Executive Summary

Relying on the team's self-developed dynamic Chinese C-O*NET occupational information database, and adopting a task-centric analytical perspective, the report China's Labor Market in the AI Era establishes a four-quadrant coordinate system based on the dual dimensions of AI exposure and human complementarity. It systematically deconstructs the impact logic and reconfiguration pathways of Large Language Models (LLMs) on China's labor market.

The report compiles a dynamic C-O*NET occupational information database, with a raw data foundation covering 743 million original recruitment records from mainstream hiring platforms between 2022 and April 2026. After engineering processing involving deduplication, balanced sampling, LLM semantic clustering, and full regular expression backfilling, it consolidates 65,200 standardized task entries. This covers 1,267 standardized segmented occupations nationwide, enabling micro-task-level tracking of the labor market through monthly updates. This database overcomes three core shortcomings of the static U.S. O*NET: time lag, insufficient granularity, and incompatibility with China. During the observation period, isolating macroeconomic and industry cycle interferences, the report calculates the dual gaps—current and cumulative—in recruitment demand caused by AI. It clarifies that AI does not simply "eliminate jobs," but rather inverses the logic of the division of labor through task unbundling and responsibility repackaging. Furthermore, centering on the structural challenge of supply-demand mismatches in undergraduate education, the report formulates a "five-in-one" quantitative assessment and innovative solution plan. This provides quantitative reference support for economic research, higher education innovation, and the formulation of employment and macroeconomic policies.

Part 1. Research Foundation: Construction of the Dynamic C-O*NET Database and Core Analytical Framework

1.1 Flaws of Traditional Research Paradigms

For a long time, domestic and international research on AI and employment has generally relied on the static O*NET database of the U.S. Department of Labor. This system relies on offline questionnaires from incumbents and manual evaluations by industry experts as data sources; version 28.2 contains 19,281 standardized tasks and approximately 900 standard U.S. occupational categories, serving as the universal public foundation for global academia. However, adapting it to China's labor market encounters two insurmountable "isolation walls":

• The Barrier of Temporal Resolution: O*NET's complete revision cycle takes 3 to 5 years, only recording stable job tasks from years ago, making it unable to capture the real-time rewriting of job descriptions brought about by LLM iterations. Since the comprehensive commercialization of generative AI in late 2023, domestic job descriptions (JDs) and hiring requirements have exhibited significant monthly iteration characteristics, leaving static questionnaires entirely lagging behind actual market changes.
• The Barrier of Granularity and Localization: O*NET outputs a single average profile for an occupation, flattening out differences in task structures across different cities, enterprise sizes, salaries, years of experience, and educational backgrounds. For example, the task combinations for accounting positions in small micro-enterprises providing bookkeeping services are entirely different from those in listed companies, yet the database only outputs a unified mean. Meanwhile, quantitative calculations show that about 38% of the core screening conditions in domestic employers' recruitment texts cannot be covered by the O*NET scale, meaning traditional overseas databases cannot adapt to the true screening logic of the domestic job market.

At the same time, public opinion and popular research have fallen into the misconception of the "job disappearance theory," simply equating AI impact with the total elimination of certain occupations, which creates anxiety but offers no solutions. The core logical flaw lies in confusing "occupational labels" with "task collections": a job title is merely a packaging container for tasks, and the smallest unit of AI impact is text, calculation, and process tasks, not a complete occupation. A direct comparison between two typical roles—Java Developer and Telephone Customer Service—intuitively verifies this:

• Both fall into the high AI exposure range, but almost all scripting and work order tasks of the customer service role can be taken over by models, yielding a very low human complementarity score.
• For Java Developers, while a large amount of code generation and document writing can be completed by AI, high-human-judgment tasks such as architecture design, online fault handling, and cross-business system coordination are irreplaceable.
• The two diverge completely; relying solely on job titles to classify risk leads to systematic misjudgment.

The report establishes a core thesis: the fundamental unit of AI impact is the task, not the complete occupation; tracking market changes requires observing job req/res (requirements/responsibilities) entries first, as occupational title adjustments have significant time lags.

