Rising AI shifts hiring toward more-educated staff in China’s listed firms: each unit increase in firm-level AI exposure cuts low-education labor share by 0.007 and raises high-education share by 0.006, driven by firms’ technological innovation and strongest in high-tech sectors.

Main Finding

Firm-level AI development in China (2014–2024) systematically reshapes firms’ educational composition of labor: higher AI activity reduces the share of low-educated workers and raises the share of high-educated workers. The author interprets this as “substituting for low-educated labor while complementing high-educated labor.” Technological innovation capability mediates this effect, and the impacts vary markedly by region and industry.

Key Points

• Quantitative effects (estimated): a one-unit increase in the firm-level AI index is associated with a 0.007-unit decrease in the share of low-educated labor and a 0.006-unit increase in the share of high-educated labor.
• Mechanism: AI development enhances firms’ technological innovation capability (TCE), which in turn reduces demand for low-educated labor and increases demand for high-educated labor (mediation support for the TCE channel).
• Heterogeneity:
  • Regional: substitution (reduction of low-educated share) is stronger in developed regions; complementarity (rise in high-educated share) is stronger in less-developed regions.
  • Industry: effects are much larger in high-technology industries — roughly 2.5 times the magnitude in non-high-tech industries.
• Theoretical framing: applies skill-biased technological change and the task-model literature — AI acts as a general-purpose technology that both substitutes routine/low-skill tasks and complements high-skill cognitive/innovative tasks.
• Limitation noted by author: focus restricted to educational composition (not other worker attributes such as specific skills, age, or gender).

Data & Methods

• Sample: Chinese A‑share listed firms on Shanghai and Shenzhen exchanges, 2014–2024.
• Core explanatory variable (AI): firm-level AI development index constructed from annual report and patent text:
  • Initial seed list: 52 AI-related terms from literature and industry reports.
  • Semantic expansion via Word2Vec (skip-gram) trained on firms’ texts; top similar words per seed word retained, deduplicated and filtered → final AI dictionary of 73 keywords.
  • AI indicator = ln(1 + total frequency of AI keywords in a firm’s annual report).
• Dependent variables (educational composition):
  • Share of high-educated labor: proportion of employees with bachelor’s degree or above (from annual reports).
  • Share of low-educated labor: proportion with bachelor’s degree or below.
• Mechanism variable (TCE): firm technological innovation capability measured via text analysis and ML using an initial seed-word set (45 digital-technology terms) and annual report texts (details of the exact index construction are described in the paper).
• Controls: firm size, age, ownership type (SOE), leverage, profitability (ROA), growth, ownership concentration (Top1), CEO duality, etc.
• Empirical strategy: panel-level empirical analysis (regressions) estimating the association between AI index and labor-education shares, mediation analysis for TCE, and subgroup/heterogeneity tests across regions and industries. (The paper uses text- and ML-based variable construction and standard econometric controls to identify relationships.)

Implications for AI Economics

• Micro evidence for skill-biased technological change: this firm-level study supports the view that AI raises relative demand for higher-educated workers while displacing lower-educated workers, with measurable magnitudes in listed firms.
• Innovation as a transmission channel: strengthening firm R&D and technological capabilities is a key pathway by which AI reshapes labor composition — implying policies that influence firm innovation will also affect labor demand composition.
• Policy targeting:
  • Labor-market policy should be differentiated by region and industry given heterogeneity (e.g., stronger substitution in developed regions and larger effects in high-tech sectors).
  • Active re-skilling/up-skilling and education policy (STEM and interdisciplinary training) are important to enable worker transitions and to supply firms’ demand for higher-educated talent.
  • For equity and social stability, complementary measures (retraining, income support, mobility programs) are needed to mitigate short-run displacement of low-educated workers.
• Measurement and research methods: the paper demonstrates a replicable firm-level approach to measuring AI exposure using annual-report text mining augmented by Word2Vec — useful for future empirical work on AI’s economic effects.
• Broader economic considerations: shifting firm-level labor composition toward higher-educated workers may amplify wage dispersion and require attention to complementary institutions (education, social insurance, regional development) to manage distributional consequences while capturing productivity gains.