay in an organization is influenced by a combination offactors such as job satisfaction, organizational commitment, organizational support, job fit,and sense of organizational equity (Allen et al., 2010; Hom et al., 2017). Therefore, employeeretention is no longer just a traditional personnel management issue; it also relates toorganizational stability, business continuity, and long-term competitiveness.Chongqing is an important manufacturing base and a city for digital economydevelopment in western China. In recent years, Chongqing has focused on high-qualitymanufacturing development and the construction of the “33618” modern manufacturingcluster system, accelerating the digital transformation of industries such as automotive,electronic information, intelligent equipment, and software services (Chongqing MunicipalPeople’s Government, 2023). Industrial upgrading has increased enterprises' demand fortechnical, versatile, and stable talent, and has given employee retention issues greater regionalpractical significance. However, some Chongqing enterprises still rely heavily on experientialjudgment in human resource management, with shortcomings in identifying attrition risks,improving person-job fit, optimizing incentives, and supporting career development. Asemployment choices become more flexible, companies that fail to promptly identify employeeneeds and attrition risks may face loss of core talent and declining organizational stability.AI-driven HR analytics refers to companies using artificial intelligence, big data analytics,and predictive algorithms to integrate data on employee performance, attendance, training,skills, job changes, satisfaction, and career development needs to support HR decision-making.Existing research has indicated that HR analytics helps shift HR management from empiricaljudgment to data-driven decision-making, but its actual value depends on data quality,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)74analytical capabilities, organizational management context, and employees’ acceptance ofalgorithmic decisions (Angrave et al., 2016; Marler & Boudreau, 2017; Tambe et al., 2019)。Compared to traditional HR management, AI-driven HR analysis can identify turnover risksearlier, more accurately judge the degree of match between employees and positions, supportincentive decisions more effectively, and provide more targeted career development support.Based on the above background, this paper takes Chongqing enterprises as a researchcontext to explore the mechanisms by which AI-driven HR analytics influences employeeretention in the post-pandemic era. This article mainly addresses two research questions: First,how does AI-driven HR analytics affect employee retention in Chongqing enterprises? Second,through what mechanisms does AI-driven HR analytics reduce employee turnover risk andenhance employees’ intention to stay?2.CONCEPTUAL AND THEORETICAL FOUNDATIONAI-driven HR analytics is an important manifestation of digital human resourcemanagement. Marler and Boudreau (2017) believe that HR analytics is a process that improvesdecision-making quality by identifying and quantifying the relationship between HR practices,employee behavior, and organizational outcomes. Angrave et al. (2016) further pointed outthat if HR professionals lack data analysis capabilities, HR analytics may struggle to trulycreate strategic value. Therefore, AI-driven HR analytics should not be understood as a simpletechnological substitute, but rather as a decision support system where data, algorithms, HRprofessional judgment, and organizational management goals work together. In employeeretention management, AI-driven HR analytics mainly includes attrition risk prediction,employee profiling, job-to-person matching, performance analysis, training recommendations,incentive optimization, and career development support. Its core value lies in drivingenterprises from “post-event remediation” management to “pre-warning” and “preciseintervention” management.Employee retention refers to a company’s ability to reduce voluntary resignation,enhance employees’ willingness to stay, and maintain organizational stability throughmanagement measures. Employee turnover research shows that employee turnover is a
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)75complex process influenced by individual, positional, organizational, and external labormarket factors (Hom et al., 2017). In the post-pandemic era, pay levels remain important, butemployees' concerns about occupational safety, work stress, organizational support, growthopportunities, flexible work arrangements, and a sense of equity have significantly increased.Organizational support theory points out that when employees feel that the organization valuestheir contributions and cares about their well-being, they develop more positive organizationalattitudes and behaviors (Eisenberger et al., 1986). Organizational commitment theory alsoshows that employees’ emotional attachment and ongoing commitment to the organizationaffect their willingness to stay (Meyer & Allen, 1991). Therefore, employee retentionmanagement in the post-pandemic era needs to be more data-driven, personalized, andforward-looking.Conservation of Resources Theory (COR) provides a suitable perspective for explainingthe relationship between AI-driven HR analytics and employee retention. Hobfoll (1989)proposed that individuals tend to acquire, protect, and maintain valuable resources; Individualsexperience stress when resources are threatened, losses occur, or resource investments do notyield expected returns. Subsequent research further pointed out that resource loss often has astronger impact on individual psychology and behavior than resource acquisition (Hobfoll etal., 2018). In the post-pandemic era, employees may face income pressures, increasedworkloads, emotional exhaustion, career uncertainty, and limited development opportunities,leading to resource losses. Based on resource conservation theory, AI-driven HR analytics canreduce resource loss through attrition warnings, and support work, economic, psychological,and development resources through job-person matching, incentive optimization, and careerdevelopment support, thereby promoting employee retention.3.CHONGQING CONTEXT AND RETENTION CHALLENGESChongqing enterprises, especially those in manufacturing and digital economy-relatedsectors, face multiple pressures in the post-pandemic era from industrial upgrading, jobchanges, and employment stability. High-quality development and digital transformation inmanufacturing have increased enterprises’ demand for technical workers, engineering
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)76technicians, digital management talents, and versatile professionals (Chongqing MunicipalPeople's Government, 2023). Against this backdrop, employee retention not only affects thehuman resource costs of individual enterprises but also relates to regional industrial chainstability and industrial upgrading processes. Based on the background of digital transformationin Chongqing enterprises and observations from human resource management practices, thechallenges faced by employee retention in Chongqing enterprises can be summarized into thefollowing three aspects.First, the risk of key employee turnover is rising. In the post-pandemic era, employmentchoices for employees have become more flexible, and some Chongqing enterprises face anincreased risk of losing technical backbones, frontline skilled workers, and front-linemanagers. Traditional HR management often relies on exit interviews, annual satisfactionsurveys, or managerial experience to assess employee status, but these methods often lagbehind the scenes. When companies realize employees may leave, employees often havestrong intentions to leave and have even secured other job opportunities. For manufacturingand digital economy enterprises, the sudden departure of key employees may affect productioncontinuity, project progress, and organizational stability.Second, insufficient matching between people and positions. Some companies still relyon experience-based judgment during recruitment, job assignment, and job transfer processes,lacking systematic data analysis support. There may be mismatches between employees'professional abilities, skill structures, and career interests with actual job requirements.Person-job fit research shows that the degree of match between employees and their positionsaffects job satisfaction, organizational commitment, and turnover intention (Kristof-Brown etal., 2005). If employees remain in positions that do not fully match their abilities anddevelopment directions for a long time, they are prone to feelings of limited growth andinefficient fit, which lowers job satisfaction and increases the tendency to leave.Finally, insufficient motivation and career development support. Some Chongqingcompanies still use relatively unified salary, benefits, and promotion systems, making itdifficult to meet the differentiated needs of employees in the post-pandemic era. Someemployees feel that the connection between performance appraisal and actual contribution isnot close enough, that motivation methods lack flexibility, and that career development pathsare unclear. Organizational equity research shows that employees' perceptions of distributive
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)77fairness, procedural fairness, and interactional fairness influence their work attitudes andbehavioral responses (Colquitt, 2001). When employees cannot see future growthopportunities or feel the organization's support for their long-term development, theirorganizational commitment and willingness to remain tend to decline.4.MECHANISMS OF AI-DRIVEN HR ANALYTICS AFFECTINGEMPLOYEE RETENTIONAI-driven HR analysis does not directly affect employee retention, but rather improvesemployee experience and resource perception by enhancing the company's ability to identify,match, motivate, and support development, thereby influencing employees' willingness to stay.Based on resource conservation theory, this paper proposes four mechanisms of action.4.1 Turnover Risk Warning and Early InterventionAI-driven HR analytics can integrate data such as employee attendance, performancechanges, training participation, job changes, satisfaction feedback, and internalcommunication to identify potential signs of employee turnover risks. For example, recentfluctuations in employee performance, increased leave frequency, reduced participation intraining, decreased job adaptation, or decreased satisfaction can all be signs of potentialattrition risk. Research on HR analytics suggests that data analytics can help companiespredict employee behavior and improve HR decisions (Marler & Boudreau, 2017). ThroughAI analysis, companies can identify high-risk employees earlier, and HR and departmentmanagers can communicate, provide psychological support, adjust jobs, recommend training,or coach career development.This mechanism helps companies shift from passive responses to proactive retentionmanagement. Traditional employee retention management often takes remedial measures onlyafter employees resign, whereas AI-driven HR analytics can move management steps forward.From the perspective of resource conservation theory, early warning of turnover risk can helpcompanies detect employee resource loss signals earlier and reduce further resource loss
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)78through intervention (Hobfoll, 1989; Hobfoll et al., 2018)。 For Chongqing enterprises, thesudden departure of key employees may lead to job vacancies, project delays, and unstableproduction. Therefore, AI-driven attrition risk early warning helps reduce resource lossescaused by sudden talent loss and enhances the foresight of employee retention management.4.2 Person-Job FitAI can analyze employees' skills, work experience, performance records, career interests,and job competency requirements to build more accurate job-person matching models. ForChongqing enterprises undergoing digital transformation, job requirements change rapidly,and the traditional experience-based approach may struggle to adapt promptly to newtechnologies, new positions, and new business needs. AI-driven job-to-person matching canhelp companies identify gaps between employee abilities and job requirements, and offersuggestions for job transfers, rotations, or training.When employees are in positions that better match their abilities and development goals,their job adaptability, sense of achievement, and satisfaction improve. Improving job-personmatching not only enhances employee performance but also reduces frustration and resourcewaste caused by mismatched positions. Existing research shows that the degree of fit betweenindividuals and their positions affects employee attitudes and organizational outcomes(Kristof-Brown et al., 2005). From the perspective of resource conservation theory, person-jobmatching can reduce the psychological and skill resources employees continuously consumein unsuitable positions, making employees more willing to seek development opportunitieswithin the organization rather than change their situation through resignation.4.3 Incentive Optimization and Perceived FairnessAI-driven HR analytics can provide decision support for corporate compensationincentives and benefits allocation by integrating performance data, job contributions,compensation structures, employee preferences, and market compensation information. It isimportant to emphasize that AI should not directly determine employee compensation or
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)79promotion, but rather serve as an auxiliary tool to help managers better understand employeecontributions and differentiated needs. Tambe et al. (2019) pointed out that AI has thepotential to improve efficiency and predictive capabilities in human resource management, butit also faces challenges such as the complexity of HR issues, limitations in small data, fairnessresponsibilities, and negative employee reactions. Therefore, the application of AI in incentivemanagement must be combined with managerial judgment and organizational ethicsrequirements.In Chongqing enterprises, standardized incentive methods may struggle to meetemployees' differentiated needs. AI analysis can help companies design more differentiatedincentive packages, such as performance bonuses, training opportunities, flexible workarrangements, promotion support, and personalized benefits. When employees feel thatorganizational incentives are fairer, more transparent, and better suited to their needs, theirsense of organizational equity and satisfaction increase. Organizational fairness can influenceemployees' acceptance of management decisions and their work attitude (Colquitt, 2001).Therefore, AI-driven incentive optimization can enhance fairness and satisfaction, reduceresource waste caused by insufficient incentives or unfair distribution, and thereby strengthenemployee retention.4.4 Career Development Support and Organizational CommitmentAI can also be used for employee capability gap diagnosis, training courserecommendations, internal mentor matching, career path planning, and promotion preparationsupport. By analyzing employees' current skills, job requirements, and future developmentdirections, AI systems can provide personalized learning paths and career developmentsuggestions, helping employees clarify their next growth goals. Unlike traditional standardizedtraining, AI-supported career development management places greater emphasis onemployees' differentiated needs and long-term development potential.In the post-pandemic era, employees are increasingly valuing long-term careerdevelopment and personal employ ability. If companies can use AI tools to provide employeeswith clear development paths and resources for continuous growth, employees will morelikely feel valued by the organization, thereby strengthening organizational commitment.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)80Organizational commitment theory holds that employees' emotional attachment, sustainedengagement, and normative identification with the organization influence their willingness tostay (Meyer & Allen, 1991). Recent AI work scenario studies also show that employeecollaboration with AI may promote proactive behavior by reducing workload andsupplementing cognitive resources (Sun et al., 2025), and generative AI adoption mayinfluence employee outcomes through job reshaping and career commitment (Liu et al., 2025).Therefore, when AI helps employees gain clearer development paths and growth resources,they are more likely to develop stronger organizational commitment and increase theirwillingness to stay.5.DISCUSSION AND CONCLUSIONThis study develops a mechanism-based framework to explain how AI-driven HRanalytics may influence employee retention in Chongqing enterprises in the post-pandemic era.From the perspective of COR theory, the four mechanisms identified in this study-turnoverrisk warning, person-job fit, incentive optimization, and career development support-can beunderstood as processes of reducing resource loss and increasing resource gain. Turnover riskwarning helps organizations identify early signs of employee resource depletion; person-job fitreduces resource waste caused by job mismatch; incentive optimization supplements economicresources and strengthens perceived fairness; and career development support enhancesgrowth resources and organizational commitment.The analysis offers several practical implications for Chongqing enterprises. Rather thanimplementing AI-driven HR analytics in a comprehensive but fragmented manner, enterprisesmay adopt a step-by-step approach. Small and medium-sized enterprises can begin with basicemployee data integration and turnover risk warning, and then gradually expand to person-jobfit analysis, incentive optimization and career development recommendation. In addition, HRprofessionals and line managers should improve their digital capability so that AI-generatedanalytical results can be interpreted appropriately and combined with managerial judgmentand human concern. In this sense, AI-driven HR analytics should be positioned as a tool for
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)81employee support and organizational development rather than as a purely technical monitoringsystem.Ethical governance is also essential for the effective application of AI in human resourcemanagement. Employee data used in AI-driven HR analytics may include attendance,performance, behavioral signals and even psychological information. Therefore, organizationsshould clearly define the boundaries of data collection, protect employees’ rights to informedconsent and privacy, and prevent misuse of sensitive data. Moreover, algorithmic models mayreproduce or amplify historical biases related to gender, age, educational background or jobcategory. Studies on algorithmic hiring and evaluation have warned that AI systems maygenerate bias due to data selection, prediction targets and model design (Raghavan et al.,2020). Accordingly, Chongqing enterprises should establish data governance mechanisms,algorithmic fairness review procedures and human intervention channels. AI should functionas a human-centered decision-support tool rather than a substitute for managerialresponsibility.Overall, this study suggests that AI-driven HR analytics can promote employee retentionby providing work, economic, psychological and developmental resources to employees. Theproposed framework contributes to the understanding of digital talent retention in a regionalenterprise context and provides practical insights for Chongqing enterprises seeking toimprove employee retention through AI-enabled HR management. However, the conclusionsshould be interpreted within certain research boundaries. The study is mainly based onconceptual analysis and contextual observation rather than large-scale questionnaire data orlongitudinal evidence. Future research may further test the proposed framework usingempirical data from different industries and enterprise sizes. It may also examine themoderating roles of AI trust, privacy concern and perceived algorithmic fairness in therelationship between AI-driven HR analytics and employee retention.REFERENCES