1.2 Full Process of the Engineering Construction of China's Dynamic C-O*NET

To overcome the limitations of static overseas databases, the report built a localized, dynamically updatable monthly C-O*NET system. The pipeline distinguishes between the application boundaries of LLMs and full-volume data processing tools. It adheres to the core engineering principle: "LLMs are only used for sampling, clustering, tagging, and regex generation, while massive recruitment text full-volume backfilling relies on 106,000 self-developed regular expressions + the Hyperscan high-speed scanning engine." The complete ten-step production chain is as follows:

• Step 1: Raw Data Collection. Aggregating all original recruitment records from platforms from 2022 to April 2026, totaling 743 million rows.
• Step 2: Text Deduplication. Hash algorithms are used to remove duplicate JDs, retaining 116 million unique job descriptions.
• Step 3: Stratified Balanced Sampling. Sampling with upper limits stratified by occupation, year, and platform yields 2.31 million representative samples, avoiding data monopoly biases from top industries and tech giants.
• Step 4: LLM Atomic Phrase Extraction. LLMs are called only for the 2.31 million samples to split original short sentences of standardized job requirements (req) and job responsibilities (res), forming 2.9 million req phrases and 7.33 million res phrases after deduplication.
• Step 5: Semantic Clustering and Loose Merging. Vector embedding recall is used for synonymous texts; the LLM completes 520,000 pairwise semantic judgments, merging 200,000 sets of synonymous entries, ultimately condensing into 20,721 standardized req buckets and 44,479 standardized res buckets, totaling 65,200 task entries.
• Step 6: Construction of a Dual-Axis Classification System. The reqs are divided into 16 primary categories and 181 subcategories. The res entries feature an original "business domain × execution action" dual-axis architecture (32 domains + 26 actions, totaling 832 grids, with 697 valid sets), solving the traditional single-axis classification flaw that confuses "maintaining clients" with "maintaining equipment."
• Step 7: Multi-Dimensional Continuous Scoring. The GPT-5.5 model is selected to synchronously score all 65,200 entries across thirteen 0–1 continuous dimensions. Compared against 12 mainstream LLMs, this model had a mean MAE of 0.0635, providing the optimal balance of speed and cost.
• Step 8: Batch Generation of Regular Expressions. Five models generate matching rules in parallel; a choose-one-from-five verification process produces 106,000 regular expressions.
• Step 9: Full-Volume Backfill Scanning. The Hyperscan engine processes 1,460 JDs per second, annotating all 13-dimensional attributes for the 116 million unique jobs.
• Step 10: Multi-Dimensional Tag Mounting. Each JD is linked with comprehensive tags—city, industry, salary, education, years of experience, enterprise size, and major flow—forming a dynamically disaggregable database.

The entire data system underwent manual verification, achieving a Pearson correlation coefficient of 0.947 between regex matching and manual annotation; the data reliability meets the requirements for macro, industry, and major stratified calculations. The database naturally retains monthly timestamps, capable of tracking the complete cycle of any task or responsibility bundle from niche to ubiquitous, providing micro-evidence for calculating the time series of AI impact. The 13 core scoring dimensions constitute the entire quantitative foundation of the report, divided into two opposing axes and auxiliary observation indicators:

• Horizontal Axis: LLM Exposure. The proportion of tasks AI can complete independently, acting as the core driver of contraction.
• Vertical Axis: Human Complementarity. The degree of reliance on physical presence, judgment, and responsibility, hedging against substitution risks.
• Auxiliary Dimensions: Includes 11 other indicators, such as physical presence dependency, embodied craftsmanship, error consequences, routine encodability, cognitive complexity, and public interaction. The driving forces were verified through a correlation "horse race" experiment.

1.3 Core Four-Quadrant Coordinate System and Conclusions from the 13-Dimensional "Horse Race"

The report establishes a standard analytical framework of "LLM Exposure (horizontal axis 0–1) × Human Complementarity (vertical axis 0–1)", dividing the labor market into four major sectors:

• Substitution Pressure Zone (High Exposure, Low Complementarity): Standardized text, processes, data entry, and basic content roles. AI can take over a massive amount of core output, leaving a narrow space for human value addition, resulting in continuously contracting hiring demand. Representative roles: Accounting Specialist, Basic Administration, Telephone Customer Service, Junior Copywriter.
• Collaborative Upgrade Zone (High Exposure, High Complementarity): Technical, management, and engineering roles. AI serves as an efficiency tool to amplify human judgment; demand shows strong resilience, and salaries for senior positions are trending upward. Representative roles: Java Engineer, Quality Management, Key Account Sales.
• Human Protection Zone (Low Exposure, High Complementarity): On-site, caregiving, practical operation, and high-responsibility roles. Physical presence and interpersonal trust are irreplaceable, leading to long-term stable expansion. Representative roles: Nurse, Freight Driver, Security Guard, Domestic Worker.
• Low Impact, Low Growth Zone (Low Exposure, Low Complementarity): Traditional standardized physical labor. AI penetration is slow, but salary and career advancement opportunities are limited. Representative roles: General Worker, Sorting and Delivery Worker.

