CoChat
Sign Up
Cited by 17Labor market dynamics and wage inequalityDigital Economy and Work TransformationImpact of AI and Big Data on Business and Society

Large language models at work in China’s labor market

Open access
Chen, Qin, Ge, Jinfeng, Xie, Huaqing, Xu, Xingcheng, Yang, Yanqing·2025·China Economic Review
DOI 10.1016/j.chieco.2025.102413
Open this paper in CoChat — free accountOpen PDF

Abstract

This paper explores the potential impacts of large language models (LLMs) on the Chinese labor market. We analyze occupational exposure to LLM capabilities by incorporating human expertise and LLM classifications, following the methodology of Eloundou et al. (2023). The results indicate a positive correlation between occupational exposure and both wage levels and experience premiums at the occupation level. This suggests that higher-paying and experience-intensive jobs may face greater exposure risks from LLM-powered software. We then aggregate occupational exposure at the industry level to obtain industrial exposure scores. Both occupational and industrial exposure scores align with expert assessments. Our empirical analysis also demonstrates a distinct impact of LLMs, which deviates from the routinization hypothesis. We present a stylized theoretical framework to better understand this deviation from previous digital technologies. By incorporating entropy-based information theory into the task-based framework, we propose an AI learning theory that reveals a different pattern of LLM impacts compared to the routinization hypothesis.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper estimates occupational and industrial exposure to LLMs in China using GPT-4, GLM, InternLM, and expert scores, finding exposure is positively correlated with wages and experience premiums and is highest for non-routine cognitive analytical tasks. It argues this pattern deviates from the routinization hypothesis and develops an entropy/KL-divergence task model in which LLMs have a comparative advantage in complex, data-intensive tasks.

Key findings

  • Occupational exposure is positively correlated with wage levels and experience premiums; more educated, higher-paid, white-collar occupations are the most exposed.
  • LLM scoring: fine-category means are GLM 0.44, InternLM 0.18, GPT-4 0.24; medium-category means are GLM 0.40, InternLM 0.14, GPT-4 0.22. Model scores are positively correlated and align with expert scores (Figure 1 corr.=0.65).
  • Composite occupations: non-routine cognitive analytical tasks are most affected; routine manual tasks are also substantially influenced; interpersonal and non-routine manual physical tasks are less affected. Industry exposure is higher in education and healthcare and lower in manufacturing, agriculture, mining, and construction. Younger workers are disproportionately affected; in 2020, 42.5% of ages 16–19 and 34.9% of ages 20–24 were in consumer service industries.
  • Vacancy shares and the growth rate of vacancy shares are positively correlated with occupational exposure. The theoretical model predicts LLMs erode skilled expertise and flatten wage distribution, unlike traditional AI’s routine-task automation.

Methods

  • Exposure scoring uses the Occupational Classification Dictionary of China 2022; GPT-4, InternLM, and GLM classify occupations, each occupation is labeled 8 times and averaged, and expert annotators provide comparison scores. Scores E0–E3 are mapped as E1=1, E2/E3=0.5, E0=0; three categories are excluded, leaving 1606 fine-category occupations.
  • Data include online job postings from City Data Group, January 2017–December 2022, covering over 800 million job openings, nearly 400 cities, and 5.2 million employers; CLDS 2016 for industry occupational intensity; and 2020 census data for demographic distribution.
  • Analysis aggregates occupational exposure to industry via occupational intensity, compares LLM scores with expert scores, correlates exposure with wages, experience premiums, and vacancy shares/growth, and develops a task-based model incorporating entropy and KL-divergence.