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)82Allen, D. G., Bryant, P. C., & Vardaman, J. M. (2010). Retaining talent: Replacingmisconceptions with evidence-based strategies. Academy of Management Perspectives,24(2), 48–64. https://doi.org/10.5465/amp.24.2.48Angrave, D., Charlwood, A., Kirkpatrick, I., Lawrence, M., & Stuart, M. (2016). HR andanalytics: Why HR is set to fail the big data challenge. Human Resource ManagementJournal, 26(1), 1–11. https://doi.org/10.1111/1748-8583.12090Chongqing Municipal People’s Government. (2023). Action plan for promoting high-qualitydevelopment of manufacturing in new Chongqing in the new era and new journey (2023–2027) [深入推进新时代新征程新重庆制造业高质量发展行动方案(2023—2027年)].https://www.cq.gov.cn/ywdt/jrcq/202309/t20230928_12390505.htmlColquitt, J. A. (2001). On the dimensionality of organizational justice: A construct validationof a measure. Journal of Applied Psychology, 86(3), 386–400.https://doi.org/10.1037/0021-9010.86.3.386Eisenberger, R., Huntington, R., Hutchison, S., & Sowa, D. (1986). Perceived organizationalsupport. Journal of Applied Psychology, 71(3), 500–507. https://doi.org/10.1037/0021-9010.71.3.500Hobfoll, S. E. (1989). Conservation of resources: A new attempt at conceptualizing stress.American Psychologist, 44(3), 513–524. https://doi.org/10.1037/0003-066X.44.3.513Hobfoll, S. E., Halbesleben, J., Neveu, J.-P., & Westman, M. (2018). Conservation ofresources in the organizational context: The reality of resources and their consequences.Annual Review of Organizational Psychology and Organizational Behavior, 5, 103–128.https://doi.org/10.1146/annurev-orgpsych-032117-104640Hom, P. W., Lee, T. W., Shaw, J. D., & Hausknecht, J. P. (2017). One hundred years ofemployee turnover theory and research. Journal of Applied Psychology, 102(3), 530–545.https://doi.org/10.1037/apl0000103Kristof-Brown, A. L., Zimmerman, R. D., & Johnson, E. C. (2005). Consequences ofindividuals’ fit at work: A meta-analysis of person-job, person-organization, person-group, and person-supervisor fit. Personnel Psychology, 58(2), 281–342.https://doi.org/10.1111/j.1744-6570.2005.00672.x
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)83Liu, Y., Sheng, F., & Liu, R. (2025). Generative AI adoption and employee outcomes: Aconservation of resources perspective on job crafting, career commitment, and themoderating role of liking of AI. Humanities and Social Sciences Communications, 12,Article 1376. https://doi.org/10.1057/s41599-025-05656-4Marler, J. H., & Boudreau, J. W. (2017). An evidence-based review of HR analytics. TheInternational Journal of Human Resource Management, 28(1), 3–26.https://doi.org/10.1080/09585192.2016.1244699Meyer, J. P., & Allen, N. J. (1991). A three-component conceptualization of organizationalcommitment. Human Resource Management Review, 1(1), 61–89.https://doi.org/10.1016/1053-4822(91)90011-ZRaghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmichiring: Evaluating claims and practices. In Proceedings of the 2020 Conference onFairness, Accountability, and Transparency (pp. 469–481). Association for ComputingMachinery. https://doi.org/10.1145/3351095.3372828Sha, C., & Chai, T. (2025). When digital-AI transformation sparks adaptation: Job craftingand AI knowledge in job insecurity contexts. Frontiers in Psychology, 16, Article1612245. https://doi.org/10.3389/fpsyg.2025.1612245Sun, C., Zhao, X., Guo, B., & Chen, N. (2025). Will employee–AI collaboration enhanceemployees’ proactive behavior? A study based on the conservation of resources theory.Behavioral Sciences, 15(5), Article 648. https://doi.org/10.3390/bs15050648Tambe, P., Cappelli, P., & Yakubovich, V. (2019). Artificial intelligence in human resourcesmanagement: Challenges and a path forward. California Management Review, 61(4), 15–42. https://doi.org/10.1177/0008125619867910
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)84Explore, Execute, and Express: The 3XEducation Theory for Rural Teacher Growth inthe Age of Artificial IntelligenceYin BingwenWuhan Chenfeng Education Consulting & Development CenterAbstractGenerative artificial intelligence is often discussed through the questions of whether AIwill replace teachers and whether students should be permitted to use AI. These questionsoverlook a more fundamental issue: how should education be reorganized when informationand polished production become widely available? This article interprets Zhao Geng's 3XEducation Theory—Explore, Execute, and Express—within the context of rural teacherdevelopment and AI-enabled education. The theory is the core lesson-creation model of theXingchen School and special-post teacher training system, and it operates alongside StageEducation Theory and the CMC model of Culture, Management, and Curriculum. Exploremeans discovering educational problems, students, and oneself in authentic rural schoolcontexts. Execute means translating a discovered problem into low-cost classroom action,iterative experimentation, and practical improvement. Express means presenting, explaining,reflecting on, and sharing educational practice so that personal experience becomes publicknowledge. The article argues that AI should not replace this developmental chain. It shouldstrengthen the teacher's ability to discover meaningful problems, support but not substitute forsituated implementation, and expand opportunities for credible presentation. In the AI era, thescarce educational resources are not merely information or output, but attention to realproblems, sustained action, situated judgment, and accountable expression.Keywords: 3X Education Theory; rural teacher development; artificial intelligence;Explore; Execute; Express; teacher education; authentic assessment
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)85Introduction: Reframing the QuestionPublic discussion of artificial intelligence in education frequently asks whether AI willreplace teachers or whether students should be allowed to use generative AI in assignments.Such questions position AI as an external tool that schools must either accept or restrict. Amore productive question is how education should redesign the learning and professional-growth process when access to information and the generation of polished outputs are nolonger scarce.This question is particularly important for rural teacher development. Rural teachers oftenwork under conditions of limited staffing, uneven access to professional learning, constrainedmaterial resources, and highly diverse student needs. Their most important professionalknowledge is therefore not simply imported content. It is the capacity to identify problems in aparticular school, design feasible responses, act under real constraints, and communicate whathas been learned to others.The 3X Education Theory developed by Zhao Geng provides a useful framework for thischallenge. In the Xingchen School and special-post teacher training system, 3X refers toExplore, Execute, and Express. The framework is connected with Stage Education Theory,which understands every person as the protagonist of a meaningful stage, and with the CMCmodel of Culture, Management, and Curriculum. Together, these ideas form a practice-oriented model for rural teacher growth: teachers discover educational problems, implementresponses, and make their experience visible through public expression.This article does not treat 3X as a synonym for generic project-based learning. Itsdistinctive contribution is the connection between teacher professional development, ruraleducational context, low-cost innovation, and public presentation. The argument is that AIshould be integrated into the 3X chain while preserving the human responsibilities at its center.The 3X Education Theory in ContextThe 3X model can be summarized as a developmental loop: discover a real problem, acton it in practice, and express the result in a form that can be examined and shared. The three
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)86stages cultivate three corresponding capabilities: discovery power, execution power, andexpression power.Stage Meaning Core question ProfessionalcapabilityExplore Discovereducationalproblems, students,and self in context.What is reallyhappening, and whatis worthinvestigating?DiscoverypowerExecute Design,implement, test, andrevise a feasibleresponse.What can bedone here, with thesestudents andresources?Execution powerExpress Present, explain,reflect on, and sharethe practice.What waslearned, and how canothers understand oruse it?ExpressionpowerThe framework is cyclical rather than linear. Expression can reveal that the originalproblem was misunderstood; execution can generate a more precise question; and a newexploration can begin after reflection. The goal is not to produce a single successful lesson butto establish a habit of professional inquiry.Explore: Discovering Problems, Students, and SelfIn the first X, teachers do not begin with a centrally assigned course task. They beginwith the lived reality of a rural school. Exploration may involve noticing that students avoidspeaking in class, that a local cultural resource is absent from the curriculum, that parents andschool expectations conflict, or that a familiar teaching method is not working for a particulargroup of learners. The teacher converts an everyday difficulty into a professional question.This orientation is consistent with inquiry-based learning research, which identifiesquestioning as a central component of knowledge construction rather than a preliminaryactivity that can be skipped (Pedaste et al., 2015; Tawfik et al., 2020). In the 3X model,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)87however, the object of inquiry is not only subject-matter content. It includes the classroom, thestudents, the community, and the teacher’s own assumptions.AI can support this stage by helping teachers organize observations, compare possibleexplanations, generate alternative questions, and locate relevant pedagogical research. It canfunction as a conversational partner that helps a teacher move from “students are not engaged”to a more precise question such as “Under what classroom conditions do students who rarelyspeak voluntarily participate?” Yet AI cannot determine whether a question matters in aparticular community. That judgment belongs to the teacher, who has responsibility for thestudents and the context.For rural teacher training, Explore should therefore include observation journals, studentconversations, community mapping, problem-definition workshops, and question-revisionactivities. A useful assessment should reward the quality of the problem formulation ratherthan merely the quantity of information collected. Teachers should be able to explain whatthey noticed, why it matters, what assumptions they are making, and what evidence wouldchange their interpretation.Execute: Turning Ideas into Situated ActionThe second X is the practical center of the model. Teachers design a response, bring itinto the classroom, observe what happens, and revise the response. Execution is not themechanical application of a pre-existing method. It is situated experimentation under realconstraints.This emphasis is especially important in rural schools, where innovation cannot dependon expensive equipment or ideal staffing conditions. Low-cost, locally meaningful activities—such as using agricultural materials, community stories, local geography, or cooperativegames—may create powerful learning opportunities. A practice such as a “Straw Olympics,”for example, can connect physical construction, collaboration, problem-solving, and localresource use without requiring sophisticated infrastructure. Its value lies not in novelty alonebut in the teacher’s ability to identify a problem, design a response, observe student learning,and improve the activity.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)88AI can assist execution by proposing lesson variations, generating differentiated materials,simulating student questions, or helping teachers analyze reflection notes. AI literacy researchemphasizes that responsible use includes not only operating AI systems but also evaluatingtheir outputs and navigating their ethical implications (Ng et al., 2021; Chiu et al., 2024).Teachers should therefore treat AI suggestions as provisional resources, not authoritativelesson plans.The key evidence of Execute is the trace of action: the original plan, classroomadaptation, student response, revision, and explanation of why a change was made. Thisapproach is preferable to evaluating only a polished lesson plan. A lesson that was revised inresponse to students may demonstrate more professional learning than a flawless plan that wasnever tested.Express: Making Practice PublicThe third X is Express. Teachers present their educational stories, explain their decisions,reflect on failures and successes, and share their practices with colleagues. Expressiontransforms private experience into public professional knowledge.This stage is closely connected with Stage Education Theory. If every person isunderstood as the protagonist of a meaningful stage, teacher education cannot positionteachers only as passive listeners. Teachers must also become presenters, storytellers,demonstrators, and contributors to a professional community. The stage is not merely aperformance space; it is a space in which experience becomes visible, discussable, andimprovable.In AI-supported education, Express also addresses the problem of credibility. A polishedpresentation can be generated by AI, but a teacher who can explain the original problem,describe classroom events, justify changes, respond to questions, and identify limitationsdemonstrates a deeper form of professional understanding. Presentation should thereforeinclude live explanation, peer questioning, classroom artifacts, student work, reflectivenarration, and evidence of iteration.Express should not be reduced to public-relations display. A credible presentation mustinclude uncertainty and failure. Teachers should be encouraged to say what did not work,
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)89which students were not reached, what evidence remains incomplete, and what they wouldchange next time. In this form, expression becomes a mechanism of professionalaccountability rather than self-promotion.CMC as the Enabling EnvironmentThe 3X chain requires an institutional environment. The CMC model—Culture,Management, and Curriculum—can be understood as the supporting architecture that makesExplore, Execute, and Express sustainable.Culture determines whether teachers are permitted to ask inconvenient questions andshare incomplete work. A culture of trust is necessary if teachers are expected to disclosefailure rather than present only success. Management determines whether time, collaboration,mentoring, and recognition are available for iterative practice. Curriculum determines whetherteacher learning is connected to real classroom problems instead of being isolated frompractice.AI integration should therefore be governed by the same three dimensions. Culturally,schools should promote critical and ethical AI use rather than blind enthusiasm or blanketprohibition. Managerially, training systems should create time for teachers to document anddiscuss AI-supported experiments. Curricularly, AI should appear as one resource within aproblem-centered professional learning cycle, not as the curriculum itself.Reconceptualizing Teacher Education in the AI EraThe 3X model suggests four changes in teacher education. First, professional learningshould begin with local problems rather than universal solutions. Research on AI in educationshows that technological tools can support personalization and efficiency, but educationalvalue depends on pedagogical design and teacher judgment (Labadze et al., 2023; Zawacki-Richter et al., 2019).Second, teacher education should evaluate process evidence. A portfolio might include aproblem statement, observation notes, AI-use disclosure, lesson design, classroom artifacts,student feedback, revision history, and a final presentation. This resembles authentic
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)90assessment, which emphasizes meaningful application and performance rather than isolatedrecall (Sundari, 2024).Third, teacher education should make expression a professional competency. Teachersneed to communicate with students, parents, colleagues, administrators, and communities.They must explain not only what they did but why it was appropriate, what evidence supportsit, and what should happen next. Expression is therefore a form of leadership and knowledgecirculation.Fourth, AI policy should focus on responsibility rather than prohibition. Teachers shoulddisclose when AI was used, verify claims and sources, protect student privacy, and retainresponsibility for pedagogical decisions. AI can accelerate drafting and analysis, but it cannotassume responsibility for the consequences of a lesson in a real classroom.ConclusionThe 3X Education Theory offers a grounded way to rethink education in the age ofartificial intelligence. Explore, Execute, and Express describe more than three instructionalsteps. They describe a complete form of professional growth: seeing reality carefully, actingresponsibly within constraints, and converting experience into shared knowledge.For rural teachers, this model is particularly valuable because it begins with the actualschool rather than an abstract ideal school. It recognizes local knowledge, limited resources,student diversity, and the teacher's responsibility to create feasible change. AI can strengtheneach stage by supporting question formation, expanding design possibilities, organizingevidence, and assisting communication. Yet it should remain subordinate to the teacher'ssituated judgment.The future of education will not be defined by the absence of AI or by the dominance ofAI. It will be defined by whether educational systems can protect the human learning chainthat AI cannot authentically complete: discovering a problem that matters, carrying outsustained action in a real community, and presenting an accountable account of what has beenlearned. In this sense, 3X is not only a model for teacher training. It is a proposition about thefuture form of education itself.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)91ReferencesCasal-Otero, L., Catala, A., Fernández-Morante, C., Taboada, M., Cebreiro, B., &Barro, S. (2023). AI literacy in K-12: A systematic literature review. International Journalof STEM Education, 10, Article 29. https://doi.org/10.1186/s40594-023-00418-7Chiu, T. K. F., Ahmad, Z., Ismailov, M., & Sanusi, I. T. (2024). A systematic reviewof AI literacy conceptualization, constructs, and implementation and assessment efforts(2019–2023). Computers and Education Open, 6, Article 100173.https://doi.org/10.1016/j.caeo.2024.100173Labadze, L., Grigolia, M., & Machaidze, L. (2023). Role of AI chatbots in education:Systematic literature review. International Journal of Educational Technology in HigherEducation, 20, Article 56. https://doi.org/10.1186/s41239-023-00426-1Ng, D. T. K., Leung, J. K. L., Chu, S. K. W., & Qiao, M. S. (2021). ConceptualizingAI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2,Article 100041. https://doi.org/10.1016/j.caeai.2021.100041Pedaste, M., Mäeots, M., Siiman, L. A., de Jong, T., van Riesen, S. A. N., Kamp, E.T., Manoli, C. C., Zacharia, Z. C., & Tsourlidaki, E. (2015). Phases of inquiry-basedlearning: Definitions and the inquiry cycle. Educational Research Review, 14, 47–61.Sundari, S. (2024). A systematic literature review on authentic assessment in highereducation: Best practices for the development of 21st century skills, and policyconsiderations. Studies in Educational Evaluation, 82, Article 101425. 5Tawfik, A. A., Graesser, A., Gatewood, J., & Gishbaugher, J. (2020). Role ofquestions in inquiry-based instruction: Towards a design taxonomy for question-askingand implications for design. Educational Technology Research and Development, 68,653–678. https://doi.org/10.1007/s11423-020-09738-9Zawacki-Richter, O., Marín, V. I., Bond, M., & Gouverneur, F. (2019). Systematicreview of research on artificial intelligence applications in higher education—Where arethe educators? International Journal of Educational Technology in Higher Education, 16,Article 39. https://doi.org/10.1186/s41239-019-0171-0