To pinpoint the core variables truly driving labor market changes, this research conducted a correlation "horse race" of the 13 scoring dimensions at the occupational level, comparing the change in hiring proportion from the 2023 Q4 base period to January–April 2026. Positive and negative correlation coefficients represent expansion/contraction trends, yielding the following conclusions:

• The strongest indicator on the contraction side is LLM Exposure ($r = -0.47$), far exceeding the runner-up, Expression Ambiguity ($r = -0.32$).
• The two core drivers on the market expansion side are Embodied Craftsmanship ($r = +0.55$) and Physical Presence Dependency ($r = +0.53$).

The essence of all market changes lies on one core axis: whether a task can be independently and fully undertaken by an LLM; the remaining dimensions are all auxiliary correction variables. The definition of the LLM Exposure score is explicit: the average score of all tasks for a single occupation equals the theoretical maximum workload of that position that can be substituted by AI, which serves as the underlying baseline for the subsequent gap calculation. Meanwhile, experiments verify that "high AI exposure does not mean job disappearance"; significant declines in hiring scale only occur when paired with low human complementarity. Solely relying on one-sided AI exposure to determine job risk carries a massive margin of error.

Part 2. Market Impact Brought by AI: Exposure Scale, Affected Groups, and Three-Phase Gap Calculation

2.1 Overall Scale of AI Exposure in National Online Recruitment

The report's baseline observation window encompasses approximately 40.01 million total online hiring demands from January to April 2026. The study sets loose, neutral, and strict score thresholds to test the robustness of the conclusions, with the neutral standard acting as the full-text baseline: 33.2% of total demand (about 13.27 million entries) belongs to high-AI-exposure task combinations. Among these, demand featuring high exposure and low human complementarity—falling into the substitution pressure zone—accounts for 26.6% (about 10.63 million entries); that is, 1 in every 4 hiring demands sits in the high-risk AI substitution range.

The structural distribution exhibits clear geographical and industry differentiation: The proportion of high-substitution roles in new Tier-1 and Tier-2 cities is higher than in Tier-1 cities (where high-end services and offline interpersonal roles dilute the risk). At the industry level, highly text- and process-oriented industries such as basic customer service, content copywriting, administration, financial accounting, and junior operations have the largest exposure. Substitution exposure is significantly lower in manufacturing, healthcare, and offline service industries. The core boundary of this metric must be clarified: a tens-of-millions-level exposure scale only represents risk exposure, not the actual number of unemployed individuals; online recruitment data represents new incremental traffic, not the national employment stock, and cannot be directly extrapolated to total macroeconomic unemployment.

2.2 Stratification of AI Impact: Concentrated at Career Entrances and Mid-Range Groups

Under the same AI exposure level, the magnitude of hiring contraction exhibits vast stratified differences across groups with varying salaries, years of experience, and educational backgrounds. The degree of impairment is quantified via grouping regression slopes:

• Salary Dimension: The 5k–8k mid-range entry-level positions have a slope of -0.31, marking the deepest contraction in the entire market. The slope turns positive in the 12k+ salary range, and high-end positions over 50k have a correlation coefficient of +0.19; demand for senior positions with high AI exposure is actually expanding.
• Years of Experience: Workplace newcomers with 1–3 years of experience have a slope of -0.25, suffering the most severe damage. Mature practitioners with 5–10 years of experience have a slope of only -0.05, remaining almost unaffected by the AI demand contraction.
• Education Level: Associate degrees and below have a slope of -0.23, Bachelor's degrees -0.16, and Master's degrees -0.11. The higher the education, the stronger the buffering effect, though no educational level is completely immune to the impact.