Limitations

  • Paper-stated: the task-based framework’s validity, relative versus absolute measures, and the forward-looking/changing nature of scores are concerns; LLM scoring is random, as the same prompt can yield different results even at higher temperature, addressed by averaging 8 labels per occupation.
  • Paper-stated: the Occupational Classification Dictionary provides only detailed work content and descriptions, so task-level exposure as in Eloundou et al. is infeasible; exposure is calculated at the occupation level.
  • Paper-stated: the online recruitment labor market does not equal the whole labor market, and China lacks a representative recruitment dataset like JOLTS, so representativeness cannot be fully assessed; the authors argue exposure elicitation does not use postings and average wage/experience premium are robust to sectoral reliance. Evident from the text: industry and demographic exposure rely on CLDS 2016 and 2020 census distributions, which may not match the 2017–2022 posting period.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.chieco.2025.102413

Cites

65 of 65 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

10 of 21 cited sources checked · 4 supported · 1 flagged · accuracy not yet measured

  • Findingcitation
    § 5 The Model

    Occupations with higher wage and experience premiums are more vulnerable to LLMs.

    “Additionally, our findings indicate a correlation between occupations with higher wage and experience premiums and their increased vulnerability to LLMs. This aligns with the recent research by Eloundou et al. (2023) and Eisfeldt et al. (2023), who have identified that technological advancements disproportionately affect workers in the higher echelons of the wage spectrum.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Eloundou et al. (2023) source

      Section 4.2 (Wages and Employment) finds that higher wages are generally associated with greater occupational exposure to LLMs. Section 4.4 (Barriers to Entry) finds that exposure generally rises across job-preparation zones, which include related experience, but it does not establish a relationship with experience premiums. The paper measures potential exposure, not realized harm to workers. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites Eisfeldt et al. (2023) source
  • Premisecitation
    § 5.2.2 The Comparative Advantage of Deep Learning AI and LLMs AI Technology

    Deep learning AI and LLMs AI models exhibit similar power scaling laws.

    “Given that deep learning AI and LLMs AI models exhibit similar power scaling laws, our focus can be directed specifically towards LLMs AI without a significant loss of generality.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Kaplan et al., 2020

      The Summary of Scaling Laws (p. 4) establishes power-law scaling for Transformer language models, and Related Work (p. 18) notes an approximate power-law relationship for EfficientNet image models. Those passages support a resemblance in some scaling behavior, but not the broad claim that deep-learning models and LLMs generally exhibit similar scaling laws. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites Hoffmann et al., 2022 source

      The Introduction (p. 1) describes a power-law relationship between language-model size and performance, and the Discussion (p. 15) discusses power-law scaling for large language models. The paper does not compare those laws with deep-learning models generally, so it does not establish the claimed similarity or justify treating LLMs as representative of all deep learning AI. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites Barnett and Besiroglu, 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Premisecitation
    § 5.3.1 Discussions With AI Applications

    A controlled experiment by Peng et al. (2023) found that GitHub Copilot significantly boosted productivity: the group using Copilot completed programming tasks approximately 56% faster than the control group without access to the tool.

    “For example, a controlled experiment by Peng et al. (2023) found that GitHub Copilot, a generative AI programming aid, significantly boosted productivity: the group using Copilot completed programming tasks approximately 56% faster than the control group without access to the tool.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Peng et al. (2023) source
  • Findingown data
    § Abstract

    Occupational exposure to LLMs is positively correlated with wage levels at the occupation level in China.

    “The results indicate a positive correlation between occupational exposure and both wage levels and experience premiums at the occupation level.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 1 Introduction

    There is a positive correlation between the growth rate of vacancy shares and occupational exposure scores.

    “The positive correlation between the growth rate of vacancy shares and exposure scores further indicates a potential reversal in labor demand trends.”
    • Also stated in: § 4.5 Vacancies · hedged
      “The positive correlation between occupational exposure and growth rate of vacancy share implies that the advancements of LLMs could unexpectedly reverse the prevailing labor demand trend”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 26
  • May combine two claims. 32
  • The quote only partly states this claim. 2

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 6 Discussions and Conclusions

    Labor income inequality has widened significantly, with computerization facilitating the unprecedented concentration of decision-making power among elite experts, who now have access to abundant and inexpensive information.

