This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.

Large Language Models at Work in China's Labor Market

Chen, Ge, Xie, Xu, Yang

2025China Economic Review4 citations

Exposure / measurementTheoretical model

LLM / Generative AIGeneral automation

View Repository DOI: 10.1016/j.chieco.2025.102413

Summary

Chen et al. use LLM classifiers (GPT-4, InternLM, GLM) and expert annotations to construct occupational exposure scores for 1606 Chinese occupations, correlate these with wage and vacancy data from online job postings, and develop an information-theoretic task model to explain why LLMs disproportionately affect high-wage, non-routine cognitive occupations.

Main Finding

LLM exposure positively correlates with occupational wages and experience premiums in China, with non-routine cognitive analytical tasks most affected—contradicting the routinization hypothesis—and education and healthcare industries showing highest exposure while manufacturing and agriculture show lower exposure.

Primary Datasets

None (conceptual paper)

Secondary Datasets

None

Key Methods: LLM-based occupation classification using GPT-4, InternLM, and GLM with expert validation; descriptive correlation analysis between exposure scores and wage/vacancy data from online job postings; theoretical modeling integrating information theory (entropy, KL-divergence) into task-based framework
Sample Period: Not applicable
Geographic Coverage: China
Sample Size: 1636 fine-category occupations (1606 excluding military/government); 800 million online job postings from 5.2 million employers
Level of Analysis: Occupation, Industry
Occupation Classification: None
Industry Classification: None
Replication Package: Partial

Notes

China Economic Review, vol. 92 [Claude classification]: This is a conceptual/policy piece, not an empirical research paper. It explores potential scenarios and makes policy recommendations but does not test hypotheses or analyze data. Published in MIT's online collection on generative AI applications. The authors draw on economic theory (search models, signaling) and make parallels to prior technological transitions (online job boards) but do not present formal models or empirical evidence. [Claude classification]: This is a conceptual/policy piece, not an empirical research paper. It explores potential scenarios and makes policy recommendations but does not test hypotheses or analyze data. Published in MIT's online collection on generative AI applications. The authors draw on economic theory (search models, signaling) and make parallels to prior technological transitions (online job boards) but do not present formal models or empirical evidence. [Claude classification]: This is a conceptual/policy piece, not an empirical research paper. It explores potential scenarios and makes policy recommendations but does not test hypotheses or analyze data. Published in MIT's online collection on generative AI applications. The authors draw on economic theory (search models, signaling) and make parallels to prior technological transitions (online job boards) but do not present formal models or empirical evidence. [Claude classification]: This is a conceptual/policy piece, not an empirical research paper. It explores potential scenarios and makes policy recommendations but does not test hypotheses or analyze data. Published in MIT's online collection on generative AI applications. The authors draw on economic theory (search models, signaling) and make parallels to prior technological transitions (online job boards) but do not present formal models or empirical evidence. [Claude classification]: This is a conceptual/policy piece, not an empirical research paper. It explores potential scenarios and makes policy recommendations but does not test hypotheses or analyze data. Published in MIT's online collection on generative AI applications. The authors draw on economic theory (search models, signaling) and make parallels to prior technological transitions (online job boards) but do not present formal models or empirical evidence. [Claude classification]: This is a conceptual/policy piece, not an empirical research paper. It explores potential scenarios and makes policy recommendations but does not test hypotheses or analyze data. Published in MIT's online collection on generative AI applications. The authors draw on economic theory (search models, signaling) and make parallels to prior technological transitions (online job boards) but do not present formal models or empirical evidence. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores. [Claude classification]: This paper is primarily a measurement/exposure construction study rather than empirical causal analysis. The theoretical model (Section 5) incorporates entropy and KL-divergence from information theory into a task-based framework to distinguish traditional AI from deep learning/LLM automation logic. The paper uses LLMs as methodological tools (GPT-4, InternLM, GLM for classification) but studies LLMs as the object of analysis. Correlations reported are descriptive, not causal. The online job posting data may not be representative of China's entire labor market, as acknowledged by authors. Expert validation shows correlation of 0.65 with LLM scores.