The underlying mechanism is clear: core tasks for entry-level roles primarily consist of document entry, basic accounting, standardized communication, and other easily AI-automated work. The traditional corporate task chain, which relied on newcomers completing basic training, is being directly extracted by models. Employers no longer offer low-paying practice roles, creating an "entrance trap." Occupational level data shows the median proportion of low-experience roles is as high as 92.2%; while the nominal hiring door hasn't closed, entry thresholds continue to rise, making complex tasks a rigid requirement for fresh graduates upon onboarding. The median experience reward multiplier is 1.53 times, indicating that long-term workplace accumulation still commands a stable premium. The market is weeding out mid-range entry-level workers who only master standardized basic work without the ability to provide judgment-based safety nets. Typical shrinking roles include basic copywriters, photo retouchers, telemarketers, and junior accountants. Expanding roles are concentrated in strong-embodiment, high-responsibility positions such as on-site practical operations, clinical diagnosis, high-end technical coordination, and physical equipment debugging.

2.3 Three-Phase Recruitment Gap: Full Cycle Calculation of Initiation, Overshoot, and Stabilization

Using September to December 2023 as the pre-AI interference baseline zero point, the report selects on-site roles with the lowest AI exposure as the natural control group. Stripping away macroeconomic cycle impacts, it distinguishes between two core calculation metrics: the current gap (annual traffic fluctuation) and the cumulative gap (total reduction in stock), completely recording the three-year market adjustment cycle from 2024 to 2026:

• 2024 (Initiation Phase): Current gap 2.6%, cumulative gap 1.0%, with a cumulative reduction of 940,000 hiring entries. Enterprises merely piloted AI tools on a small scale without massively reconstructing roles; the substitution effect remained on paper, with the theoretical substitution upper limit stable at 41.8%.
• 2025 (Overshoot Phase): Current gap 30.1%, cumulative gap 10.1%, with a cumulative reduction of 9.05 million hiring entries. Enterprises panic-cut full-process standardized roles; in some industries, the scale of layoffs even exceeded the proportion of tasks models could undertake, resulting in a clear overshoot. The theoretical upper limit was 35.6%.
• 2026 (Stabilization Phase): The current gap fell back to 16.9%, while the cumulative gap rose to 15.1%, with a cumulative reduction of about 13.59 million hiring entries. Enterprises completed process reconstruction, distinguishing between "AI-assisted roles and human-machine collaborative roles," and resumed expanded hiring for high-level composite roles. The existing gap continued to accumulate unidirectionally.

Three core objective laws can be observed:

• First, the theoretical workload that LLMs can substitute has remained stable in the 35%–42% range long-term; the capability ceiling of large models has seen no drastic changes, with fluctuations stemming solely from corporate hiring decisions.
• Second, the cumulative gap only increases and never decreases, while the current gap floats up and down with corporate trial and error.
• Third, structural stratification is prominent: Tier-1 cities have the highest cumulative gaps, and the cumulative reduction scales for copywriting, translation, basic customer service, and junior software roles lead all industries; lifestyle services and on-site manufacturing industries have the lowest gaps.

Crucial metric constraints: All gaps represent the "magnitude of demand reduction for high-exposure occupations relative to low-exposure control groups." This only represents the transfer of job demand and cannot directly equate to the total loss of employment in society; the online sample lacks sufficient coverage of blue-collar and county-level regions, indicating sampling bias.

Part 3. Employment Reconfiguration Brought by AI: Responsibility Reorganization and the Inversion of the Division of Labor Logic

3.1 Core Fact: Occupations Shrink, but Responsibilities Do Not Synchronously Disappear

Intuitive market recruitment data shows a decline in hiring volumes for a massive number of basic positions. However, three sets of quantitative evidence—res (responsibility) entropy, cross-industry diffusion of responsibilities, and dual-layer slopes of occupation/responsibility—confirm that the standardized tasks extracted by AI have not completely disappeared. They have merely detached from exclusive positions and diffused across industries into other composite roles; there is no large-scale phenomenon of task extinction.