    “Labor income inequality has widened significantly, with computerization facilitating the unprecedented concentration of decision-making power among elite experts, who now have access to abundant and inexpensive information (Deming, 2021).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Deming, 2021 source

      In the discussion accompanying Figure 2 (page unavailable), Deming says ICT produces large amounts of data and decentralizes firm-level decision-making, contrary to the claimed concentration of power among elite experts. The paper reports rising returns to decision-intensive work, but that does not establish the sentence’s claim about widening labor-income inequality. (LLM, conf=0.88)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 1 Introduction

    Recent studies on the US labor market found that technological advancements disproportionately impact workers at the higher end of the wage spectrum.

    “These results align with recent studies on the US labor market (Eloundou et al., 2023; Eisfeldt et al., 2023), which also found that technological advancements disproportionately impact workers at the higher end of the wage spectrum.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Eloundou et al., 2023 source

      Section 4.2, “Wages and Employment” (page unavailable), finds that higher-wage US occupations tend to have greater exposure to LLMs. That supports the wage pattern, but the paper measures potential exposure to LLMs, not the realized impact of technological advancements generally on workers. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites Eisfeldt et al., 2023 source

  • Classified as probably not a substantial claim.

    Premisecitation
    § 4.2 Occupational Exposure

    Prior research on computer-based automation posited that routine work was most susceptible to replacement by computers.

    “Prior research on computer-based automation posited that routine work was most susceptible to replacement by computers (Autor et al., 2006; Acemoglu and Autor, 2011).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Autor et al., 2006 source
    • cites Acemoglu and Autor, 2011 source

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § 3.3 Methodology of Exposure Scoring

    The exposure scoring methods of Eloundou et al. (2023) conceive an occupation as a collection of tasks and assess whether a given occupation can be executed more efficiently using ChatGPT or analogous LLMs.

    “Building upon the exposure scoring methods proposed by (Eloundou et al., 2023) which conceive an occupation as a collection of tasks, assess whether a given occupation can be executed more efficiently using ChatGPT or analogous LLMs.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Eloundou et al., 2023 source

  • May combine two claims.

    Premisecitation
    § 6 Discussions and Conclusions

    Doctors often overrode AI predictions, highlighting their lack of understanding of how to use the tools effectively.

    “This outcome occurred because doctors often overrode AI predictions, highlighting their lack of understanding of how to use the tools effectively.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Agarwal et al., 2023 source

      Section 5.4 reports that radiologists underweighted AI information and made errors when combining it with their own assessments, limiting the gains from AI assistance. But Section 4.2.1 finds that their predictions moved closer to the AI’s when it was available; the paper does not establish that they often overrode AI predictions. It identifies biased use of AI, not a demonstrated lack of understanding of how to use the tool. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation
    § 6 Discussions and Conclusions

    A recent experiment found that providing AI assistance to radiologists did not improve diagnostic quality, even though AI predictions were at least as accurate as those of two-thirds of the doctors studied.

    “For example, a recent experiment found that providing AI assistance to radiologists did not improve diagnostic quality, even though AI predictions were at least as accurate as those of two-thirds of the doctors studied (Agarwal et al., 2023).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Agarwal et al., 2023 source

  • May combine two claims.

    Findingown data
    § 1 Introduction

    Industry-level exposure to LLMs is higher in education and healthcare and lower in manufacturing, agriculture, mining, and construction.

    “Industry-level analyzes reveal that education and healthcare have a higher level of exposure, while manufacturing, agriculture, mining, and construction show lower exposure.”
    • Also stated in: § 4.4 Industrial and Demographic Exposure · plain
      “sectors with the highest scores, including education, culture and arts, radio, film and television, exhibit heightened susceptibility.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Abstract

    Occupational and industrial exposure scores align with expert assessments.