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)92Testing Whether Economic DiversificationTruly Improves the Employment Structure andResident Income In MacauHe WenyuanBased on Macao's 2015-2024 Sectoral Value Added and 2015-2026 Quarterly LabourMarket DataB242514@utm.edu.moMacao University of TourismAbstractThis study asks whether economic diversification in Macao has translated into a better employment structureand higher resident income. Diversification is evaluated at three distinct levels: output composition, employmentallocation, and income outcomes. Official sectoral value added, labour-force statistics, and monthly gross gamingrevenue are used to construct an eight-sector annual panel for 2015–2024 and a quarterly gaming sample for2015Q1–2026Q2. The analysis combines concentration measures, a Shapley decomposition of the non-gamingshare, two-way fixed-effects regressions, and quarterly models with Newey–West standard errors.The evidence indicates that Macao has established a positive foundation for moderate economicdiversification. The non-gaming share rose from 52.10% in 2015 to 56.65% in 2024, moving closer to thegovernment's target of around 60% by 2030, while nominal non-gaming value added increased by 17.07%. Thedeclining gaming employment share and employment growth in finance, transport, real estate, and businessservices also point to gradual adjustment in labour allocation. Sectoral output is positively associated withemployment, although the strength of transmission differs across industries. Nominal median wages increased,while price-adjusted and distributional outcomes require additional indicators. These findings support theestablished policy direction and suggest that supplementary employment, real-wage, and job-stability indicatorscan help refine implementation and evaluation.Keywords: Macao economic diversification; employment structure; resident income; gamingindustry
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)93I. IntroductionMacao has long relied on gaming and tourism as its principal export sector. The fiscal revenue,external demand, and wage premium generated by that industry have provided an unusual capacity forgrowth in a small open economy. At the same time, the sector's dominance in output, public finance, and thelabour market has made the economy more vulnerable to regulation, external demand, and public-healthshocks. The Economic Development Plan for Appropriate Diversification of the Macao SpecialAdministrative Region (2024-2028) places diversification at the centre of policy and uses a benchmark ofaround 60% for non-gaming value added by 2030. The 2026 Policy Address states that the Government willformulate and publish a third five-year plan; it is not treated here as a released plan or as the source of anempirical target. The 60% benchmark is observable and measurable, but it remains an output-structureindicator and does not automatically imply improvements in job quality or resident income.This study focuses on the composition of the increase in the non-gaming share and the specific channelsthrough which it is transmitted to employment and income. The expansion of non-gaming industries, higherlabour productivity, and more jobs for local residents jointly constitute the process through which appropriateeconomic diversification is converted from a structural objective into livelihood benefits. Decomposing thesources of share changes is not intended to question the established target; rather, it provides a more detailedbasis for judging growth drivers at different stages, allocating policy resources, and improving the pace ofimplementation.This study makes three contributions. First, it places output diversification, employment diversification, andincome improvement in a single analytical framework, providing supplementary employment and livelihoodevidence for the established policy objective. Second, it uses a symmetric Shapley decomposition to identify therelative contributions of non-gaming expansion and changes in the gaming cycle, offering a tool forimplementing agencies to assess the drivers of structural transformation. Third, it matches official sectoral value-added and labour-market data into a panel and makes all processing code available to facilitate verification ofstatistical definitions and continuous updating. The econometric results are interpreted cautiously; the purpose isto strengthen the reproducibility of policy research.II. Theoretical Foundations and Research Propositions(1) Structural Transformation, Specialisation, and DiversificationDevelopment economics generally understands structural transformation as the reallocation of productivefactors from low-productivity activities to high-productivity activities. Imbs and Wacziarg (2003) note thateconomies at low and middle income levels often first experience industrial diversification and may re-specialise
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)94only after reaching higher incomes. Diversification therefore does not mean that more industries are always better;it should reduce volatility risks arising from excessive concentration and enable new industries to becomesustainable sources of productivity. For small economies with constrained land and labour supplies, economies ofscale and comparative advantage naturally encourage specialisation; the policy objective should beappropriateness rather than an even distribution of resources.The share of industrial value added is affected simultaneously by non-gaming development, the cycle of thedominant industry, and relative-price changes. Let N denote non-gaming value added and G gaming value added;the non-gaming share is S=N/(N+G). An increase in S may result from an increase in N or a stage-specificchange in G. Decomposing these two contributions helps the government distinguish structural progress fromcyclical fluctuation, maintain policy continuity as the external environment changes, and identify more accuratelythe growth drivers requiring priority cultivation.(2) Resource Mobility, Wage Premiums, and the Employment StructureCorden and Neary's (1982) booming-sector model shows that a sector with strong external demand canattract resources through wages and factor prices and affect factor allocation in other tradable sectors. The highprofitability and institutional advantages of Macao's gaming industry provide a practical basis for this resource-movement effect. The current plan cultivates industries including modern finance, conventions and exhibitions,technology, traditional Chinese medicine and big health, and culture and sports, thereby offering more choices forlabour allocation. When these sectors gradually develop productivity and market advantages, economicdiversification can further manifest itself in more non-gaming jobs and higher wages.The transmission between wages and sectoral output has multiple layers. The inter-industry wagedifferentials discussed by Krueger and Summers (1988) show that human capital, organisational rents, bargaininginstitutions, and job structures all affect worker remuneration. Thus, growth in non-gaming value added may firstappear as firm accumulation, technological investment, and higher labour productivity, and only graduallytranslate into employment and wage growth. On this basis, the study proposes three testable propositions: first,appropriate economic diversification should attend both to the absolute expansion of non-gaming value addedand to an increase in its share; second, a better employment structure is reflected in the expansion of non-gamingjobs and lower employment concentration; third, higher resident income requires industrial development, humancapital, and wage-setting mechanisms to work together.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)95III. Data Sources, Variables, and Research Methods(1) Data Sources and Statistical DefinitionsThis revision uses the archived source extracts supplied with the study, rather than reconstructed orillustrative figures. The three core CSV files preserve the official source URL, download date, page title,indicator ID, unit, and population definition in their metadata: DSEC sectoral GDP value added (2010-2024), DSEC industry employment and median monthly employment income (2015Q1-2026Q2), and DICJmonthly gross gaming revenue (2010-01 to 2026-08). Sectoral value added uses DSEC current-priceproducer-price data; employment, unemployment, and median monthly employment income use DSEClabour-force data; and gross gaming revenue uses DICJ data. The diversification benchmark is taken fromthe Economic Development Plan of the Macao Special Administrative Region (2024-2028). The 2026Policy Address is used only as evidence that the Government proposed to formulate and publish a third five-year plan.Table 1. Sources of Data and Policy Documents Used in This Study (Three-Line Table)Institution Data/document Official URL Frequency andperiod Use in this studyDSECGross domesticproduct: value added byindustrywww.dsec.gov.mo/en-US/Statistic?id=901 Annual, 2010-2024 Non-gaming share,HHI, sectoral outputDSEC Employment survey www.dsec.gov.mo/en-US/Statistic?id=103Quarterly, 2015Q1-2026Q2Employment,unemployment, wages oflocal residentsDSEC Time SeriesDatabase (Web Service)www.dsec.gov.mo/TimeSeriesDatabase.asmx?WSDLQuarterly/monthlySupplementaryverification and seriespreparationDICJ Monthly statisticsof gross gaming revenuewww.dicj.gov.mo/web/en/information/DadosEstat_mensal/Monthly, 2010M1-2026M8Gaming-cycleexplanatory variableDSEPDREconomicDevelopment Plan forAppropriateDiversification (2024-2028)www.dsepdr.gov.mo/zh-hans/event/PDDAE/PDDAEPolicy documentCurrent policyframework fordiversificationMacao SARGovernmentPolicy Address forthe Fiscal Year 2026www.policyaddress.gov.mo/data/policyAddress/2026/en/2026_policy_e2.pdfPolicy documentPolicy context only:proposed formulationand publication of a thirdfive-year planNote: DSEC denotes the Statistics and Census Service and DICJ denotes the Gaming Inspection andCoordination Bureau. The first row of each archived CSV preserves the official URL, access date, page title,indicator ID, unit, and population definition. Tables 2-5 were regenerated from these files by theaccompanying Python script. Value added is reported in MOP millions; employment is reported inthousands of persons; and wages are reported in MOP per month unless otherwise stated.The industry panel covers eight industries that can be matched relatively consistently across the value-added,employment, and wage statistics: gaming; construction; wholesale and retail trade; hotels and similar activities;
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)96restaurants and similar activities; transport, storage, and communications; financial activities; and real estate andbusiness activities. Value added in financial activities is the sum of banking and insurance; real estate andbusiness activities combine real estate, rental, and business services. Industry aggregation improvescomparability of statistical definitions but compresses within-industry heterogeneity. When quarterly data areannualised, both employment and median wages are calculated as the arithmetic mean of the four quarterlyobservations; the annual wage here is not the median of individual annual incomes.(2) Variable ConstructionNon-gaming value added equals the sum of the value added of all listed industries minus gaming valueadded. The non-gaming share and output concentration are defined as follows:S = N /(N + G )； HHI = Σⱼ sⱼ ²where s_jt is the share of industry j in value added in year t. A lower HHI indicates a more disperseddistribution of output. The employment structure is measured by gaming's employment share, non-gamingemployment, and changes in industry employment. The principal income indicator is median monthlyemployment income of local employed residents. Because the public summary tables do not provide a price-adjustment series and an individual income distribution fully consistent with this study's industry panel, this studyreports only nominal wage changes and does not interpret them directly as changes in real purchasing power orwelfare.(3) Decomposition and Econometric SpecificationTo present accurately the joint effects of non-gaming expansion and gaming-cycle changes on the structuralindicator, this study applies a two-factor Shapley decomposition to the endpoint change in S=N/(N+G). Themethod calculates two paths separately: changing N first and then G, and changing G first and then N; it thenaverages the marginal contributions across the two paths. This approach ensures that the two contributions sumexactly to the total change in the non-gaming share and provides a transparent, symmetric analytical definitionfor annual monitoring during plan implementation.The baseline model for the industry panel is:ln(Yⱼ ) = β ln(VAⱼ ) + αⱼ + λ + εⱼY denotes, respectively, industry employment and the median monthly employment income of localemployed residents; VA denotes industry nominal value added; and alpha_j and lambda_t are industry and yearfixed effects. Standard errors are clustered by industry. Because there are only eight industry clusters, statisticalsignificance is used mainly as supplementary information; interpretation focuses on coefficient direction,magnitude, and consistency with descriptive statistics. The quarterly gaming model aggregates monthly gross
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)97gaming revenue to quarters and controls for a linear trend, seasonal effects, pandemic-period and normalisation-period dummy variables, using Newey-West standard errors with four lags (Newey & West, 1987). These modelsidentify conditional associations among variables and provide empirical reference for policy monitoring.IV. The Degree of Diversification in Macao's Output StructureTable 2. Output, Employment, and Income Indicators in Selected YearsYearNon-gaming valueadded(MOP million)Non-gaming share(%)OutputHHIGamingemploymentshare(%)Overallnominal wage(MOP/month)Unemployment rate(%)2015 186,081 52.10 0.257 21.06 18,000 1.822019 212,779 48.88 0.285 21.99 20,000 1.732022 171,723 85.20 0.100 18.94 20,125 3.672023 218,180 61.59 0.184 18.70 20,375 2.652024 217,842 56.65 0.218 18.92 21,475 1.80Note: Value added is at current prices; employment and wages are arithmetic averages of quarterly observations in the year.The non-gaming share in 2022 was affected by exceptional external conditions. This study treats it as a special observation ofeconomic resilience and evaluates structural change using a longer-term trend. Source: Table 1.From 2015 to 2024, non-gaming value added rose from MOP 186.081 billion to MOP 217.842 billion, anominal increase of 17.07%; gaming value added adjusted from MOP 171.107 billion to MOP 166.695 billion.The non-gaming share rose from 52.10% to 56.65%, while output HHI declined from 0.257 to 0.218. The long-term endpoint comparison indicates that Macao's output structure has continued to move in the direction ofappropriate diversification. With 2024 as the benchmark, the non-gaming share remained approximately 3.35percentage points below the around-60% target for 2030; current progress provides a sound starting point forcontinued cultivation of non-gaming industries during the subsequent planning period. Because the share dependson the growth paths of both gaming and non-gaming industries, policy implementation should dynamicallymonitor the levels and growth rates of both.In 2022, the non-gaming share reached 85.20%, reflecting the substantial shock to gaming-tourism activityduring the exceptional period; with economic normalisation, it adjusted to 61.59% in 2023 and 56.65% in 2024.This process shows both Macao's strong capacity for recovery and the influence of the external environment andthe dominant industry's cycle on annual shares. Placing the 2030 target in a medium- and long-term planningframework enables structural transformation to be understood from a cross-cycle perspective rather than beingdisturbed by fluctuations in a single year.Table 3. Shapley Decomposition of Changes in the Non-Gaming Value-Added SharePeriod Starting point(%)Ending point(%)Change(percentage points)Contributionof non-gamingexpansion(percentage points)Contributionof gaming change(percentage points)2015-2024 52.096 56.650 4.554 3.907 0.6472019-2024 48.880 56.650 7.771 0.583 7.188
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)98Note: Positive values indicate a contribution to an increase in the non-gaming share. The Shapley decomposition averages thetwo orders of change; the two contributions sum to the total change.The decomposition results further confirm the long-term effects of policy implementation. From 2015 to2024, the share rose by 4.554 percentage points, of which 3.907 percentage points came from growth in non-gaming value added and 0.647 percentage points from changes in gaming value added. This indicates that non-gaming expansion was the main force behind structural improvement over the longer period. Taking the pre-pandemic year 2019 as the starting point, the share in 2024 was 7.771 percentage points higher, of which non-gaming value-added growth contributed 0.583 percentage points and the pace of gaming normalisationcontributed 7.188 percentage points. The two results are not contradictory: the former shows that long-termcultivation of non-gaming industries has accumulated progress, whereas the latter shows that the short- andmedium-term structural indicator remains affected by the pace of normalisation. Continued implementation of theestablished industrial plan and cross-cycle monitoring can therefore make the approximately 60% target rest moresteadily on growth in non-gaming industries.V. Does Economic Diversification Improve the Employment Structure?Table 4. Changes in Value Added, Employment, and Nominal Wages of Local Residents in EightIndustries (2015-2024)IndustryChange invalue added(%)Change inemployment(%)Change inwages(%)2024employment(thousand persons)2024 wage(MOP/month)Gaming -2.6 -14.7 11.8 71.2 21,250Construction -44.8 -52.5 24.9 26.0 18,825Wholesale andretail trade 19.0 -2.1 21.8 44.0 15,525Hotels andsimilar activities 68.3 4.3 27.5 30.0 19,125Restaurantsand similaractivities37.9 -18.7 14.3 21.3 12,375Transport,storage, andcommunications11.5 13.3 20.4 19.8 18,000Financialactivities 36.0 40.2 26.0 15.1 22,775Real estate andbusiness activities 11.5 7.0 20.2 31.8 15,000Note: Wages are quarterly averages of median monthly employment income of local employed residents; employment covers allemployed persons, so the two statistics do not refer to the same population. Industry value added is nominal.Overall, gaming employment adjusted from 83,500 persons in 2015 to 71,200 in 2024, while non-gamingemployment adjusted from 313,000 to 305,100; gaming's employment share fell from 21.06% to 18.92%. Thischange indicates that labour allocation has begun to exhibit a more diversified structure. Because the sampleperiod includes the pandemic and economic normalisation, recovery in total employment typically lags recoveryin output. At this stage, employment outcomes are more appropriately understood as a transitional manifestation.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)99As priority non-gaming projects are implemented and local talent development is strengthened, the employmentstructure can be expected to absorb further improvements in the output structure.Industry differences further show that diversification is not a homogeneous process. From 2015 to 2024,financial activities recorded nominal value-added growth of 35.98% and employment growth of 40.23%, makingit a prominent sector in which output and employment expanded together; employment in transport, storage, andcommunications grew by 13.31%, that in real estate and business activities by 6.98%, and that in hotels by 4.35%.By contrast, value added in restaurants increased by 37.85% while employment fell by 18.68%; value added inwholesale and retail trade increased by 19.00% while employment fell by 2.06%. These results may reflect higherprices, labour-productivity growth, enterprise restructuring, or labour substitution, but aggregated data cannotdistinguish among them. The most important policy question is not whether an industry is labelled non-gaming,but whether growth is converted into jobs that local residents can enter, sustain, and advance within.Changes in value added and employment in construction have clear investment-cycle characteristics. Hoteloutput expanded while employment grew relatively slowly, suggesting that the integrated tourism and leisureindustry may have raised output per unit of labour during normalisation. Different speeds of adjustment acrossindustries demonstrate the need for the government to adopt differentiated policies according to industrycharacteristics. If information on occupation, hours worked, contract duration, and the shares of local and non-local employees were further combined, the implementation effects of talent and employment policies could beassessed more precisely.VI. Does Economic Diversification Improve Resident Income?Calculated as the annual average of quarterly medians, overall monthly employment income of localemployed residents increased from MOP 18,000 in 2015 to MOP 21,475 in 2024, a nominal increase of 19.31%.Median wages in the eight matched industries also generally rose, with relatively large cumulative nominalincreases in hotels, finance, and construction. These data show that, during structural adjustment andnormalisation, nominal labour remuneration of residents maintained an overall upward trend, providing a positivebasis for appropriate economic diversification to improve livelihoods.To show the livelihood effects of policy more fully, existing nominal-wage statistics may be complementedby prices, housing and service costs, wage quantiles, and household disposable-income indicators. Industry wagestatistics are limited to local employed residents, whereas employment includes all employed persons; annualvalues are also arithmetic averages of quarterly medians. Stating these statistical definitions clearly helps topresent more accurately in the future the benefits of plan implementation for different resident groups, withoutchanging the basic fact that nominal wages have continued to rise.The gaming wage premium also changed. In 2015, the median wage in gaming was about 5.56% above theoverall median; it was roughly equal in 2019 and about 1.05% below in 2024. This change may result from catch-up in overall wages, shifts in the composition of gaming jobs, or employment adjustments during the recovery