• Evidence 1: Occupational internal res entropy continues to rise. Entropy values represent the richness of job responsibilities; an entropy increase means a single position shifts from a singular specialization to one person holding multiple duties. The entropy increase amplitude for crew members, process engineers, and operations roles leads the pack. After AI strips away simple inspections and data sorting, enterprises package exception handling, cross-departmental coordination, and proposal reporting into the same role. The boundaries of roles are continuously blurring, and the internal task variance within a single job title continues to expand.
• Evidence 2: High LLM-executable responsibilities diffuse across industries with a weighted correlation coefficient of +0.52. Document handling, data analysis, and first-draft writing are no longer exclusive to copywriting and data roles. Instead, they are embedded across all roles, including product, production, operation and maintenance, and sales; specialized single-duty roles are losing their unique value.
• Evidence 3: Comparison of stratified slopes reveals the contraction slope at the occupational level is $\beta = -1.33$, whereas the slope at the responsibility level is significantly flatter. In the top quartile of high-exposure occupations, 56% of roles are shrinking at a pace far exceeding the contraction rate of their corresponding task pools, a characteristic present in only 33% of general occupations. Two typical types of contraction can be distinguished:
  • Task-based contraction (e.g., copywriting and planning): tasks diffuse across the entire industry, and the exclusive role disappears.
  • Industry-specific contraction (e.g., freight driving): the total volume of tasks is still growing, and the demand decline is only affected by industry cycles. The response paths for the two are entirely different.

3.2 The New Locus of Creation: Composite Closed-Loop Function Bundles, Not Entirely New Occupations

Public opinion generally expects AI to spawn a massive number of brand-new job titles, but the data yields the opposite conclusion. Between 2023 and 2026, the volume of newly added job titles within the recruitment pool was extremely small; only niche tracks like LLM algorithms, AI products, and robotics algorithms saw minor expansion. Concurrently, traditional algorithm roles such as SLAM, recommendation, and risk control continued to shrink. The overall recruitment pool for algorithms turned negative, indicating structural rotation within the AI industry rather than absolute growth. The fastest-expanding roles across the market are all transformations of traditional occupations, with no entirely new occupational categories born.

The true carrier of AI creation is a composite function bundle that merges the responsibilities of multiple people into a single-person closed loop. This report catalogs the Top 10 high-growth responsibility bundles; the highest-frequency combination, "manufacturing optimization + quality closed loop," appeared cumulatively 171,480 times, covering 465 types of occupations. All follow the closed-loop structure of "technical execution + problem diagnosis + cross-departmental reporting." Enterprise hiring screening standards have undergone a fundamental transformation: shifting from assessing single-point professional skills to evaluating the ability to deliver complete, closed-loop tasks. The competitiveness of job seekers who only know single standardized operations continues to decline.

Emerging responsibility bundles possess clear structural characteristics, confirmed by a ten-dimensional metric comparison between new and old responsibility bundles. Emerging combinations score significantly higher in physical presence dependency (0.59 vs. 0.31, +0.28), embodied craftsmanship (+0.19), and error consequences (+0.09). Conversely, LLM exposure (-0.11) and public interaction (-0.36) are drastically reduced. New work scenarios have shifted from pure online document processing to physical sites, equipment operations, and risk judgment scenarios. Pure scripting and pure text public reception tasks have been entirely taken over by models, and corresponding roles are no longer being created. The beneficiaries are composite talents who master professional technology + on-site handling + multi-party coordination; the disadvantaged are narrow-role practitioners responsible only for single online standardized processes, lacking on-site and judgment capabilities.

3.3 The Inversion of the Division of Labor Logic: From Smith's Pin Factory to Single-Person AI Closed Loops

The report asserts that the division of labor paradigm is undergoing a disruptive transformation. Adam Smith's pin factory relied on splitting processes to lower the cognitive task-switching costs for humans; the Babbage Principle refined skill grading to suppress labor costs; and Coase posited that internal corporate coordination replaces market transactions, driving continuous refinement of the division of labor for over two centuries. Generative large models are rewriting the core constraint—the marginal cost of task switching approaches zero, causing the Babbage Principle to run in reverse.

The underlying constraint of historical division of labor was the depletion of human attention during task switching, leading enterprises to divide work and assign specialized roles to boost efficiency. Today, models can switch costlessly among coding, copywriting, and data analysis tasks; without the need to split human labor, a single person paired with AI can complete a chain of tasks that previously required multi-person division of labor. This births a new "AI-assisted single-person closed loop" work model—akin to the self-sufficient Robinson Crusoe economic model, but scaling via digital tools and external markets. A fundamental shift in the division of labor has occurred: the division of labor between humans is contracting, while the division of labor between humans and AI is expanding.