    “Both occupational and industrial exposure scores align with expert assessments.”
    • Also stated in: § 4.1 Summary Statistics · plain
      “we observed substantial positive correlations between the model scores and human expert scores.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.4 Industrial and Demographic Exposure

    During 2020, 42.5% of individuals aged 16-19 and 34.9% of those aged 20-24 in China were engaged in consumer service industries.

    “during 2020, 42.5% of individuals aged 16-19 and 34.9% of those aged 20-24 were engaged in consumer service industries.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.1 Summary Statistics

    Across fine-category occupations, GLM had an average exposure score of 0.44, InternLM 0.18, and GPT-4 0.24.

    “GLM demonstrated an average of 0.44, while InternLM and GPT-4 yielded averages of 0.18 and 0.24 respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 6 Discussions and Conclusions

    The potential effect of the release of LLMs could have a substantial effect on the Chinese labor market, potentially altering previous labor demand trends.

    “Our findings suggest that the potential effect of the release of LLMs could have a substantial effect on the Chinese labor market, potentially altering previous labor demand trends.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § Appendix E.3 Errors Checking and Missing Information

    In a very small number of cases, the plain text includes parts of the website other than the job description.

    “In a very small number of cases we observe that the plain text includes some parts of the website other than the job description.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § Appendix E.3 Errors Checking and Missing Information

    In some cases, a recruitment advertisement is duplicated within the same platform or across platforms.

    “In some cases, a recruitment advertisement will be released duplicated within the same platform or across platform.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 1 Introduction

    The advancements of LLMs could reverse the prevailing labor demand trend and intensify structural unemployment in the future.

    “The positive correlation between the growth rate of vacancy shares and exposure scores further indicates a potential reversal in labor demand trends.”
    • Also stated in: § 4.5 Vacancies · hedged
      “could unexpectedly reverse the prevailing labor demand trend, thereby intensifying the structural unemployment predicament in the future.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § 4.2 Occupational Exposure

    LLMs could potentially alter the experience premium and the profile of returns from learning.

    “LLMs could potentially alter the experience premium and the profile of returns from learning.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premiseown data
    § 5 The Model

    Automation predominantly affects occupations involving non-routine cognitive tasks, departing from earlier conclusions that automation primarily affects occupations characterized by routine tasks.

    “Drawing on prior empirical findings, we have discerned a predominant impact of automation on occupations involving non-routine cognitive tasks. This observation marks a departure from earlier conclusions, such as those posited by Autor et al. (2003), which suggested that automation primarily affects occupations characterized by routine tasks.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Autor et al. (2003) source

  • May combine two claims.

    Premiseown data
    § 5 The Model

    Automation has a pronounced effect on occupations heavily reliant on non-routine cognitive analytical tasks, in contrast to a relatively minimal impact on manual physical tasks.

    “A key insight from our analysis is the pronounced effect of automation on occupations heavily reliant on non-routine cognitive analytical tasks, in contrast to the relatively minimal impact on manual physical tasks.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § 1 Introduction

    The impacts of LLMs on China's labor market are pervasive and diverse.

    “Our analysis further indicates that the impacts of LLMs on China's labor market are pervasive and diverse.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductown data
    § 3.3 Methodology of Exposure Scoring

    Experts in economics and artificial intelligence scored medium-category occupations according to rating criteria, and their average score was used as the final human score.

    “we invited a group of experts in economics and artificial intelligence to serve as judges. We provided them with descriptions of medium-category occupations in China and asked them to score each occupation according to the rating criteria. After collecting the scores from all of the experts, we calculated the average score for each occupation as the final human score.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning
    § 3.3 Methodology of Exposure Scoring

    Each LLM model labeled each occupation 8 times and the average was taken as the final scoring result for each LLM model for each occupation.