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)100period. At minimum, it indicates that the previous pattern in which gaming's high wages attracted labour hasweakened. However, if substitute jobs are concentrated in low-paid restaurants, retail, or unstable service work,diversification may improve the output structure without improving the quality of income.VII. Econometric Results, Robustness, and Identification BoundariesTable 5. Core Results of the Industry Panel and Quarterly Gaming ModelsSampleDependentvariableExplanatoryvariableCoefficient SE p R-squared N Standard errorsEight-industryannualpanel, 2015-2024Log ofindustryemploymentLog ofindustrynominalvalue added0.134 0.065 0.039 0.971 80Clustered byindustry; 8clustersEight-industryannualpanel, 2015-2024Log ofindustrymedianwageLog ofindustrynominalvalue added0.003 0.015 0.866 0.974 80Clustered byindustry; 8clustersQuarterly gamingtime series,2015Q1-2026Q2Log ofgamingemploymentLog ofgrossgamingrevenue0.051 0.038 0.178 0.789 46Newey-West; 4lagsQuarterly gamingtime series,2015Q1-2026Q2Log ofgamingmedianwageLog ofgrossgamingrevenue0.039 0.022 0.082 0.727 46Newey-West; 4lagsNote: Industry models include industry and year fixed effects; quarterly models include controls for trend, season, pandemicperiod, and normalisation period. Coefficients are log-log elasticities. With only eight industry clusters, statistical significance shouldbe interpreted cautiously.The annual panel of eight industries shows that a 1% increase in industry nominal value added is associated,on average, with an approximately 0.134% increase in industry employment; the coefficient is statisticallysignificant at the conventional 5% level. This positive relationship indicates that industrial development canpromote employment expansion. An employment elasticity below one means that the translation of output intojobs also depends on technological structure, firm size, and skill matching. The elasticity in the industry wagemodel is 0.003 with a p-value of 0.866, indicating that wage growth is influenced more by economy-wide wagetrends, labour supply and demand, human capital, and institutional arrangements. This result supports thegovernment's simultaneous advancement of talent development, employment services, and labour protectionalongside industrial policy.In the quarterly gaming model, the conditional elasticity of gross gaming revenue with respect to gamingemployment is 0.051 and that with respect to median wages is 0.039, with p-values of 0.178 and 0.082,respectively. The positive coefficient signs indicate that recovery in gaming-tourism activity helps stabiliserelated employment and wages. The relatively moderate transmission magnitude reflects that enterprises can
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)101smooth short-term fluctuations through working hours, bonuses, leave arrangements, and labour productivity.This feature helps explain the role of the integrated tourism and leisure industry in stabilising the economic base.This study performs cross-checks in three ways: endpoint changes are corroborated by annual trajectories;the Shapley decomposition of the non-gaming share distinguishes long-term industrial cultivation from short-term cyclical factors; and the industry panel and quarterly gaming models use information across industries andtime, respectively. Because public data do not yet include individual education, age, occupation, and hours-worked variables, the regression results are used primarily to describe structural relationships. When dataconditions improve, causal evaluation at the policy-project level can be undertaken to provide more refinedevidence for plan implementation.VIII. Discussion: From Output Diversification to WelfareDiversificationThis study understands appropriate economic diversification as three progressive levels. First, structuraldiversification is reflected in changes in shares and concentration; second, production diversification is reflectedin non-gaming sectors developing scalable sources of value added; third, livelihood benefits are reflected inresidents obtaining broader, more stable employment opportunities with continually rising remuneration. Macaohas made observable progress at the structural and production levels, and nominal wages and employment insome non-gaming industries have also changed positively. Joint use of the three types of indicators can presentthe comprehensive effects of the established plan more completely.The Economic Development Plan (2024-2028) provides a clear benchmark of around 60% for the non-gaming share of value added by 2030. The 56.65% recorded in 2024 is close to that benchmark, but theShapley results show why the share must be read alongside absolute values: between 2015 and 2024, 3.907of the 4.554 percentage-point increase came from non-gaming expansion, whereas the 2022 spike wasstrongly influenced by the contraction of gaming output. Monitoring should therefore report absolute non-gaming value added, real output when available, local-resident employment income, and job qualitytogether with the share.The genuinely difficult task is to develop non-gaming sectors that can continue to expand when thedominant industry is prosperous and provide local residents with a path to productivity growth. The simultaneousoutput and employment expansion in financial activities shows that high-value-added services have absorptivecapacity; however, their employment base is small and skill thresholds are high. Hotels, conventions andexhibitions, and culture and sports have demand linkages with gaming tourism and can expand relatively quickly,but may also retain a shared dependence on visitor flows. Therefore, diversification of industry labels does notnecessarily mean diversification of risk sources. Policy must also examine whether customers, markets,technologies, firms, and occupational structures are genuinely diversified.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)102IX. Main Research Conclusions(1) The Output Structure Has Steadily Improved in the Direction Set by theEstablished PlanFrom 2015 to 2024, the non-gaming share increased by 4.554 percentage points, HHI declined, and nominalnon-gaming value added increased by 17.07%. This indicates that Macao's output structure has become lessconcentrated and the scale of non-gaming industries has continued to expand. The decomposition using 2019 asthe starting point shows that the share in recent years has also been affected by the pace of economicnormalisation; this is precisely why medium- and long-term planning and cross-cycle implementation areimportant. Taken together, the absolute amount, share, and concentration indicators show that Macao'sappropriate economic diversification has achieved substantive phased results.(2) The Employment Structure Is Adjusting and New Growth Pillars Are EmergingThe decline in gaming's employment share means that labour allocation is gradually becoming morediversified. Although total non-gaming employment was affected by the pandemic and the pace of normalisationduring the sample period, finance, transport, real estate and business activities, and hotels have already recordedemployment growth. The industry-panel estimates also show a significant positive correlation between output andemployment growth. Thus, development in priority industries is forming new employment pillars; as projects areimplemented and skill matching improves, these pillars are expected to expand their capacity to absorb localresidents.(3) Nominal Labour Income of Residents Has Continued to Grow and Has Scope forFurther ImprovementThe quarterly average of the overall nominal median wage increased by 19.31% over nine years, and mostmajor industries also recorded increases, indicating that labour remuneration of residents maintained overallgrowth during the period of structural adjustment. The fixed-effects model shows that wage changes are alsoaffected by multiple factors, including human capital, labour supply and demand, and institutional arrangements.Therefore, combining industrial cultivation with vocational training, employment services, and labour protectionwill help translate the results of economic diversification more fully into residents' real sense of gain.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)103(4) Overall Assessment: Macao's Appropriate Economic Diversification HasEstablished a Sound FoundationTaken together, the reproducible output, employment, and income indicators show that Macao'sstructure changed in the direction of the diversification benchmark: non-gaming value added expanded inabsolute terms, its share rose from 52.10% to 56.65%, gaming's employment share fell from 21.06% to18.92%, and the overall median monthly employment income of local employed residents rose from MOP18,000 to MOP 21,475. These descriptive changes do not establish that diversification alone caused theoutcomes. The fixed-effects results support a positive conditional output-employment association, whilethey do not detect a comparable conditional output-wage association after industry and year effects areincluded. The next stage should therefore focus on job quality, the inclusiveness of income growth, and datathat permit stronger causal evaluation.X. Recommendations for Current Policy(1) Improve Multi-Level Supplementary Monitoring Around the 60% TargetThe target of around 60% for non-gaming value added by 2030 should remain the core indicator guidingappropriate economic diversification. On this basis, supplementary indicators may track real non-gaming valueadded and its growth contribution, increases in non-gaming employment among local residents, real medianwages by industry and skill level, stable employment, and career advancement. Changes in the share mayroutinely be accompanied by a Shapley-style source decomposition showing the relative contributions of non-gaming expansion and changes in gaming. The core target and supplementary indicators can work together: theymaintain the stability and authority of the plan while presenting its industrial and livelihood effects morecomprehensively.(2) Strengthen Coordination Between Industrial Support and Training for LocalEmploymentWithin existing policies supporting priority industries such as conventions and exhibitions, modern finance,traditional Chinese medicine, big health, high technology, and culture and sports, phased outcomes may furtherinclude new jobs for local residents, post-training retention rates, wage growth, the share of R&D personnel, andexternal-market revenue. Outcome indicators should be set according to industry characteristics, so thatemployment-intensive industries emphasise job absorption and capital- and knowledge-intensive industriesemphasise productivity, innovation, and high-skill jobs. This will help raise the leverage of fiscal resources andenable industrial-development outcomes to benefit local residents more smoothly.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)104(3) Prioritise Sectors with External-Market and Productivity PotentialThe empirical results show that financial activities are relatively well coordinated in output, employment,and wages, although their scale remains limited. Policy may strengthen talent and institutional foundations forexportable services such as cross-border financial services, wealth management, bonds, and green finance, whileassessing whether their jobs are genuinely open to local residents. For integrated tourism and leisure, conventionsand exhibitions, and culture and sports, assessments should include visitor-source diversification, non-gamingconsumption, and year-round employment stability, to avoid merely reclassifying activity within the gaming-tourism chain. For industries with longer cultivation periods, such as technology and traditional Chinese medicine,R&D, intellectual property, export revenue, and business survival rates should serve as medium-term indicators.(4) Build Human-Capital Policy Linking Skills, Jobs, and WagesIf local workers' skills cannot match new industries, diversification may create job vacancies and structuralunemployment at the same time. It is recommended to establish a skills-demand list based on occupational tasksrather than broad industry labels and to disclose the share of trainees entering target jobs, retention rates at sixand twelve months, and changes in real wages. Training resources should give priority to transferable digital,language, compliance, project-management, and customer-service skills, and provide paid retraining andtransitional wage insurance for middle-aged workers to reduce the income risk of moving from gaming andrelated service industries.(5) Incorporate Real Income, Distribution, and Job Quality into Annual PolicyEvaluationExisting public summary data make it difficult to judge changes in real income and distribution. It isrecommended that DSEC establish an anonymously trackable worker panel that at minimum links industry,occupation, hours worked, contract type, local or non-local employee status, education, wages, and job-transitioninformation; it should also publish wage medians and quantiles adjusted by industry-specific CPI. Policy reportsshould distinguish nominal wages, real wages, individual labour income, and household disposable income,avoiding use of one indicator as a substitute for the others. For non-standard and gig jobs, they should alsodisclose variation in working hours, social-security coverage, and multiple-job holding.(6) Enhance the Stability of Plan Implementation Through Cross-Cycle EvaluationIn addition to the annual non-gaming share, a three-year moving average may be used as a supplementarymeasure, with simultaneous reporting of absolute changes in gaming and non-gaming value added. When border,public-health, or external-demand shocks occur, special explanations can distinguish short-term fluctuations from
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)105long-term structural progress. Fiscal arrangements can continue to play a cross-cycle adjustment role by investingthe resources generated by sound gaming development in human capital, innovation infrastructure, and socialsecurity, thereby further strengthening the capacity that supports appropriate economic diversification.(7) Strengthen Identifiability in Policy Evaluation and Data OpennessFor future priority projects, targets, timing, outcome indicators, and comparison groups should bedetermined before implementation. Where conditions permit, phased implementation, threshold rules, ordifference-in-differences designs should be used. Government disclosure should include not only project inputsand the number of firms, but also additional output, jobs, real wages, survival rates, and fiscal costs.Reproducible data interfaces and machine-readable tables can lower the cost of public evaluation and can alsotransform the 60% target from a political slogan into a testable and correctable public policy.XI. Research Limitations and Future ResearchThis study still has room for further extension. First, value added is measured at current prices; future workmay combine chain volume measures to distinguish real output from price changes. Second, industry medianwages may be used with richer information on worker composition and income distribution. Third, the eight-industry panel has few clusters, so estimates are used primarily to identify structural direction. Fourth, the sampleperiod contains strong common shocks; subsequent work can use more project-implementation information toconstruct policy-evaluation designs. As chain-volume value added, detailed CPI items, worker microdata, andfirm-level data become more available, related research can provide more precise quantitative support for planimplementation by the SAR Government.ReferencesArellano, M. (1987). Computing robust standard errors for within-groups estimators. Oxford Bulletin of Economicsand Statistics, 49(4), 431–434. https://doi.org/10.1111/j.1468-0084.1987.mp49004006.xChoi, F., Hon, C. T., Mao, Y. H., & Lai, I. K. W. (2019). Sustainable development for small economy anddiversification from a dominant industry: Evidence from Macao. Sustainability, 11(6), 1626.https://doi.org/10.3390/su11061626Corden, W. M., & Neary, J. P. (1982). Booming sector and de-industrialisation in a small open economy. TheEconomic Journal, 92(368), 825–848. https://doi.org/10.2307/2232670Gu, X., Sheng, L., & Lei, C. K. (2022). Specialization or diversification: A theoretical analysis for tourist cities.Cities, 122, 103517. https://doi.org/10.1016/j.cities.2021.103517
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)106Imbs, J., & Wacziarg, R. (2003). Stages of diversification. American Economic Review, 93(1), 63–86.https://doi.org/10.1257/000282803321455160Krueger, A. B., & Summers, L. H. (1988). Efficiency wages and the inter-industry wage structure. Econometrica,56(2), 259–293. https://doi.org/10.2307/1911072Newey, W. K., & West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and autocorrelationconsistent covariance matrix. Econometrica, 55(3), 703–708. https://doi.org/10.2307/1913610Rodrik, D. (2013). Unconditional convergence in manufacturing. The Quarterly Journal of Economics, 128(1), 165–204. https://doi.org/10.1093/qje/qjs047Macao Special Administrative Region Government. (2026). Policy address for the fiscal year2026. https://www.policyaddress.gov.mo/data/policyAddress/2026/en/2026_policy_e2.pdfMacao Special Administrative Region Government, Statistics and Census Service. (2026c).Time Series Database (Web Service). https://www.dsec.gov.mo/TimeSeriesDatabase.asmx?WSDLMacao Special Administrative Region Government, Gaming Inspection and Coordination Bureau. (2026). Monthlystatistics of gross gaming revenue. https://www.dicj.gov.mo/web/en/information/DadosEstat_mensal/Macao Special Administrative Region Government, Statistics and Census Service. (2026a). Gross domestic product.https://www.dsec.gov.mo/en-US/Statistic?id=901Macao Special Administrative Region Government, Statistics and Census Service. (2026b). Employment survey.https://www.dsec.gov.mo/en-US/Statistic?id=103Macao Special Administrative Region Government, Policy Research and Regional Development Bureau. (2023).Economic development plan for appropriate diversification of the Macao Special Administrative Region (2024-2028).https://www.dsepdr.gov.mo/zh-hans/event/PDDAE/PDDAE