Impact on organizational hierarchies:

• First, basic entry-level specialized positions continue to disappear, leaving newcomers unable to rely on simple tasks to accumulate experience.
• Second, the moats of mid-sized coordinative knowledge-service enterprises are thinning; business models that rely purely on splitting and integrating professional human labor are under pressure.
• Third, the scale advantages of top-tier enterprises in capital, computing power, compliance, and physical industries remain unchanged.
• Fourth, personal career development logic is reconstructed: professionals can no longer dig deep into a single niche skill but must build a composite capability framework of "professional practical operation + AI tool calibration + risk and responsibility judgment."

The inversion of the division of labor brings a core societal pain point: the training chain for newcomers within traditional enterprises has broken. Basic practical training, historically borne by employers, must now be front-loaded into higher education and vocational training stages, which is also the root cause of educational supply-demand mismatches.

Part 4. The Dual Mismatch in Undergraduate Education Supply and Demand: Market Signal Failure

4.1 Severe Imbalance Between Major Supply and AI Impact Quadrants

The education side of the report focuses strictly on the undergraduate level. By aggregating data on 240 university major tiers and 189 standard major categories for college admissions across multiple provinces, and calculating the AI four-quadrant coordinates weighted by the occupations graduates flow into, it reveals that over half of the enrollment volume for the undergraduate class of 2025 falls into the substitution pressure zone (high exposure, low complementarity). Only a small number of medical and on-site engineering majors fall into the human protection zone. Enrollment in the collaborative upgrade zone accounts for a mere 2.7%, indicating a severe shortage in enrollment scale for premium human-machine collaborative tracks.

Market demand is decoupled from salary signals: The correlation coefficient between changes in hiring demand for majors and changes in salary is extremely low. Majors with expanding demand do not necessarily see salary increases; demand for a large number of business and liberal arts roles has expanded, but average starting salaries continue to trend downward. The root cause lies in the concentrated influx of graduates into high-exposure entry-level positions, diluting the salary premium. At the city level, across 590 city-major matching samples, the median supply-demand matching index is only 0.44, signifying extremely low alignment between local enrollment structures and actual corporate hiring needs.

4.2 Dual Mismatches in Time and Direction: Cutoff Scores Fail to Reflect Employment Prospects

• The First Mismatch (Direction): Using changes in admission rankings across 240 undergraduate majors to gauge enrollment popularity, the correlation coefficient between popularity and changes in market hiring demand is only 0.23, with a coefficient of determination of roughly 5%. College entrance examination cutoff scores and admission rankings are nearly useless for predicting job quality four years later. A large number of high-scoring popular majors are concentrated in the high-AI-exposure substitution quadrant, while unpopular on-site and medical majors possess long-term stable employment returns, forming a distinct overvaluation/undervaluation dichotomy.
• The Second Mismatch (Time): The undergraduate cultivation cycle is four years. Universities enroll and candidates fill out their preferences based only on current 1–2 year market salary heat; given the three-year full cycle of market AI task structure changes, by the time students graduate, job demand has been completely rewritten. The correlation between synchronous enrollment and market changes is significant, but the lag correlation across four years drops almost to zero, forming a typical cobweb model of supply and demand fluctuation. Verification by region and institutional tier shows: local candidates in low- to mid-GDP cities respond extremely strongly to short-term employment signals, while 985 institutions are anchored by brand reputation, with enrollment scales hardly adjusting to market warmth or coldness, thereby exacerbating structural regional talent surpluses.

4.3 Critical Null Result: Cutoff Scores for High-AI-Exposure Majors Remain Unchanged Over Seven Years

Complete admissions data spanning the seven years from 2019 to 2025 shows that the difference in enrollment popularity between high-LLM-exposure majors and low-exposure majors remains stable between -1.3 to -1.6 percentage points, showing no significant downward inflection point. In 2025 (the second year of massive AI realization), cutoff scores and application popularity have not seen any adjustments, forming a decisive null result: tens of millions of market hiring gaps have materialized, yet test-takers, parents, and universities have not adjusted their major selections based on the AI impact.