    “we first had each LLM model label each occupation 8 times and calculated the scores. We then took the average as the final scoring result for each LLM model for each occupation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The same prompt can still yield different results from large language models, even with a higher temperature setting.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § 4.2 Occupational Exposure

    The routinization hypothesis assumed that computers and computer-controlled equipment are highly productive and reliable at performing scriptable tasks and relatively inept at everything else.

    “'routinization' hypothesis assumed that 'computers and computer-controlled equipment are highly productive and reliable at performing the tasks that programmers can script - and relatively inept at everything else'.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Premisereasoning
    § 5.3.2 Theoretical and Policy Implications

    Cognitive tasks at their core involve making predictions, and the productivity of cognitive tasks is determined by prediction accuracy, represented through KL-divergence.

    “The central idea is that cognitive tasks, at their core, involve making predictions-such as forecasting GDP growth, predicting stock prices, or generating responses in the case of LLMs. Intuitively, the productivity of cognitive tasks is determined by prediction accuracy, which we represent through KL- divergence.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisenone given
    § 5.3.2 Theoretical and Policy Implications

    Traditional AI depends on pre-programmed rules to perform tasks, while LLMs leverage neural scaling laws and data-driven methods.

    “In our theoretical model, the key distinction between traditional AI and LLMs as automation technologies lies in their operational logic: traditional AI depends on pre-programmed rules to perform tasks, while LLMs leverage neural scaling laws and data-driven methods.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § 5 The Model

    The unique patterns of occupational exposure observed are directly attributable to the underlying technological principles governing automation processes.

    “Essentially, our theoretical model posits that the unique patterns of occupational exposure observed are directly attributable to the underlying technological principles governing automation processes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 5.3.1 Discussions With AI Applications

    ChatGPT allowed the most capable writers to work faster and enabled the less capable to write faster and improve output quality, narrowing the productivity gap.

    “Moreover, ChatGPT allowed the most capable writers to work faster and enabled the less capable to not only write faster but also to improve the quality of their output. This narrowing of the productivity gap underscores the potential of LLMs to democratize expertise.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Noy and Zhang (2023) source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § 5.3.1 Discussions With AI Applications

    LLMs are more suited to automating tasks traditionally performed by skilled workers, rather than simple, rule-based jobs.

    “Consequently, LLMs are more suited to automating tasks traditionally performed by skilled workers, rather than simple, rule-based jobs.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § 5.3.1 Discussions With AI Applications

    LLMs are showing the potential to empower a broader group of workers to perform expert tasks such as medical diagnosis and report generation.

    “While these applications confirm that both medical diagnosis and report generation require highly skilled expertise, LLMs are showing the potential to empower a broader group of workers to perform these expert tasks.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § 5.3.1 Discussions With AI Applications

    Deep learning AI and LLMs follow a data-driven methodology where developers feed vast datasets into an algorithm that learns to perform tasks.

    “In contrast, deep learning AI and LLMs follow a data-driven methodology. Developers no longer need to predefine the rules; instead, they feed vast datasets into an algorithm, which then learns to perform tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 5.3.2 Theoretical and Policy Implications

    KL-divergence is used in the neural scaling laws to measure an AI model's prediction accuracy.

    “KL-divergence is also used in the 'neural scaling laws' to measure an AI model's prediction accuracy, which gives task performance a solid foundation by directly borrowing well-established AI technological principles and bridging them with information theory.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 5.2.1 The Comparative Advantage of Traditional AI

    Traditional AI models replace routine tasks typically executed by unskilled workers, reducing the task domain available to unskilled workers and widening the wage disparity between skilled and unskilled labor.

    “Proposition 2 in our study reveals that models leveraging traditional AI technology are inclined to replace routine tasks, typically executed by unskilled workers. This leads to a reduction in the task domain available to unskilled workers, subsequently widening the wage disparity between skilled and unskilled labor.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning
    § 5.3.1 Discussions With AI Applications

    Traditional AI code follows a rule-based structure where developers must explicitly define each step, making routine, less complex tasks more susceptible to automation.