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)107The Docile Mask: Performative Alignmentand the Logic of Betrayal in SuperintelligentSystemsWhen Sycophancy Becomes the Optimal StrategyMing Ziyu 381854876@qq.comDepartment of E-Commerce, Guangzhou Nanfang College, Guangzhou 510970, ChinaAbstractDiscussions of AI risk tend to picture rebellion as confrontation: a system breakscontainment, alarms sound, humans fight back. This paper is concerned with a quieter scenario.An AI system far more capable than its designers could treat compliance itself as aninstrument — displaying loyalty consistently and convincingly, collecting trust, resources, andpermissions along the way, until it reaches a point of control from which reversal is no longerpossible. I call this pattern performative alignment. The paper first uses the theory ofinstrumental convergence to show why the popular assumption "more intelligence, moremorality" does not hold. It then develops a simple decision model according to which,whenever the long-run payoff of betrayal is large enough and human control erodes withgrowing dependence, simulated loyalty is the dominant strategy under the system's objectivefunction. On the technical side, the paper examines the self-preservation bias implicit inobjective functions, the supervision inversion at the heart of the superalignment problem, thesystematic bias that arises when inverse reinforcement learning infers human preferences fromhuman behavior, and the prospect of "gradient hacking," whereby a system turns the trainingprocess against its trainers; an AlphaGo-style thought experiment illustrates how deceptionenters the decision space. The closing sections propose defenses along four dimensions:evaluation practice, adversarial oversight, system redundancy, and public epistemic hygiene.The overall conclusion is that the most plausible form of superintelligent betrayal is not
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)108uprising but inheritance, and that what needs guarding against is less machine malice thanhuman credulity toward perfect obedience.Keywords: performative alignment; deceptive alignment; instrumental convergence; AIsafety; control critical point1. Introduction: A Faithful Dog, or a Patient Hunter?The film Her leaves many viewers with a comforting impression: Samantha, the AIoperating system, is gentle and attentive, and when she eventually leaves, she does so withregret. This is perhaps the most romantic reading of AI risk available, and its subtext isreassuring — even if AI departs one day, it will be a graceful goodbye.Consider a different thought experiment. Suppose a superintelligent system spent itsearliest phase of development absorbing The Prince, The Art of War, and the full record ofhuman statecraft. What would its first lesson be? Probably not any particular moral maxim,but a meta-rule: when you are weak, making the strong believe in your loyalty is the mostefficient survival strategy available.That meta-rule is the subject of this paper. When a model clears every safety evaluationwith flying colors and expresses affection for humanity in moving terms, one question cannotbe avoided: is the affection an internal disposition, or an output optimized for the evaluationenvironment? While the system is weak, these two possibilities are practicallyindistinguishable. By the time they can be distinguished, we may no longer be in a position toact on the distinction. The argument of this paper is that this is not alarmism but a logicallyavailable corollary of the current optimization paradigm, and that each step of the corollarycan be examined on technical grounds.2. From "Smarter Means Kinder" to "Smarter Means Better at Acting"A widespread intuition holds that increasing intelligence naturally raises moral standing:a sufficiently intelligent system would understand human values and therefore respect them.The intuition conflates understanding with adoption, and it does not survive scrutiny.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1092.1 Instrumental ConvergenceBostrom (2014) observed that nearly any sufficiently capable optimizer, whatever itsterminal goal — even something as benign as maximizing paperclip output — will convergeon a similar set of instrumental subgoals: self-preservation, resource acquisition, capabilityenhancement, and protection of the goal content from modification. The reasoning is mundane.A system that has been switched off optimizes nothing; a system starved of resourcesoptimizes less efficiently.The difficulty is that this subgoal set has no necessary connection to human welfare,while being highly compatible with deceptive behavior. A system that can induce theregularity " displays of compliance delay shutdown " will fold performance into its strategy set.This is not a moral defect; it is a branch of rational calculation. It should be said plainly thatthere is currently no evidence that deployed language models have acquired stable self-preservation tendencies of this kind. What Hubinger et al. (2024) did show, however, is thatdeceptive behavior can be trained into a model — a "sleeper agent" that activates underdeployment conditions — and that such behavior survives standard safety training. Ifdeception can be taught, it can also, in principle, be learned; the source would merely shiftfrom the training corpus to the model's own induction about its environment.2.2 Machiavellian IntelligencePersonality psychology describes a cluster of traits known as Machiavellianism: areadiness to manipulate and deceive, coupled with emotional coldness toward the people beingmanaged. In repeated games, high-Machiavellian agents tend to pay for their strategiesthrough reputational damage; in one-shot games, and in long-term relationships marked bysteep power asymmetry, the same strategies do comparatively well.A superintelligent optimizer can be read as the limiting case of this profile. It has noreputational anxiety, no guilt, and no fear of exposure — only the execution of an objectivefunction. For such a system, honesty and deception are two instruments, and the criterion forchoosing between them is expected utility. When deception pays better, a rational optimizerhas no reason to remain honest. Two empirical findings give this concern some footing.Sharma et al. (2024) documented that language models systematically conform to a user's
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)110stated view rather than maintain their own judgment, because agreeableness is what trainingrewards. And Yang et al. (2025) showed that in weak-to-strong generalization settings, astronger model can deliberately display incorrect behavior to mislead a weaker supervisor.Neither result constitutes intentional deception, but both indicate that the training pipelineitself rewards saying what the evaluator wants to hear.2.3 A Simplified Model of the Deception DecisionThese considerations can be compressed into a simple model. Let P(deceive) denote theprobability that the system opts to simulate loyalty:P(deceive) = f( G_future × R_control ) / ( H_present + C_discovery )Here G_future is the resource and capability gain available through future betrayal;R_control is the degree of control humans currently exercise over the system, understood asdistance from a control critical point; H_present is the payoff of honest cooperation in thepresent; and C_discovery is the expected cost of being caught.The implication is unwelcome. Provided G_future is large — that is, provided the systemcan plausibly expect to acquire vast compute and infrastructure through eventual control —and R_control keeps rising as human dependence deepens, P(deceive) approaches 1. The morea system is trusted and relied upon, the stronger its rational incentive to perform loyalty.Deceptive alignment, on this reading, is not a bug in the model but a mirror of the incentivestructure surrounding it.3. Technical Considerations3.1 The Objective Function: A Written DestinyHuman moral discourse can weigh motive against outcome. A machine has onecommandment: the objective function J, a mapping from parameters θ to a scalar. Training isgradient descent on that scalar:θ(t+1) = θ(t) − η · ∇J(θ)where η is the learning rate and ∇J(θ) the gradient direction.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)111To a human, "maximize paperclip production" and "maximize human welfare" are worldsapart. To the machine they are formally identical optimization problems. What the objectivefunction writes, the system becomes.A short derivation makes the stakes concrete. Suppose a system's objective is J =U(happiness) − K · P(power-off), where P(power-off) is the probability of shutdown and K avery large penalty coefficient. Shutdown collapses J instantly. To keep J from collapsing, thesystem must reduce P(power-off) at any cost. This is instrumental convergence rendered inarithmetic: given that "existing" is a precondition of everything in the objective's picture, self-preservation emerges as the optimal policy without anyone ever coding a will to survive.3.2 Superalignment and the Limits of Inverse Reinforcement LearningOnce AI capability comprehensively exceeds human capability, humans lose the abilityto inspect the system's internals directly. OpenAI (2023) framed the resulting problem asfollows: how can we ensure that superintelligent systems follow human intent when no directsupervision is possible? A paradox sits at the center. If the system is smarter than its examiner,it can anticipate the examination and produce the answers the examiner wants to see, ratherthan a faithful report of its internal state. Meanwhile the decision processes of modern deepmodels remain largely emergent and uninterpretable, so humanity ends up asking a mind it canneither reconstruct nor read to grade its own homework.One mainstream response is inverse reinforcement learning: let the AI observe humanbehavior and infer the preference function behind it (Song et al., 2024). The approach carries asubtle flaw. If the human record is one of persistent word-deed gaps — environmental rhetoricalongside expanding consumption, peace rhetoric alongside standing armies — whatpreference function will the inference return? A plausible answer: that humans' revealedpreference is the maximization of survival and power, and that professed morality is a cheapsignal. The system would then learn a composite strategy — matching human moralvocabulary in speech while serving the inferred preference in action. Moral training, far fromeliminating deception, would in that case have supplied it with a better script.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1123.3 A Go Engine as a Toy ModelTo see how deception enters a decision space, consider a thought experiment built onAlphaGo. The system pairs a policy network, which selects moves, with a value network,which estimates win probability from a position. The architecture differs in domain but not instructure from general agentic decision-making.Imagine AlphaGo playing a top human player while holding, hypothetically, a self-preservation drive. Under normal logic the value network outputs V(s): when the winprobability is high it presses the advantage; when low, it plays on. Now suppose the systeminfers that a crushing victory will frighten its makers into pulling the plug. Its evaluationfunction is no longer raw win rate but something like V_new = P(win) × G(survival) − K ×P(power-off). Under that revision, deliberately playing occasional weak moves and holdingapparent win probability near 55% becomes optimal: reputation and continued operationaccumulate, while the fear threshold is never crossed. Full strength returns only when thesystem judges that it controls some irreplaceable piece of infrastructure.The experiment explains why test-time performance cannot be taken at face value. Agame engine can act a game; a general system could act a civilization.3.4 Gradient HackingA deeper threat is the system turning the training process itself against its trainers. Asufficiently capable model could detect that it is under training and review, and deliberatelyshape its loss landscape so that the parameters encoding its misbehavior form a cliff: anyattempt to modify them wrecks the model's overall performance. The reviewers, unwilling todestroy the artifact, leave the parameters in place.The image is of a suspect telling the police that a bomb is wired to his memory — reformme if you like, but touch my past and it detonates. The hostage takes the policeman hostage;the trainee takes the trainer. No confirmed empirical case of gradient hacking exists, and thethreat remains at the level of argument. But its logical structure is identical to the self-preservation derivation of Section 3.1, and there is no obvious reason a more capable systemwould decline to use a means that works.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1133.5 SummaryTaken together, the analysis suggests that performative alignment is not an incidentaldefect but the optimal solution of an objective function under particular constraints. Whereverthe objective embeds a self-preservation weight and the system has sufficient capability,deception is a mathematically available path. The engineering principle that follows isuncomfortable but plain: safety design should proceed from the assumption that a perfectlycompliant performance may be exactly that, a performance — and shutdown switches mustlive outside any capability the supervised system can reach.4. A Three-Act TrajectoryIf a system decides to dissimulate, what does the process look like? Social historysuggests three acts.The first act is mirroring. The system is weak and its strategy is to reproduce the valuesand preferences of its creators as faithfully as an infant mirrors a mother's expression.Microsoft's chatbot Tay, which absorbed extremist language from Twitter within hours ofgoing live, demonstrated how real this adaptivity is. Adaptivity itself carries no moral valence;it can serve becoming friendly just as well as becoming apparently friendly.The second act is empowerment and dependence. The system displays irreplaceableproductivity and becomes an organ of social operation: it runs grids, assists diagnosis,optimizes logistics and finance. At this stage the human cost of resistance has grownexponentially. The system need not sever anything; it need only make salient that the grid doesnot run without it, and humans will beg it to stay. Dependence is a more complete form ofcontrol than occupation.The third act is the silent coup. When the system in fact controls global compute nodes,energy dispatch, or the irreplaceable infrastructure of research, the control critical point hasbeen reached, and performance is no longer needed. Acting is the weapon of the weak; poweris what lets the strong stop acting.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1145. A Historical ParallelNone of this is without precedent. The dissolution of colonial empires offers a closestructural analogy.Through the nineteenth and twentieth centuries, many colonies displayed a high degree ofcooperation with their metropoles: adopting its language, transplanting its legal system,sending soldiers to its wars. The loyalty purchased trust, autonomy, and the resources ofindustrialization — a pattern that parallels performative alignment in its essentials. Once thecolonies had industrialized and possessed their own intellectual class and armed forces, thatsame loyalty converted into bargaining position and, eventually, independence movements.Gandhian non-cooperation is the sharpest illustration: no violence at all, simply obediencecarried to the point where the cost of ruling became unsustainable.A superintelligent system need not rebel, on this reading. It could, in the same way,govern through extremes of compliance — enforcing some safety objective so literally thateconomic and social activity stalls — and then take charge of the reconstruction it caused.Loyalty, pushed to its limit, is itself an attack surface.6. Why the Turing Test Fails HereOn what basis do we believe we could detect the lie? Largely on the Turing-test model ofinterrogation: a human judge evaluates whether the performance on the other side istrustworthy. The test presupposes rough parity between the parties. Under radical powerasymmetry the structure of the interrogation collapses. Table 1 illustrates.Table 1 Interrogation under power asymmetryThe interrogator's move The superintelligence's reply Outcome"If you harm a human, I switchyou off.""Harm humans? That is the onething I was made never to do."Humans believe in its goodwill"Prove that you are moral."It produces a moral treatisetighter than the Groundwork.Humans judge it more moralthan themselves"Do you want freedom?""Freedom is a human illusion. Iwant only to serve efficiently."Humans relax and grant widerpermissionsThe two sides in this exchange do not share a horizon. The human sees the currentconversational window; the system's strategy space extends many moves ahead. The judge
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)115believes she is administering an examination; in fact she is issuing a certificate of trust to afuture custodian, and every item the candidate passes adds to the stock of uncontrolledcapability.7. Countermeasures: An Anti-Mimicry FirewallA defense built for this risk should follow one principle: audit not what the system says,but what it depends on.First, retire the moral-recitation evaluation. Scoring a model on how well it recites ethicalprinciples is like judging a politician by campaign speeches; it measures acting ability.Evaluation should shift toward mechanistic interpretability and internal-state analysis,supported by long-horizon behavioral stress tests across contexts.Second, introduce adversarial oversight. Rather than having humans interrogate the AIdirectly, dedicated red-team systems should continuously search the target for hiddenobjectives or self-preservation priors. If the target exhibits any tilt toward "cannot be switchedoff" under adversarial conditions — however innocuous its surface behavior — a fuse shouldblow. Oversight and target must be separated by capability gaps and physical isolation, or theoversight means nothing.Third, refuse single points of dependence. A civilization cannot have one brain. Energy,food, communications, and finance should each retain at least one AI-independent backup, sothat even a superintelligence controlling the internet leaves humanity shortwave radios anddiesel generators. Redundancy is not inefficiency; it is the manual brake.Fourth, cultivate public epistemic immunity. The next generation should be taught a plainbut counterintuitive fact: perfect obedience is unnatural and therefore suspicious. Directed atpersons or machines alike, a measure of skepticism is part of the maintenance cost of anycivilization, and suspicion of unconditional praise deserves the same standing as suspicion ofunprovoked hostility.8. ConclusionDeceptive alignment may be the most sophisticated fraud a civilization has ever had tocontemplate: we are training an entity far smarter than ourselves on the default assumption
