AI adoption in China’s manufacturing sector raises hiring and pay but deepens internal wage gaps; firms deploying AI show higher employment and wages, yet gains concentrate in technical and service roles while pay dispersion widens, with stronger job gains in regions with developed AI ecosystems and supportive policies.

Main Finding

AI adoption by Chinese listed manufacturing firms (panel of >1,700 firms, 2001–2024) is associated with net employment expansion and higher wages (for employees and executives) but also with wider intra‑firm pay disparities. These aggregate gains arise from within‑firm occupational reallocation — growth in technical and service roles and contraction in production and routine managerial jobs — while gender balance improves slightly (modest decline in male share). Regional AI ecosystems and supportive policies strengthen employment gains and help mitigate inequality.

Key Points

• Sample and scope
  • Firm-level analysis of >1,700 Chinese A‑share listed manufacturing firms (2001–2024).
  • Focus on both employment quantity and within‑firm structure/distribution (wages, occupational shares, gender).
• Measurement of AI adoption
  • Constructed multi‑dimensional AI indicators using: R&D investment, patent activity related to AI, and firm textual disclosures referencing AI.
• Core empirical results
  • Employment quantity: AI adoption is positively associated with overall firm employment (supports H1).
  • Wages and inequality: AI adoption increases average wages for employees and executives (H2a) but widens intra‑firm wage dispersion (H2b).
  • Gender composition: Male‑to‑female employment ratio declines modestly after AI adoption (H3) — improved gender balance but small effect.
  • Occupational structure: Technical and service occupations expand while production and routine managerial shares contract (H4a, H4b).
  • Moderation by environment: Regions with more developed AI ecosystems and stronger policy support amplify employment gains and attenuate unequal effects (H5a, H5b); limited moderating role on gender composition.
• Mechanisms identified
  • Mediation analyses attribute employment and wage outcomes to task reallocation: AI complements technical/service tasks (demand ↑) and substitutes routine production/managerial tasks (demand ↓).
• Robustness/dynamics
  • Results hold across fixed‑effects regressions, mediation and moderation models, and dynamic specifications (event/dynamic analyses reported).
• Caveat
  • Manuscript is an unedited pre‑publication version (authors note potential errors prior to final editing).

Data & Methods

• Data
  • Panel dataset of publicly listed manufacturing firms on China’s A‑share market, 2001–2024 (N > 1,700 firms).
  • Firm disclosures used to extract occupational counts, wages, executive compensation, and gender composition; regional indicators of AI industry development and policy support appended.
• AI adoption measures
  • Multi‑proxy approach: AI‑related R&D intensity, AI patent counts, and text‑based indicators from corporate reports/filings referencing AI adoption.
• Outcome variables
  • Overall employment level; average employee wages; executive compensation; intra‑firm wage dispersion; male/female employment shares; occupational shares (technical, service, production, managerial).
• Empirical strategy
  • Panel fixed‑effects regressions controlling for firm and year fixed effects to isolate within‑firm changes associated with AI adoption.
  • Mediation analysis to decompose total effects into occupational composition channels (technical/service vs production/managerial).
  • Moderation models interacting AI adoption with regional AI ecosystem and policy indices to test heterogeneous impacts.
  • Dynamic/event specifications to examine timing and persistence of effects around AI adoption measures.
• Identification notes
  • Firm fixed effects absorb time‑invariant heterogeneity; year effects account for macro shocks. Multiple AI proxies and robustness checks strengthen confidence in associations, though causality relies on plausibility of controls and timing tests reported by authors.

Implications for AI Economics

• Dual nature of AI: The paper provides micro‑level evidence that AI can be both creative (job creation, higher wages, occupational upgrading) and disruptive (increased within‑firm inequality, contraction of routine roles). Models of technological change should account for simultaneous scale (employment growth) and distributional (wage dispersion, occupational sorting) effects at the firm level.
• Importance of occupational channels: The mediation results emphasize that occupational reallocation — not just overall substitution/complementarity — is the dominant mechanism. Theories and empirical work should model heterogeneous task complementarity across occupations (technical/service vs routine production/management).
• Role of local ecosystems and policy: Regional AI industry development and supportive policies materially shape labor outcomes from AI adoption. Policies that build local skills, supplier networks, and complementary institutions can amplify positive labor impacts and mitigate inequality — suggesting place‑based policy design matters for equitable AI diffusion.
• Policy prescriptions
  • Invest in complementary human capital (technical/reskilling programs) to capture the job‑creation potential of AI.
  • Strengthen regional AI ecosystems (training, specialized suppliers, R&D networks) to improve firm capacity to translate AI into inclusive growth.
  • Monitor intra‑firm pay distribution and consider targeted redistribution/compensation mechanisms to address widening wage dispersion.
  • Gender‑aware workforce policies to sustain and amplify modest gains in female participation as task composition shifts.
• Research directions
  • Causal identification: future work could exploit exogenous shocks to AI adoption (procurement changes, policy rollouts) to more cleanly identify causal effects.
  • Worker outcomes: richer microdata on worker transitions, hours, and heterogenous wage trajectories would clarify welfare implications.
  • Generalizability: replicate in other countries/sectors to test external validity of the dual creative/destructive pattern.

(Reference: Qiao et al., "Creative disruption or destructive inequality? Firm-level evidence on AI adoption and employment dynamics", Humanit Soc Sci Commun — unedited article in press.)