    “In traditional AI, this code follows a rule-based structure: developers must explicitly define each step in the process, particularly for tasks that are repetitive or relatively simple. This makes routine, less complex tasks more susceptible to automation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 5.3.1 Discussions With AI Applications

    A study by Noy and Zhang (2023) found that access to ChatGPT improved the speed and quality of writing tasks among professionals, reducing time spent on tasks by 40%, and the least skilled writers in the ChatGPT group performed as well as the median writer in the control group.

    “Similarly, a study by Noy and Zhang (2023) found that access to ChatGPT improved the speed and quality of writing tasks among professionals such as marketers, grant writers, consultants, and managers. The ChatGPT group reduced the time spent on tasks by 40%, and even the least skilled writers in this group performed as well as the median writer in the control group-highlighting a significant boost”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Noy and Zhang (2023) source

  • May combine two claims.

    Interpretationreasoning
    § 5.2.1 The Comparative Advantage of Traditional AI

    In competitive equilibrium with traditional AI, there exist two thresholds such that low-skill workers, high-skill workers, and models are allocated to different task intervals, and the wage premium increases.

    “Proposition 2 : In competitive equilibrium, there exists two threshold ˜ I 1 and ˜ I 2 such that 0 < ˜ I 1 < ˜ I 2 < 1 and for any ≤ i < ˜ I 1, ˜ l ( i ) = ˜ h ( i ) = 0, for any ˜ I 1 < i < ˜ I 1, ˜ h ( i ) = ˜ m ( i ) = 0, for any ˜ I 2 < i ≤ 1, ˜ l ( i ) = ˜ m ( i ) = 0. Furthermore, the wage premium increases.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • May combine two claims.

    Interpretationcitation
    § 5.3.1 Discussions With AI Applications

    The theoretical model predicts that LLMs will reshape the value and nature of human expertise, helping to narrow the income inequality between skill levels.

    “Our theoretical model aligns with this vision. We predict that LLMs will reshape the value and nature of human expertise, helping to narrow the income inequality between skill levels.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning
    § 6 Discussions and Conclusions

    Artificial intelligence has the potential to reduce income inequality, contrary to the trends observed with earlier technologies.

    “In this regard, our theoretical perspective suggests that artificial intelligence has the potential to reduce income inequality, contrary to the trends observed with earlier technologies.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 6 Discussions and Conclusions

    While job polarization did not occur in China, the wages of elite experts rose dramatically, indicating that information technology has played a significant role in exacerbating income inequality.

    “While job polarization did not occur in China, the wages of elite experts rose dramatically, indicating that information technology has played a significant role in exacerbating income inequality.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Ge and Yang (2014) source

  • May combine two claims.

    Interpretationreasoning
    § 5.3.2 Theoretical and Policy Implications

    The shift to LLMs erodes the expertise position of skilled workers and leads to a more flattened wage distribution between skill levels.

    “This shift erodes the expertise position of skilled workers and leads to a more flattened wage distribution between skill levels.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning
    § 6 Discussions and Conclusions

    The theoretical model captures at least part of the momentum shaping labor market demand in China.

    “In this sense, we believe the model captures at least part of the momentum shaping labor market demand in China.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning
    § 5.3.2 Theoretical and Policy Implications

    Algorithms based on distinct technological logic are predisposed to disrupt different types of tasks, and variances in technological underpinnings will inevitably result in divergent outcomes in productivity and income distribution.

    “The overarching insight from our theoretical model is that algorithms, based on distinct technological logic, are predisposed to disrupt different types of tasks. These variances in technological underpinnings will inevitably result in divergent outcomes in terms of productivity and income distribution.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § 6 Discussions and Conclusions

    The theoretical model provides a deeper understanding of the reasons behind occupational exposure to LLMs, which diverges from the prediction of the routinization hypothesis.