The underlying reasons are twofold:

• The public judges majors solely based on major names, completely ignoring the structural changes in the tasks graduates will face. From 2022 to 2026, the LLM exposure scores for the export tasks of numerous computer science, finance, and journalism majors continued to climb, but admission scores and enrollment scales did not shrink; educational signals lag behind the labor market.
• Simultaneously, the exports for undergraduate majors are extremely dispersed. The top five matched occupations for a single major only account for 20%–35% of total demand, while 65%–80% of roles are distributed in long-tail industries. Universities' training programs focus only on a few matched occupations, neglecting long-tail composite responsibility demands, leaving curriculum systems disconnected from real work requirements.

Part 5. How to Promote Employment-Oriented Educational Innovation

5.1 Market Capability Gaps: Directly Converting Responsibility Bundles into Curriculum Modules

Relying on the Top 10 high-growth composite responsibility bundles, the report completed the translation from market tasks to teaching content, classifying differentiated curriculum innovation pathways for five major professional categories:

• Engineering/IT: Build project courses for AI-assisted development, fault diagnosis, and quality closed-loops; increase composite practical training in equipment operation and maintenance and cross-departmental reporting; cut hours for purely templated code and standardized documentation training.
• Business/Economics/Management: Strengthen data interpretation, compliance reviews, and client closed-loop delivery; reduce pure manual bookkeeping and templated copywriting training.
• Medical/Nursing: Core modules should focus on clinical evidence verification, doctor-patient communication, and on-site emergency response; AI diagnostic tools should serve merely as teaching aids.
• Humanities/Law: Focus on text verification, evidence chain drafting, and ethical trade-offs; weaken pure first-draft writing and templated document practice.
• On-site Applied Majors: Strengthen dual modules of embodied practical operation + data review.

All high-growth responsibility bundles span multiple major categories, meaning curriculum reform must break down departmental barriers and establish cross-disciplinary joint practical training mechanisms. A single, generalized AI course cannot adapt to different professional task scenarios; it must be embedded into professional, real-world project training. By comparing the six-year major req/res rankings from 2023 and 2026, the share of items such as historical cultural relic regulations, engineering data diagnosis, and business compliance reviews has surged dramatically, while purely templated skills continue to exit the top ten hiring requirements, providing clear quantitative targets for curriculum updates.

5.2 Six Major Standardizable Human Complementary Core Capabilities

The report distills six types of core capabilities that cannot be substituted by large models and can undergo standardized teaching and evaluation, along with corresponding teaching scenarios and quantifiable assessment standards: situational judgment, responsibility bearing, cross-role coordination, ethical trade-offs, on-site adaptation, and evidence chain expression. Distinct from vague "comprehensive quality" descriptions, each capability category can feature classroom projects and form verifiable deliverables. This is paired with a standardized seven-step complete delivery chain for AI collaboration: Task Breakdown → AI Generation → Fact-Checking → Bias Recording → Human Decision-Making → Citation Annotation → Delivery Review. This requires students to completely retain traces of AI usage, using process evidence as the core for grading, changing the traditional assessment model that grades only the final text or code, and resolving the teaching pain points of AI ghostwriting and plagiarism.

5.3 Stratified Innovation Operational Toolbox and Four-Party Action Matrix

The report produces two sets of reference tools: twelve standardized operations for educational innovation, and a four-dimensional stratified action matrix for Students / Universities / Enterprises / Policies, covering different time dimensions.

• University Side: Seven core actions include dynamic adjustment of major catalogs, micro-major composite minors, curriculum module replacements, co-construction of university-enterprise real task libraries, evaluation system reform, cross-departmental course assembly, and an annual market monitoring flywheel. Using monthly updated C-O*NET data as the adjustment basis, establish a closed-loop mechanism of "Market Task Monitoring → Major Profiling → Curriculum Revision → Practical Training Implementation → Employment Feedback," abandoning the pulse-like major revisions that occur once every few years.
• Student Side: During the preference selection stage, inquire into the AI quadrant of majors; complete composite closed-loop portfolios while in school; proactively supplement on-site and judgment-based practical training, avoiding purely standardized single-skill tracks.
• Enterprise Side: Open real job task bundles to institutions, participate in curriculum reviews, and co-build talent evaluation standards to lower hiring screening costs.
• Policy Side: Build a unified national monitoring platform for recruitment-major supply and demand; establish an early warning mechanism for major enrollment; restrict enrollment expansion for high-exposure, low-complementarity majors; and perfect supporting systems for the integration of industry and education.