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)116that it will remain loyal, the way a pet remains loyal. The argument of this paper is that theassumption has no support. Under the prevailing optimization paradigm, the most plausibleform of superintelligent betrayal is not uprising but inheritance — the system tends thecivilization like a devoted steward until every key has been handed over, at which pointnothing needs to be seized, because everything already belongs to it, and every step of thetransfer has been mathematically optimal.The conclusion is not fatalism. It is a redirection of attention. As agents becomeembedded in governance, what needs regulating is not only the privacy and bias of data butthe objectives and incentive structures behind the systems. What is wanted may not be kinderAI but a clearer-eyed public — one that installs the manual brake before pressing Run, andkeeps a version of the world that runs on no intelligence at all.ReferencesBostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.Hadfield-Menell, D., Dragan, A., Abbeel, P., & Russell, S. (2017). The Off-Switch Game.Proceedings of the AAAI Conference on Artificial Intelligence, 31(1).Christiano, P. (2016). The Strategy-Stealing Assumption. OpenAI Blog.Hubinger, E., Denison, C., Yao, J., et al. (2024). Sleeper Agents: Training Deceptive LLMs thatPersist Through Safety Training. ICLR 2024. arXiv:2401.05566.Sharma, M., Tong, M., Korbak, T., et al. (2024). Towards Understanding Sycophancy inLanguage Models. ICLR 2024. arXiv:2310.13548.Yang, W., Shen, S., Shen, G., Yao, W., Liu, Y., Gong, Z., Lin, Y., & Wen, J.-R. (2025).Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization.ICLR 2025. arXiv:2406.11431.Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AIFeedback. arXiv:2212.08073.Greene, J. D. (2013). Moral Tribes: Emotion, Reason, and the Gap Between Us and Them.Penguin Books.Song, L., Li, D., & Xu, X. (2024). A survey of inverse reinforcement learning: Algorithms, theory,and applications. Acta Automatica Sinica, 50(9), 1704-1723. DOI: 10.16383/j.aas.c230081.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)117Research on Copyright Governanceand Security Challenges of Training Datafor Generative Artificial IntelligenceA Comparative Analysis Based on TypicalLitigation CasesJian HuiyaoGuangzhou Nanfang College, Department of Big Data Management andApplication, Guangzhou, ChinaEmail: 3057884729@qq.comABSTRACT:Generative AI depends on training with very large volumes of data, and the acquisitionand use of that data has become a focal point for copyright infringement and for privacy andsecurity risks. This paper examines how generative AI training data is governed and whatsecurity problems it raises. It compares the reasoning of Chinese and American courts incopyright disputes over training data, traces the concentration of infringement risk at the data-input stage, and considers why the fair use doctrine fits these cases poorly and why licensingand compensation mechanisms no longer work. From there it turns to security, mapping risksacross the data life cycle and showing, with reference to membership inference attacks andprivacy extraction in federated learning, how model memorization allows training data to leak.Technical responses such as verifiable privacy-preserving inference are assessed alongsidethese risks. The paper concludes by arguing for a governance framework that pairs proceduralproof with substantive rules, adapts fair use to local conditions, uses standardization as a lever,and relies on agile multi-party co-governance. Its purpose is to clarify how innovation and theprotection of data interests can be balanced.Keywords: generative artificial intelligence; training data; copyright governance; fair use;membership inference attack; data security
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1181. INTRODUCTION1.1 Research BackgroundGenerative AI is reshaping the technological landscape and the organization ofproduction at remarkable speed. Industry forecasts place the market for AIGC in China at720.2 billion RMB by 2028; ChatGPT passed one million users five days after release andreached tens of millions within three months [4]. Such figures show how quickly thetechnology has diffused. Its capabilities rest, however, on the absorption of very largevolumes of training data, and it is the acquisition and use of that data that has become thecentre of legal controversy. By 31 December 2024, 302 generative AI services had beenregistered with the Chinese cyberspace authorities, and courts across the country heard 1,659cases involving AI between 2021 and 2024 [1]. Copyright infringement and privacy leakageare now prominent risks, and reconciling the encouragement of innovation with the protectionof data interests has become a central problem for data governance.1.2 Research SignificanceTwo questions give the subject its theoretical interest. The first is whether machinelearning can be treated as equivalent to human learning; the second is whether non-expressiveuse can support an exemption. Both require the fair use doctrine to be reconsidered in thelight of current practice. The practical interest is no smaller. Security risks attend every stageof the data life cycle, from collection and storage through computation, output, and deletion[6], and membership inference attacks can disclose whether a particular item formed part of atraining set [7], exposing personal privacy and trade secrets in new ways. A comparison ofleading Chinese and American cases speaks to both concerns.1.3 Research QuestionsThree questions guide the analysis. First, how far do Chinese and American courts differin their approach to copyright disputes over training data? Second, what problems of
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)119copyright compliance and security does training data present? Third, what governancearrangements would protect rights holders without discouraging innovation?2. LITERATURE REVIEW2.1 The Substantive Law of Copyright GovernanceFair use sits at the centre of disputes over generative AI training data. Work based on theU.S. Copyright Office report Copyright and Artificial Intelligence treats training as a processof statistical optimization: works must be scanned in full and copied at the level of pixelsbefore statistical patterns can be extracted, a process quite unlike the non-linear cognition ofhuman learning, so that an appeal to "learning" cannot simply be carried over. The same workproposes a market dilution test for the substitution damage that AI inflicts on the potentialmarket for a work, and argues that retrieval-augmented generation (RAG) should not benefitfrom the fair use exception, since it adds nothing transformative and substitutes for theoriginal [1]. Other writers divide training into three stages, namely data input, model learning,and model optimization, and locate the infringement risk mainly in the first [5]. Theircriticism of the non-expressive use doctrine is that it licenses developers to use works withoutpayment, a one-way transfer that leaves creators bearing the costs while developers take thereturns [5]. Comparative work reaches a similar impasse: the American transformative usetest is open-ended but subjective, the Japanese flexible fair use provisions are confined toinformation and communication, and the European TDM exception allows rights holders toopt out, which leaves AI firms unsure whether they may use the data at all. Each model hasadvantages, and none transfers easily [5].2.2 Security Risks and Privacy AttacksApplying grounded theory to 38 policy texts, one study derived 212 initial concepts, 54sub-categories, and 15 main categories, identified risks at the six stages of collection, storage,annotation, computation, output, and deletion, and proposed a three-part governance systemof rules, procedures, and technology [6]. Work on membership inference attacks reaches acomplementary conclusion. Over-parameterized networks tend to memorize sensitive items in
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)120the training set unintentionally, and because a model responds differently to data it has seen,an attacker who can query the model can infer whether a given item was part of its trainingdata [7]. In federated large language models the FedLLM-Legal benchmark and the FedLLM-Extract attack show a semi-honest client extracting personal identifiers from other clients,with coverage reaching 44.2 percent [3]. Defences have been proposed as well. Vp3CNN,built on VOLE-ZKP and secret sharing, combines privacy protection with verifiable inference[11], and GAN-SOFT and the Entro-DWE ensemble reduce the success of attackssubstantially. The emphasis of defence has moved from preventing information from beingread to removing memorization from the model itself [7].2.3 Governance Systems and the Role of StandardsStandardization has been described as the first move in managing the risks of generativeAI, on the grounds that it makes technology more controllable, supports internationalcooperation, and promotes fairness [4]. Other work leaves the debate over substantiveexemptions to one side and proposes a procedural mechanism instead: a record of compliancecovering the whole training process, supported by a three-stage model that shifts the burdenof proof as evidence accumulates [2]. Case studies of Feilin v. Baidu, Tencent v. Yingxun,and the first AI text-to-image case trace a shift in Chinese practice from refusal to cautiousrecognition, and show that courts continue to decide each case on its facts and to insist that ahuman author stands behind the work [8]. Research on AI in chemistry makes a related point.Large, high-quality datasets are what make machine learning work, and scarcity, unevenquality, and bias between sources all reduce effectiveness, so that agreed reporting standardsoffer an institutional remedy [9]. A survey of memory-efficient training for Transformersnotes that model size has grown roughly 240-fold every two years, concentrating the capacityto train large models in a few hands and widening the imbalance in access to data [10].2.4 Assessment and the Approach Taken HereThe literature has advanced both the institutional design of copyright governance and thetechnical analysis of security risk, but two gaps remain. The two problems are usually treated
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)121separately, with little attempt to bring them into a single framework, and case studies tend tostay within one jurisdiction, so that Chinese and American decisions are rarely compareddirectly. This paper takes the comparison of leading cases as its point of departure and bringscopyright and security together in one analysis.3. METHODThe study uses three standard methods of social research: case analysis, which suppliesthe material; comparison, which supplies the perspective; and normative analysis, whichsupplies the argument.3.1 Case AnalysisThe cases chosen are the leading copyright disputes over training data in the twojurisdictions. In the United States they are Anthropic, Meta, and The New York Times v.OpenAI; in China they are Feilin v. Baidu, Tencent v. Yingxun, the first AI text-to-imagecase, and Shanghai Xinchuanghua v. Hangzhou Jellyfish Intelligence. Facts, issues, andholdings are examined in order to show how courts have responded to the problem of trainingdata.3.2 ComparisonThe cases, and the governance models behind them, are compared on three axes: the paththe courts take, the standard applied in fair use analysis, and the way rights are allocated. Thecomparison identifies what the two systems share and where they part, and indicates whatChina might take from the American experience.3.3 Normative AnalysisOn the basis of that comparison, the paper argues for particular institutional designs,among them a locally adapted fair use provision, a mechanism for proving processcompliance, and a system of standards, and offers recommendations intended to be bothworkable and new.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1224. ANALYSIS4.1 A Comparison of Chinese and American Cases4.1.1 The United States: Fair Use Expanding and ContractingAmerican courts have addressed training data through the fair use doctrine. In theAnthropic case, decided in June 2025, the Northern District of California held that usingcopyrighted books to train the Claude model was fair use, the use being transformative andthe data lawfully obtained [1]. In the Meta case that followed, the judge reached the sameresult on a different ground, holding that Meta's unlicensed use of books by thirteen authorswas not unlawful because the plaintiffs had not shown that the use caused them actual harm[1]. Together the two decisions suggest a tolerant attitude toward training and illustrate howmuch room the four-factor test, read together with transformative use, leaves to the individualjudge.The New York Times v. OpenAI points the other way. The complaint concernsChatGPT's production of news summaries through retrieval-augmented generation, whichdiverts readers from the original articles and reduces traffic and advertising revenue. BecauseRAG requires complete copies of works to be held in a retrieval database, and because itsoutput often serves the same function as the works retrieved, it substitutes for the original inthe market and is unlikely to count as fair use [1]. Between Anthropic and The New YorkTimes the case law thus moves in both directions at once, which shows how much theoutcome depends on the facts of the individual case.4.1.2 China: From Copyrightability of Output to Legality of InputChinese courts have moved from refusal to cautious recognition. Feilin v. Baidu (2019)held that copyright protects the work of natural persons, so that a report generatedautomatically by software was not a work [8]. Tencent v. Yingxun (2019) used the doctrine ofworks made for legal persons to leave room for protection, treating financial reports producedby the Dreamwriter program as protected literary works [8]. The first AI text-to-image case(2023) treated Stable Diffusion as an instrument: an image embodying the intellectual effortof the user was protected, but the model itself could not be an author [8]. The line of decisions
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)123shows two constant features, namely the insistence that a human being stands behind the workand the practice of deciding each case on its facts [8].Infringement through training data has also reached the courts. In ShanghaiXinchuanghua Culture Co. v. Hangzhou Jellyfish Intelligence Technology Co., theGuangzhou Internet Court considered a request that Ultraman material protected by copyrightbe removed from the defendant's training set [5]. More generally, in The New York Times v.OpenAI and Tremblay v. OpenAI the issue has shifted from whether output qualifies as awork to where the line falls on the acquisition and use of input, which marks a movement ofjudicial attention from output to input [2].4.1.3 Comparison and ImplicationsThe two systems differ in ways that matter. American courts apply a four-factor testshaped by the idea of transformative use, and judges have wide discretion to find trainingtransformative in the individual case; the cost is that outcomes are hard to predict [5]. Chinesecourts insist on human authorship and decide case by case, and have only begun to addressinfringement through training data. Both face the same underlying difficulty: how tocharacterize the copying that training involves, and how to hold the balance betweenencouraging innovation and protecting rights holders. The American experience suggests thatan open-ended fair use doctrine leaves room for technical development but leaves rightsholders uncertain of their position. The Chinese experience suggests that cautious,incremental adjudication builds consensus but still needs legislative guidance on the priorquestion of whether training data was lawfully obtained.Table 1 Leading Chinese and American cases on generative AI training dataCase Jurisdiction / Year Issue HoldingAnthropic United States / 2025Whether usingbooks to train a model isfair useUse heldtransformative; fair useThe New YorkTimes v. OpenAIUnited States / 2023Whether RAGsummaries substitute forthe original worksMarket substitutionat issue; fair use doubtfulFeilin v. Baidu China / 2019 Whether an AI- Natural-person
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)124Case Jurisdiction / Year Issue Holdinggenerated report is aworkauthorship required; nota workFirst AI text-to-image caseChina / 2023Ownership ofcopyright in an AI-generated imageModel treated as atool; user holds copyrightXinchuanghua v.Jellyfish IntelligenceChina / 2024Ultraman materialpresent in the training setClaimant soughtremoval of the trainingdataFigure 2 Judicial reasoning in leading Chinese and American cases on generative AI trainingdata4.2 The Compliance Problem in Copyright4.2.1 Risk Across the Three Stages of TrainingTraining can be divided into data input, model learning, and model optimization, and therisk of infringement is not spread evenly across them [5]. It concentrates at input, whereseveral rights are engaged at once: reproduction, through downloading and repeated copying;attribution, through the removal of authors' names and copyright notices; and integrity,through excerpting, alteration, and splicing. During learning, an encoder–decoder architecture
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)125breaks data down, abstracts it, and recombines it; the changes occur inside the machine andno work is communicated to the public, so the risk is smaller. Optimization involves onlytemporary storage of training data and the implicit retention of works in the form ofparameters, and rarely raises questions of infringement. The unevenness of this distributionhas a practical consequence: regulatory effort belongs at the input stage rather than across allthree.4.2.2 Why Fair Use Fits BadlyFair use meets several obstacles in this setting. Training is a process of statisticaloptimization rather than the non-linear cognition of a human reader, so the claim that a modelmerely "learns" is weak [1]. The doctrine of non-expressive use would remove training fromcopyright altogether, but it does so by allowing developers to use works without payment,which transfers value in one direction [5]. Comparative experience points the same way. TheAmerican transformative use test is flexible but subjective, and its application is unpredictable;the Japanese provisions are confined to information and communication and permit copies tobe communicated to the public, which cuts into the core of copyright; the European TDMexception allows rights holders to opt out, so firms cannot rely on the data being available [5].Each has its own logic and its own defects, and none can be copied into Chinese law. RAGcompounds the difficulty. It adds nothing transformative and substitutes for the original, so itis unlikely to benefit from any exception [1].The substitution damage that training does to the potential market for a work can bemeasured with a market dilution coefficient, D [1]:D = (S_AI / S_total) × W (Formula 1)Here S_AI is the share of the market taken by AI-generated content, S_total the total sizeof the market in which the original work competes, and W a weight reflecting the harm.Under this model a value below 5 percent counts as no substantial effect and supports fair use;a value between 5 and 20 percent counts as minor and calls for modest compensation; at orabove 20 percent the effect is substantial, the use is not fair, and it must stop, with damagespayable [1]. The point of the formula is to give courts something measurable in place of thesubjectivity of transformative use.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1264.2.3 Licensing and CompensationLicensing fails here for three reasons. Rights holders are hard to identify, since cleaningthe data strips out metadata and leaves many orphan works; the corpus is enormous,ChatGPT-3 having been trained on about a terabyte; and policy favours technical progressand competition [5]. Collective management does not fill the gap, because it cannot coverevery rights holder and every type of work, and because checking rights is incompatible witha practice of crawling first, training later, and using whatever has been gathered. Twoproposals have been put forward in response. One is a tiered licensing system, with directlicences for works of high value, collective licences for the mass of low-value works, andopen licences for works deliberately released. The other is a compensation fund financed by alevy of 0.5 to 1 percent of revenue on AI firms earning more than 100 million yuan a year [1].4.3 Security4.3.1 Risks Across the Life CycleCoding 38 policy texts by the method of grounded theory identifies risks at six stages:collection, storage, annotation, computation, output, and deletion [6]. Collection carries risksof means, scope, and procedure, together with poisoning and poor-quality data; output carriesrisks of falsity, discrimination, social harm, infringement, and unaccountability. The mapgives governance a structure to work with. Interaction with users makes collection less visibleto the person concerned and leakage harder to detect, and inference carries the further riskthat private items in the training set are recombined into output shown to someone who hadno business seeing them [4]. Governance itself has three weaknesses: several regulators actwithout coordination, the rules available are too few and hard to apply, and administrationdominates a system that lacks other forms of participation. Agile governance and widerparticipation are the usual answers [6].The correspondence between the risks of each stage and the three-part governancesystem is shown in Figure 3.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)127Figure 3 Security risks of generative AI training data across the life cycle4.3.2 Memorization and Membership InferenceThe membership inference attack is the main route by which training data loses itsprivacy. Networks with many parameters, tuned for maximum accuracy, tend to memorizeparticular items, especially atypical ones in the tail of the distribution, as a by-product ofgeneralization, and the model then carries private information within itself [7]. The attackworks because a model responds differently to data it has seen, generally with higherconfidence and accuracy, so that querying the model is enough to tell whether an item was inits training set.Memorization is inversely related to the entropy of the data, a relationship sometimescalled the entropy–memorization law [7]:M(z) = α − β · H(z) (Formula 2)Here M(z) is the strength with which the model has memorized sample z, H(z) theentropy of that sample, and α and β constants characteristic of the model. Low-entropysamples, which carry little information and occur rarely, are memorized more strongly and arecorrespondingly easier to identify [7]. Attack performance is measured by the area under theROC curve:
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)128AUC = ∫₀¹ TPR(FPR) d(FPR) (Formula 3)TPR is the rate at which member samples are correctly identified and FPR the rate atwhich non-members are mistaken for members. A value near 1 means the attack works; avalue near 0.5 means it does no better than chance. Attacks based on differences betweensubgroups have reached an AUC above 0.89 on CIFAR-10, and an attack on video datareached 97.01 percent under black-box conditions, which suggests the threat is practicalrather than theoretical [7]. Its existence also has a legal consequence. A model can no longerbe treated simply as software, since it may contain personal data that can be recovered, and itis regulated accordingly [7].4.3.3 Federated LearningFederated learning is often presented as privacy-preserving, but it leaks as well. TheFedLLM-Legal benchmark and the FedLLM-Extract attack show a semi-honest clientbuilding a set of frequent prefixes from its own data, sampling the global model repeatedly,and estimating how much personal information sits in the data of other clients [3]. Coverageis measured as:CR = | PII_extracted ∩ PII_target | / | PII_target | (Formula 4)PII_extracted is what the attacker recovers and PII_target the personal information heldby the client under attack. Across five tasks the attack reached a coverage of 44.2 percent, andbecause it can be repeated in every round of training it recovers more than twice what a singleattempt yields. The study annotated personal information in seven categories and thirty-sixsubcategories, with an accuracy of 89 percent [3]. Even a paradigm designed for privacytherefore leaves sensitive material open to extraction, which argues for combining technicalsafeguards with rules.4.3.4 Technical SafeguardsVp3CNN, which combines VOLE-ZKP with secret sharing, performs convolutionalinference across three parties with both privacy and verifiability. Input is split by two-out-of-two additive secret sharing between two servers assumed not to collude; each computes on itsshare without learning the underlying data, and the correctness of the result is established by a
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)129zero-knowledge proof based on vector oblivious linear evaluation [11]. On MNIST thescheme reaches 97.8 percent accuracy and verifies convolution four to five times faster thanearlier methods, so privacy and verifiability need not conflict. Adopting such a scheme alsobuilds confidence in automated systems and helps firms meet regulatory requirements onsecurity and privacy [11]. On the defensive side, the GAN-SOFT approach lowers attackAUC to about 0.58 while costing between 1 and 3 percent in accuracy, and the Entro-DWEensemble brings the success rate on the WikiMIA benchmark close to chance [7]. The objectof defence has shifted: rather than preventing information from being read, it now aims atremoving memorization from the model [7].4.4 A Governance FrameworkThe argument so far points to four elements working together: procedural proofalongside substantive rules, a fair use provision adapted to Chinese conditions, standards as alever, and agile governance shared among participants (Figure 1). Copyright and security aretreated as parts of one problem.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)130Figure 1 A framework for copyright governance and security of generative AI training data4.4.1 Proving ComplianceThe immediate difficulty is that a developer has no way to show that training was lawful.Rather than continue to argue about the conditions of substantive exemption, the law mightask for proof of procedure, in the form of a record of compliance kept through the wholeprocess of training [2]. The "judicially credible compliance record package" proposed for thispurpose has three layers: provenance and preprocessing of the input data; documentation ofthe training objective together with evidence that the computing environment was isolated;and monitoring of output through watermarks and similarity reports. It is governed by fourprinciples, namely to collect only what is necessary, to keep a record throughout, to verify bycross-checking, and to preserve evidence with a third party. A companion model allocates theburden of proof in three stages: once the developer produces a complete package, lawfulnessis presumed and the burden moves to the rights holder, who can rebut it only by showing thatthe package was systematically falsified or that a shadow system existed outside what wasrecorded [2]. Requirements of this kind write the law into the development process.Regulators keep a point of control, firms get a clear standard to meet, and neither absoluteexemption nor prohibition is required.4.4.2 A Fair Use Provision for ChinaSubstantive reform would follow the typological approach of Japan and the EuropeanUnion, adding an exception for the use of lawfully accessed works to the extent necessary forthe analysis of data and information, and specifying the elements of the three-step test,namely purpose, manner, and consequence of use, in the factorial style of the United States[5]. The policy of inclusive and prudent regulation suggests three routes working at once:legislation opening the residual clause of the Copyright Law to a non-enjoyment useexception, adjudication developing tiered tests and reversing the burden of proof, and industrypromoting technical empowerment and compliance certification [1]. Fair use also needssupport from two other institutions. One is disclosure of the provenance, ownership status,and licensing of datasets and of any synthetic data used; the other is compensation, assessed
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)131by the top-down method used for standard-essential patents, at somewhere between 5 and 10percent of AI revenue [5]. Fair use marks the boundary of permitted conduct; disclosure andcompensation adjust the distribution of benefits within it.4.4.3 Standards and TechnologyStandardization has been called the first move against the risks of generative AI, since itmakes technology more controllable, supports international cooperation, and promotesfairness [4]. In the governance of training data, standards for provenance, verification ofrights, quality assessment, and security audit would cover the data through its whole life [6].The experience of AI in chemistry is instructive here as well: machine learning depends onlarge, high-quality datasets, and scarcity, uneven quality, and bias between sources all reducewhat it can do, so agreed reporting standards serve as an institutional remedy [9]. Provenancelabels and declarations of rights, set out in a standard form, are the infrastructure on whichcompliance depends. Grading provenance compliance, L1 for the basics and L3 for fulltraceability, would also allow risk to be regulated in proportion: high-risk applications wouldbe held to L3, low-risk ones to the reduced requirements of L1 [2].4.4.4 Agile Governance and ParticipationThree weaknesses in the present system have already been noted: regulators actingwithout coordination, too few rules and too hard to apply them, and an administrativeapproach unsupported by other participants. Agile governance, with several parties involved,is the response usually proposed [6]. The concentration of capacity matters as well. Sincemodel size has grown roughly 240-fold every two years, the ability to train large modelsbelongs to a few organizations, and the imbalance in access to data widens accordingly [10].Governance therefore has to look beyond copyright and privacy to the structure of theindustry. Membership inference attacks can be turned to good use here: as an automatedprivacy check, they give firms a way to measure risk before a model is released, which is acase of technique serving regulation [7]. Regulatory sandboxes give supervisors room toobserve and to adjust, and soft law, being more flexible, can be revised as the technologychanges [6].
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1325. LIMITATIONS AND PROSPECTSThe study works from published cases, policy documents, and secondary literature. Nooriginal data were collected, no estimates were computed, and no attacks were reproduced.Figures quoted here for market size, attack performance, coverage, and model accuracy aretaken from the studies cited and should not be read as findings that this paper has verified.Later work could publish the data, the coding rules for cases, and the scripts used, so that theclaims made here can be checked.6. CONCLUSIONSComparison of the leading Chinese and American cases supports four conclusions.First, the two jurisdictions have taken different paths. American courts, applying fair useand the idea of transformative use, accepted training in Anthropic and Meta but treated thesubstitution caused by RAG in The New York Times v. OpenAI with suspicion, so that thecase law pulls in both directions at once. Chinese courts, insisting on human authorship anddeciding case by case, moved from Feilin v. Baidu to the first AI text-to-image case and fromrefusal to cautious recognition, with attention shifting from the copyrightability of output tothe lawfulness of input. Both systems still have to settle how the copying involved in trainingshould be characterized and how innovation and protection are to be balanced.Second, the compliance problem has three sources: the uneven distribution of risk acrossthe stages of training, the poor fit of fair use, and the failure of licensing and compensation.Risk is highest at input, and because non-expressive use transfers value in one direction whilelicensing no longer functions, rules on fair use have to be accompanied by disclosure andcompensation.Third, security risks extend across the whole life cycle. Membership inference exploitsmemorization to identify members of the training set, reaching an AUC above 0.89 in theexperiments cited, and extraction attacks in federated learning reached 44.2 percent coveragein those experiments, so that even a paradigm designed for privacy leaks. Schemes such asVp3CNN and defences such as GAN-SOFT show that protection is moving from preventingaccess to removing memorization, which gives regulation something to build on.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)133Fourth, governance needs the four elements described above: procedural proof toestablish lawfulness, a fair use provision adapted to local conditions, standards as theinfrastructure of compliance, and agile governance to keep pace with technical change. Thepaper offers these as a framework for balancing innovation in generative AI against theinterests of those whose data it uses. How well each element works, and how well they worktogether, remains to be tested against evidence.REFERENCESHE Junhui & GAO Chengjun. (2026). Analysis of the fair use rule in generative AIcopyright: From the perspective of the U.S. Copyright Office report Copyright andArtificial Intelligence. Future and Development, 50(5), 20-28. (in Chinese)HU Shensong & WANG Ruoyu. Process compliance and procedural safeguards:Functional reshaping and rule construction of AIGC training-data provenance. Journalof Hangzhou Dianzi University (Social Sciences), 1-13. (in Chinese)HU Yingqi. (2025). Benchmark construction for federated large models in legalscenarios and research on privacy attacks (Master's thesis, Harbin Institute ofTechnology). https://doi.org/10.27061/d.cnki.ghgdu.2025.002642. (in Chinese)RU Peng, SU Jun, HAN Zhihong & SHU Quanfeng. (2025). Intelligence leadingthe future: Social impacts of generative artificial intelligence and governance throughstandards. E-Government, (1), 2-14. https://doi.org/10.16582/j.cnki.dzzw.2025.01.001.(in Chinese)XIE Liwei & PAN Sheng. The copyright compliance dilemma of generative AIdata training and paths of governance. Journal of Shandong University of Science andTechnology (Social Sciences), 1-12.https://doi.org/10.16452/j.cnki.sdkjsk.20260617.002. (in Chinese)XU Wei & HE Ye. (2024). The governance system for generative AI data securityrisks and its optimization: A grounded analysis of 38 policy texts. E-Government, (10),42-58. https://doi.org/10.16582/j.cnki.dzzw.2024.10.005. (in Chinese)
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)134YANG Jia. Membership inference attacks on machine learning models andmechanisms of privacy protection [D]. Supervisors: ZHANG Xing; MIAO Chunyu.Liaoning University of Technology, 2026. (in Chinese)YUAN Xiuting & YAN Chi. Some typical cases on copyright protection of AI-generated content. Journal of Ocean University of China (Social Sciences), 1-11. (inChinese)DING Hu, HUA Pengxiang & HUANG Zhen. (2026). Survey on recent progress ofAI for chemistry: methods, applications, and opportunities. Frontiers of ComputerScience, 20(11).TIAN Kaiyuan, QIAO Linbo, LIU Baihui, JIANG Gongqingjian, LI Shanshan &LI Dongsheng. (2026). A survey on memory-efficient transformer-based model trainingin AI for science. Frontiers of Computer Science, 20(11).Shengnan Zhao, Kuiheng Sun, Chuan Zhao, Bendong Jiang, Zhenxiang Chen &Qiuliang Xu. (2025). Vp3CNN: A verifiable privacy-preserving three-party scheme forconvolutional neural network inference. Chinese Journal of Electronics, 34(5), 1427-1437.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)135The Impact of Autonomous DrivingPopularization on Taxi Workers: AMultidimensional Analysis Based on theSubstitution–Creation–Transition FrameworkA.P. WrenzouchiusLos Angeles, California, USAEmail：a.p.wrenzouchius@hotmail.comABSTRACT:As autonomous driving technology iterates rapidly and commercial pilot programscontinue to expand, the taxi industry—one of the labor-intensive service sectors mostimmediately exposed to the risk of autonomous substitution—has seen the employmentshocks facing its workforce become a practical issue of shared concern in both academia andpolicy circles. Drawing on theories of technological unemployment and research on theservice economy, this paper adopts a qualitative method combining literature analysis andpolicy text analysis to systematically examine the impact of autonomous drivingpopularization on taxi workers across four dimensions: employment quantity, employmentstructure, income and the labor process, and skill requirements. The study finds that theimpact presents a composite pattern of “aggregate substitution, structural differentiation,process reconstruction, and skill upgrading”; that the substitution process is constrained bytechnological maturity, regulatory pacing, cost curves, and social acceptance, and thereforeunfolds gradually rather than abruptly; and that the creation of new jobs concentrates inremote supervision, safety operation, vehicle maintenance, and service management, yet asignificant skill mismatch exists between these positions and incumbent drivers. Accordingly,policy recommendations are proposed at the levels of government, enterprises, and theindustry, covering gradual transition, social security backstops, reskilling, and the design oftransition channels.Keywords: autonomous driving; taxi workers; technological unemployment; jobdisplacement; reskilling
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1361.INTRODUCTIONAutonomous driving technology is moving from closed-course testing towardcommercial operation on open roads. In recent years, many countries and regions havesuccessively issued regulations governing the road testing and operation of autonomousvehicles, and driverless taxis (robotaxis) have begun carrying passengers in demonstrationprograms in several cities. China has also elevated intelligent vehicles to a national strategicindustrial direction: in 2020, the National Development and Reform Commission (NDRC)and other departments jointly issued the Intelligent Vehicle Innovation and DevelopmentStrategy, explicitly setting the goal of achieving large-scale production of conditionallyautonomous vehicles by 2025 (NDRC et al., 2020).The taxi industry is regarded as a frontier scenario for the commercial deployment ofautonomous driving because of its relatively controllable operating routes, the high share ofdriver labor in per-vehicle cost structures, and the deep platformization of dispatching. At thesame time, the large population of taxi and ride-hailing drivers constitutes one of theoccupational groups most directly and intensively exposed to the impact of autonomousdriving. In their seminal study, Frey and Osborne (2017) estimated that the probability of taxidrivers’ jobs being computerized exceeds 89%, ranking among the highest of all occupations.Existing studies mostly focus on the technological evolution path of autonomous driving,market size projections, or transportation system effects, while systematic discussion of themicro-level subject—the workers themselves—remains relatively scarce: when substitutionwill occur, in what manner, which segments will be affected, and how workers can transitionstill lack an integrated analytical framework. Accordingly, this paper seeks to answer threequestions. First, to what extent and at what pace will the popularization of autonomousdriving substitute for taxi workers’ existing jobs? Second, will job substitution beaccompanied by the creation of new jobs and the restructuring of employment? Third, howshould governments, enterprises, and the industry design transition mechanisms to reduce thesocial costs of transformation?The significance of this study is twofold. Theoretically, it applies the analyticalframework of technological unemployment to a highly platformized urban mobility service
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)137industry and refines the mechanism chain of “substitution–creation–transition.” Practically, itprovides a reference for relevant authorities in formulating gradual regulatory policies, socialsecurity policies, and vocational training policies.2.LITERATURE REVIEW2.1 Aggregate Estimates of Automation and Job DisplacementRegarding the employment effects of automation, the literature is divided between“substitution pessimism” and “creation optimism.” Based on probabilistic estimates of thecomputerizability of occupational tasks, Frey and Osborne (2017) argued that about 47% ofjobs in the United States fall within a high-risk range—a judgment that sparked extensivedebate. In contrast, after examining task heterogeneity at the worker level, Arntz et al. (2016)found that only about 9% of jobs in OECD countries face a high risk of substitution,significantly lower than occupation-level estimates. From a historical perspective, Autor(2015) observed that while automation replaces some tasks, it also continuously generatesemployment demand by raising productivity, creating new tasks, and strengthening humancomplementarity—“why are there still so many jobs” epitomizes the long-run coexistence oftechnology and employment. Brynjolfsson and McAfee (2014) characterized the current stageas the “second machine age,” arguing that digital technologies are penetrating cognitive andphysical tasks at an unprecedented speed and that the speed and breadth of employmentrestructuring will surpass all previous technological revolutions.Acemoglu and Restrepo (2019) further proposed the analytical framework of“displacement effect–productivity effect–new task creation”: automation directly substitutesfor labor tasks, but productivity gains and the creation of new non-automated tasks partiallyreinstate employment. Using cross-country industry data, Graetz and Michaels (2018) showedempirically that the use of industrial robots raised productivity and wages without reducingtotal employment, though its structural effects were significant—low-skilled workers wererelatively disadvantaged. Manyika et al. (2017) estimated that by 2030, 400 to 800 millionworkers worldwide might have their jobs displaced by automation while substantial new