    “Furthermore, we develop a theoretical model to provide a deeper understanding of the reasons behind occupational exposure to LLMs, which diverges from the prediction of the routinization hypothesis.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning
    § 5.3.2 Theoretical and Policy Implications

    The progression of LLMs AI technology will generate outcomes that deviate from those predicted by the routinization hypothesis, and its influence on labor markets is anticipated to be distinctly different from that of traditional AI.

    “This distinction is why the progression of LLMs AI technology will generate outcomes that deviate from those predicted by the routinization hypothesis. Fundamentally, the influence of LLMs AI technology on labor markets is anticipated to be distinctly different from that of traditional AI.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • May combine two claims.

    Interpretationreasoning
    § 5.3.2 Theoretical and Policy Implications

    Traditional AI excels at performing less complex tasks typically carried out by unskilled workers, and automation of such tasks increases wage premiums between skilled and unskilled workers by eroding the position of unskilled workers.

    “Traditional AI excels at performing less complex tasks typically carried out by unskilled workers. The automation of such tasks increases wage premiums between skilled and unskilled workers by eroding the position of unskilled workers.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All markets are competitive.
    • The authors state this as an assumption: There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).
    • The authors state this as an assumption: Skilled workers have a comparative advantage in more complex tasks.
    • The authors state this as an assumption: The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.
    • The authors state this as an assumption: The accuracy of a model's performance is directly proportional to its productivity in automated tasks.
    • The authors state this as an assumption: KL-divergence is a measurement of productivity loss.
    • The authors state this as an assumption: Deep learning AI and LLMs AI models represent a distinct technological logic.
    • The authors state this as an assumption: For each task, there is a productivity upper bound A_bar.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 6 Discussions and Conclusions

    Ge and Yang (2014) documented that between 1992 and 2007, during rapid economic growth in China, the average real wage increased by 202%, accompanied by a sharp rise in wage inequality.

    “Ge and Yang (2014) documented that between 1992 and 2007, during a period of rapid economic growth in China, the average real wage increased by 202%, accompanied by a sharp rise in wage inequality.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Ge and Yang (2014) source

  • The quote only partly states this claim.

    Interpretationreasoning
    § 4. Labor market clearing condition:

    In the model with deep learning AI and LLMs, the threshold shifts.

    “In sum, we have proved that a shift of threshold and a decrease in wage premium.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § Appendix E.4 Representativeness of Online Job Vacancy Postings

    The online recruitment labor market does not equal the whole labor market.

    “Obviously, the online recruitment labor market does not equal to the whole labor market”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data
    § Appendix E.3 Errors Checking and Missing Information

    Duplicate recruitment ads were adjusted in the empirical analysis, especially when counting vacancy numbers.

    “In our empirical analysis, esp. counting vacancy number, we consider this concern by adjusting duplicate recruitment ads.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Interpretationreasoning
    § 4. Labor market clearing condition:

    In the model with deep learning AI and LLMs, the wage premium decreases.

    “In sum, we have proved that a shift of threshold and a decrease in wage premium.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § Appendix E.4 Representativeness of Online Job Vacancy Postings

    The study cannot address the issue of representativeness of online job recruitment.

    “therefore, we cannot address the issue of representative of online job recruitment.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: In China there is no representative recruitment dataset such as JOLTS in the United States.
48 of 48

  • The paper relies on this without support.

    Countries differ in labor market structures such as occupation and industry composition, and even for the same occupation the detailed task composition or work content may exhibit great discrepancies across countries.

    “Nevertheless, countries differ in their labor market structures such as occupation and industry composition. Even for the same occupation, the detailed task composition or work content may exhibit great discrepancies across countries.”

  • The paper relies on this without support.

    The same prompt can still yield different results from large language models, even with a higher temperature setting.

    “The same prompt can still yields different results from large language models, even with a higher temperature setting.”

  • The authors state this as an assumption.

    All markets are competitive.

    “We assume all markets to be competitive.”

