生成式人工智能与企业劳动力需求结构分化——基于ChatGPT发布的准自然实验

Generative artificial intelligence (AI) is profoundly reshaping the occupational task structure and labor demand allocation of enterprises. Using monthly data from online recruitment platforms in China from 2022 to 2023, this paper treats the release of ChatGPT as a quasi-natural experiment and combines occupation-level large language model exposure to construct a generalized DID model, thereby systematically examining the short-run impact of generative AI on enterprise labor demand and its underlying mechanisms. The results show that generative AI significantly triggers a structural adjustment in the recruitment demand of enterprises, with the contraction in routine occupations being substantially greater than that in non-routine occupations, indicating that technological shocks are first concentrated in cognitive jobs with a higher degree of programmability. Further task-based analysis suggests that this effect is not evenly distributed across all cognitive occupations, but mainly concentrated in routine cognitive occupations and information-processing non-routine cognitive occupations, with a relatively limited impact on non-routine cognitive occupations centered on interpersonal interaction and complex decision-making. Mechanism testing shows that enterprises, on the one hand, compress demand for routine-related positions through task efficiency improvement and, on the other hand, increase the probability of new occupations in non-routine domains through new task creation. The two channels jointly reinforce the structural differentiation of labor demand. Heterogeneity analysis further indicates that the impact is more pronounced in positions without experience requirements, private enterprises, labor-intensive industries, and traditional industries, suggesting that sectors with greater market exposure and higher task replaceability are more vulnerable to generative AI shocks. Further analysis shows that although recruitment demand for routine occupations declines more rapidly, non-routine occupations experience stronger wage premiums and significantly higher competition intensity. This implies that generative AI mainly works through substitution in routine occupations, whereas in non-routine occupations it is more closely associated with skill complementarity, productivity enhancement, and intensified competition. From the perspective of enterprise recruitment, this paper uncovers the micro-level mechanisms through which generative AI reshapes labor demand structures and provides new empirical evidence for understanding employment adjustment and occupational differentiation under technological shocks.