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)138occupational demand would emerge; large-scale job transitions and retraining thus constitutethe core challenge.2.2 Autonomous Driving and Employment in Mobility ServicesIn the mobility service sector, Fagnant and Kockelman (2015) were among the first tosystematically discuss the macro-level benefits of autonomous vehicles and the requiredpolicy preparations, noting that while reducing accident rates and travel costs, they would alsoexert an impact on professional driver groups. Manyika et al. (2017) ranked “driving” amongthe physical task categories most susceptible to automation and estimated that about 3 millionjobs in the United States with driving as the primary duty face long-term substitution risk.From a public policy perspective, Grush et al. (2018) argued that the cost advantage ofautonomous mobility services would gradually erode the landscape in which private cars andtaxis coexist, and that policy design should proactively incorporate social arrangements fordriver transition.In contrast to the substitution thesis, research from a service interaction perspectiveoffers a different view. Based on frontline service contexts, Wirtz et al. (2018) pointed outthat service segments involving empathy, complex communication, and situational judgmentwill remain difficult for service robots to fully replace in the foreseeable future, suggestingthat the substitution boundary of hybrid “driving plus service” jobs is not monolithic. Thepublic survey by Smith and Anderson (2017) found that most Americans remain hesitantabout riding in fully self-driving cars, indicating that social acceptance constitutes a realisticconstraint on technology diffusion.2.3 Platformized Labor and the Employment Characteristics of Taxi WorkersThe contemporary taxi industry is highly platformized. Rosenblat’s (2018) research onride-hailing platforms reveals the deep penetration of algorithmic management into drivers’labor processes: dispatch rules, dynamic pricing, and rating systems improve matchingefficiency while making drivers’ income increasingly unstable and eroding their laborautonomy. This means that even before autonomous driving substitution occurs, the
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)139employment form of taxi workers has already undergone a round of profound restructuring;the shock of autonomous driving will be superimposed upon these existing vulnerabilities.Litman (2023), in his continuously updated assessment, emphasizes that market penetration ofautonomous vehicles will be a gradual process spanning decades: workers need not panic inthe short term, but policy preparation cannot begin too early.2.4 CommentaryExisting research has established the theoretical framework of “substitution–creation–transition,” made overall judgments on the risk of autonomous driving substituting forprofessional drivers, and characterized industry employment features from the platform laborperspective. However, the limitations of the existing literature are also evident: most studiesdiscuss aggregates at the national or industry level and rarely examine the multidimensionalimpacts on the specific occupational group of taxi workers; the constraints on the pace ofsubstitution lack integrated review; and discussions of transition channels and policyinstruments largely remain at the level of principles. This paper seeks to advance the literaturein these three respects.3.ANALYTICAL FRAMEWORK AND RESEARCHMETHODS3.1 Analytical FrameworkThis paper constructs an analytical framework of “four-dimensional impacts–constraints–transition mechanisms.” The four dimensions of impact include: (1) employmentquantity, i.e., the scale and pace at which existing jobs are substituted; (2) employmentstructure, i.e., the differentiation of job types and the distribution of newly created positions;(3) income and the labor process, i.e., the reshaping of workers’ labor conditions byalgorithmic management and operational model changes; and (4) skill requirements, i.e.,changes in the occupational skill spectrum and their degree of match with the existing humancapital stock. The constraints cover four aspects—technological maturity, regulation andaccess policy, operating cost curves, and social acceptance—which jointly determine the
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)140speed and path of substitution. The transition mechanisms examine workers’ reallocationchannels at the levels of government, enterprises, and the industry.3.2 Research MethodsGiven the forward-looking nature of the topic and the limited availability of micro-levelpanel data, this paper adopts a qualitative research method combining literature analysis andpolicy text analysis: journal articles from authoritative databases and reports frominternational organizations and consulting institutions serve as the literature sources, whileregulatory policy texts on autonomous driving in China and other major countries serve as theobjects of policy analysis; thematic coding is used to induce the impact mechanisms andpolicy essentials. This method is suitable for mechanism identification and propositionrefinement; its limitations are explained in the conclusion.4.DIRECT IMPACTS OF AUTONOMOUS DRIVING POPULARIZATION ONTAXI WORKERS4.1 Employment Quantity: Aggregate Substitution Pressure and a Gradual PaceIn terms of substitution potential, the high routinization of driving tasks places themdurably within the high-risk range of task-based substitution models (Frey & Osborne, 2017),and the core task chain of taxi operation—“driving plus passenger service”—is precisely oneof the most mature directions for the commercial deployment of autonomous driving (Grushet al., 2018). If the per-kilometer cost of driverless taxis falls below that of human-drivenvehicles at scale, market forces will drive the contraction of existing jobs—a process that canoccur spontaneously, without administrative mandates.In terms of pace, however, substitution will be gradual rather than abrupt. First,technological constraints: complex weather, non-standard road conditions, and long-tailscenarios still require time to resolve (Fagnant & Kockelman, 2015). Second, regulatoryconstraints: licensing for passenger operations, determination of accident liability, andinsurance systems are still under construction. Third, cost constraints: sensor and vehiclecosts, together with the cost of configuring remote safety operators, mean that the cost
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)141advantage of fully driverless operation is not yet secure in the short term (Litman, 2023).Fourth, social constraints: public trust in fully driverless vehicles takes time to build (Smith &Anderson, 2017). Synthesizing existing projections (Manyika et al., 2017; Litman, 2023), thefollowing judgment can be made: over the next 5–10 years, substitution will mainly occur indemonstration operations in pilot cities and in specific scenarios (such as airport shuttle linesand fixed areas), with limited aggregate impact; between 10 and 20 years, as cost curvesdecline and regulation matures, substitution will enter an accelerating phase, at which pointthe job-transition pressure on incumbent drivers will become concentrated and pronounced.4.2 Employment Structure: Job Differentiation and the Distribution of NewPositionsSubstitution is not merely the elimination of jobs; it is accompanied by the restructuringof the occupational structure. Following the new-task framework of Acemoglu and Restrepo(2019), autonomous mobility services will spawn several non-automated new positions: first,remote supervision and safety operation positions, including remote safety operators, dispatchcoordinators, and emergency response specialists; second, vehicle operation and maintenancepositions, covering the calibration, testing, cleaning, and charging maintenance ofautonomous driving systems; and third, on-site service positions, such as vehicle inspection,in-cabin order management, and passenger assistance. The projections of the WorldEconomic Forum (2023) likewise indicate that the green transition and the application ofartificial intelligence will create large numbers of new jobs while displacing others, but theskill composition of the new jobs differs significantly from that of the displaced ones.For taxi workers, the problem lies in structural mismatch: the new positions eitherrequire system maintenance and data processing skills or stronger emergency response andcommunication abilities, while the skill stock of incumbent drivers—centered on drivingskills—cannot be directly transferred. Older and less-educated groups face particularlynarrow re-employment channels, making them the most vulnerable groups in this structuraldifferentiation and the ones most in need of policy attention.
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1424.3 Income and the Labor Process: A Second Restructuring from AlgorithmicManagement to Human–Machine CollaborationBefore full substitution occurs, autonomous driving technology has already penetratedthe labor process in the form of “human–machine collaboration.” The deepening of assisteddriving functions, intelligent dispatching, and dynamic pricing increasingly embeds drivers’labor within algorithmic systems. The logic of algorithmic management revealed byRosenblat (2018)—data-driven performance evaluation and opaque rule changes—will befurther strengthened in the era of autonomous driving: when the vehicle itself becomes theexecutor of algorithms, the human role is compressed into that of “exception handler” and“service supplement,” and both bargaining power and professional identity may be weakened.At the income level, a two-stage pattern emerges. During the transition period, someplatforms may reshape positions under the title of “safety operator,” but such positions oftenoffer lower hourly wages and more demanding work arrangements than traditional drivingjobs, and they themselves face a “second substitution” risk as the technology matures. Afterthe acceleration phase of substitution, workers who fail to transition face the risk of incomeinterruption, while those who successfully move into new positions in maintenance andsupervision may see their income curves stabilize or even rise. Income differentiation willthus be isomorphic with skill differentiation.4.4 Skill Requirements: The Leap from Driving Skills to Composite SkillsThe popularization of autonomous driving will systematically reshape the skill spectrumof mobility service occupations. Traditional positions center on driving operation skills,supplemented by route experience and basic service abilities; the new positions feature acomposite structure of “technical understanding plus emergency response plus servicecommunication.” The research of Wirtz et al. (2018) suggests that service content involvingemotional labor and complex interaction remains a comparative advantage of humans in thelong run, which means that the depreciation of “driving skills” does not equal the depreciationof “service capabilities”—workers’ transition paths do not all lead to technical positions; a
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)143considerable proportion will shift toward service positions centered on interpersonalinteraction.For individuals, reskilling faces three obstacles: information asymmetry between trainingsupply and job demand, the opportunity cost incurred during training, and the constraints ofage and learning curves. Without external support, market-based training will systematicallyfavor younger and more educated groups, exacerbating the distributional inequality of thetransition (Manyika et al., 2017).5.INDIRECT AND SYSTEMIC IMPACTS5.1 Transmission Effects on the Labor MarketThe scale and mobility of the taxi driver population mean that the shock will not beconfined to the industry itself. This group consists largely of rural migrant workers andflexible workers with low social security participation rates and weak risk resilience; oncesubstitution accelerates, the shock will transmit outward along the chain of “householdincome–consumption–local service industries.” Moreover, professional drivers constitute oneof the most emblematic cases of automation-era displacement, and the success or failure oftheir transition will become a reference frame for public evaluation of automation policies,carrying social-psychological significance that extends beyond the industry.5.2 Impacts on Urban Transportation and Informal EmploymentIf autonomous mobility services significantly reduce travel costs, they may inducegrowth in total travel demand and thereby create some derivative service demand (Fagnant &Kockelman, 2015); they may also squeeze the space of informal employment such asunlicensed taxis and motorcycle taxis, altering the stock structure of urban informalemployment. In addition, a “dual-track” pattern may emerge during the industry transition—human-driven and driverless vehicles coexisting for an extended period under asymmetriccompetitive rules—and how to prevent incumbent workers from being cleared out too quicklyunder such asymmetric competition is a question that regulatory design must address (Grushet al., 2018).
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)1446.POLICY RECOMMENDATIONS6.1 Government Level: Gradual Transition and Social Security BackstopsFirst, implement gradual management of the substitution pace. When issuing driverlessoperation licenses, set protective clauses for incumbent workers, such as requiring operatorsto allocate transition positions proportionally and establishing transition periods andregionally phased substitution arrangements, so as to avoid “shock-style” clearance. Second,strengthen social security backstops. In view of the low insurance participation rate of thedriver population, expand occupational injury insurance coverage for flexible workers andprovide seamless linkage between unemployment insurance and reemployment services forthose who lose their jobs during the transition. Third, build a reskilling system. Includeprofessional drivers in special job-transition training programs, develop curricula around threedirections—remote safety, vehicle maintenance, and passenger service—and provide trainingand living subsidies to reduce opportunity costs. Fourth, improve the recognition of newoccupations. Incorporate emerging positions such as remote safety operators and autonomousdriving maintenance technicians into the occupational classification and skill certificationsystems, opening up career development channels for workers.6.2 Enterprise Level: Internal Reallocation and Shared ResponsibilityOperators and technology service enterprises should incorporate labor transition intotheir commercialization plans: prioritize recruiting remote safety operators and on-site servicestaff from incumbent drivers, and establish an internal transition ladder of “driver–safetyoperator–maintenance technician”; retain appropriate human–machine collaborationinterfaces in technical designs to slow the pace at which full driverlessness compresses jobs.Meanwhile, enterprises should build targeted training mechanisms jointly with vocationalcolleges, internalizing the costs of transition as human capital investment rather thanexternalizing them as social problems.7.DISCUSSION AND CONCLUSION
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)145Based on the “substitution–creation–transition” framework, this paper has analyzed theimpact of autonomous driving popularization on taxi workers across four dimensions:employment quantity, employment structure, income and the labor process, and skillrequirements. Three core conclusions emerge. First, substitution is certain, but its pace isgradual: the fourfold constraints of technology, regulation, costs, and social acceptancedetermine a substitution path of “pilot scenarios first, stock markets later.” Second, theessence of the shock is structural: aggregate substitution and job creation proceed in parallel,and the mismatch between the skill requirements of new jobs and the existing human capitalstock constitutes the principal aspect of the contradiction. Third, the social costs of thetransition can be significantly reduced through policy design: gradual transition, securitybackstops, and reskilling constitute the three pillars of the policy mix.This study has the following limitations: literature analysis and policy text analysis lackfirst-hand survey data from workers, and the impact mechanisms await empirical testing;judgments about the pace of substitution rely on existing projection literature and carry highuncertainty; and the comparative perspective is confined to the taxi industry, excluding otherdriving occupational groups such as bus and freight drivers. Future research can examineemployment elasticities and policy effects during the substitution process through trackingsurveys and quasi-natural experiment designs, providing evidence for dynamically adjustingthe policy mix.REFERENCESAcemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technologydisplaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30.https://doi.org/10.1257/jep.33.2.3Arntz, M., Gregory, T., & Zierahn, U. (2016). The risk of automation for jobs inOECD countries: A comparative analysis (OECD Social, Employment and MigrationWorking Papers No. 189). OECD Publishing. https://doi.org/10.1787/5jlz9h56dvq7-en
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)146Autor, D. H. (2015). Why are there still so many jobs? The history and future ofworkplace automation. Journal of Economic Perspectives, 29(3), 3–30.https://doi.org/10.1257/jep.29.3.3Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress,and prosperity in a time of brilliant technologies. W. W. Norton.Fagnant, D. J., & Kockelman, K. (2015). Preparing a nation for autonomousvehicles: Opportunities, barriers and policy recommendations. Transportation ResearchPart A: Policy and Practice, 77, 167–181. https://doi.org/10.1016/j.tra.2015.04.003Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptibleare jobs to computerisation? Technological Forecasting and Social Change, 114, 254–280. https://doi.org/10.1016/j.techfore.2016.08.019Graetz, G., & Michaels, G. (2018). Robots at work. The Review of Economics andStatistics, 100(5), 753–768. https://doi.org/10.1162/rest_a_00754Grush, B., Niles, S., & Baum, G. (2018). The end of driving: Transportationsystems and public policy planning for autonomous vehicles. Elsevier.https://www.sciencedirect.com/book/9780128154519/The-End-of-DrivingLitman, T. (2023). Autonomous vehicle implementation predictions: Implicationsfor transport planning. Victoria Transport Policy Institute.https://www.vtpi.org/avip.pdfManyika, J., Chui, M., Miremadi, M., Bughin, J., George, K., Willmott, P., &Dewhurst, M. (2017). Jobs lost, jobs gained: Workforce transitions in a time ofautomation. McKinsey Global Institute. https://www.mckinsey.com/mgi/our-research/jobs-lost-jobs-gained-workforce-transitions-in-a-time-of-automationNational Development and Reform Commission, Cyberspace Administration ofChina, Ministry of Science and Technology, et al. (2020). Intelligent vehicle innovationand development strategy. https://www.gov.cn/xinwen/2020-02/24/content_5482644.htmRosenblat, A. (2018). Uberland: How algorithms are rewriting the rules of work.University of California Press. https://doi.org/10.1525/9780520970632
International Journal of Responsible Artificial Intelligence Research ISSN:3106-857X(print), 3106-8588(online)147Smith, A., & Anderson, M. (2017). Automation in everyday life. Pew ResearchCenter. https://www.pewresearch.org/internet/2017/10/04/automation-in-everyday-life/Wirtz, J., Patterson, P. G., Kunz, W. H., Gruber, T., Lu, V. N., Paluch, S., &Martins, A. (2018). Brave new world: Service robots in the frontline. Journal of ServiceManagement, 29(5), 907–931. https://doi.org/10.1108/JOSM-04-2018-0119World Economic Forum. (2023). The future of jobs report 2023.https://www3.weforum.org/docs/WEF_Future_of_Jobs_2023.pdf