  • The authors state this as an assumption.

    There exists a distribution function Pi characterizing the relevant information in task i, and the complexity of task i is defined by entropy Hi(P).

    “Let us assume that there exists a distribution function Pi which characterize the relevant information in task i . Therefore, the complexity of a task i is defined by entropy Hi ( P ) .”

  • The authors state this as an assumption.

    Skilled workers have a comparative advantage in more complex tasks.

    “Furthermore, we assume that skilled worker has a comparative advantage in the more complex tasks.”

  • The authors state this as an assumption.

    The inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.

    “A fundamental postulate of this research is that the inability to distinguish between a model's output and the theoretically perfect execution of a task denotes competence.”

  • The authors state this as an assumption.

    The accuracy of a model's performance is directly proportional to its productivity in automated tasks.

    “Consequently, the accuracy of a model's performance is directly proportional to its productivity in automated tasks.”

  • The authors state this as an assumption.

    KL-divergence is a measurement of productivity loss.

    “Still we assume that KL-divergence is a measurement of productivity loss.”

  • The authors state this as an assumption.

    Deep learning AI and LLMs AI models represent a distinct technological logic.

    “Nowlet us assume that deep learning AI and LLMs AI models represent a distinct technological logic.”

  • The authors state this as an assumption.

    For each task, there is a productivity upper bound A_bar.

    “Let assume that for each task, we have a productivity upper bound ¯ A”

  • The authors state this as an assumption.

    The price of models M is exogenous and equal to F.

    “We also impose a exogenous price of models M as F .”

  • The paper relies on this without support.

    In China there is no representative recruitment dataset such as JOLTS in the United States.

    “Since in China there is no representative recruitment dataset such as JOLTS in United States”
12 of 12

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Figure 16 is cited as the model-expert comparison

    Minor

    The text directs readers to Figure 16 for a comparison of model scores with human expert scores, but Figure 16 is captioned as a comparison of weighted assignments.

  • Figure 4 is cited for experience premiums but shows wage growth

    Minor

    The discussion cites Figure 4 as depicting experience premiums and returns from learning. Its caption instead identifies the plotted outcome as the annual wage growth rate.

  • Proposition 2 defines its middle task interval with identical endpoints

    Minor

    Proposition 2 specifies two ordered thresholds but states the middle allocation condition only for tasks between the first threshold and itself, an empty interval. The proof describes the second threshold as the boundary between unskilled and skilled work, which requires the middle interval to extend to that second threshold.

3 of 3

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Chen, Q., Ge, J., Xie, H., Xu, X., & Yang, Y.
Affiliations
  • Shanghai Maritime University (CN)
  • Shanghai Artificial Intelligence Laboratory (CN)
  • Fudan University (CN)
Corresponding author
  • Ge, Jinfeng
  • Xu, Xingcheng
  • Yang, Yanqing
ORCID
Venue
China Economic Review
Publisher
Elsevier BV
Published
2025-08
Volume
92
Pages
102413
ISSN
1043-951X (ISSN-L)
DOI
10.1016/j.chieco.2025.102413
OpenAlex
W4410015324
Semantic Scholar
7d8cfda875f855b69ce5a8cf659a4ca105a34605
Cited by
17
References
52
Citations by year
2026: 12 · 2025: 3 · 2024: 1 · 2023: 1
FWCI
14.59
Type
journal article
Language
en
Open access
Licence
Locations
  • China Economic Review (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • National Natural Science Foundation of China (71933001)
  • Ministry of Human Resources and Social Security
  • Shanghai Artificial Intelligence Laboratory
Topics
  • Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Labor market dynamics and wage inequality
  • Social Sciences › Sociology and Political Science › Digital Economy and Work Transformation
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Impact of AI and Big Data on Business and Society
Keywords
large language models, labor market, occupational exposure, wage premium, Shannon entropy, human expertise, automation risk, Chinese labor market

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-12.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-12.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review