Firms with stronger AI development shift income toward capital and away from labor, but the effect varies across industries and ownership types and is dampened by financing constraints.

Main Finding

The paper finds that enterprise-level artificial intelligence (AI) adoption shifts income distribution toward capital at low-to-moderate adoption levels but can reverse that effect at high adoption intensity. Empirically, AI usage (measured by AI-related keyword frequency in annual reports) has a significantly negative effect on employee income share, while a positive AI² term implies a U‑shaped relationship: early-stage AI adoption compresses labor remuneration; beyond a high threshold, further AI investment raises employee compensation through productivity and profitability gains. Results are robust to time and region splits, alternative dependent variables, and instrumental-variable estimation.

Key Points

• Core result: AI → lower labor share (negative linear term), but AI² > 0 → U-shaped effect (first decrease, then increase).
• Magnitude/time dynamics:
  • Sample: 2010–2022, China A-share non-financial listed firms (27,984 firm-year observations).
  • Negative AI effect is stronger after 2017 (post–national AI strategy), suggesting accelerating impact over time.
• Robustness:
  • Results persist when excluding municipal firms, replacing the dependent variable, and splitting pre/post-2017.
  • Endogeneity addressed via 2SLS using lagged AI as an instrument (first stage strong).
• Controls and correlates:
  • Larger firm size and higher ROE → higher employee income share.
  • Higher leverage and larger board size → lower employee income share.
• Heterogeneity & mechanism emphasized (discussed conceptually and through subgroup analysis):
  • Financing constraints (SA index) moderate effects: financially constrained firms are more likely to substitute capital for (low-skilled) labor, amplifying compression; unconstrained firms can scale productivity-enhancing AI and may share gains with employees.
  • Ownership structure and industry produce different responses (paper runs group regressions by ownership and industry).

Data & Methods

• Sample: Shanghai and Shenzhen A‑share, non-financial listed firms, 2010–2022; final N ≈ 27,984 observations after exclusions (ST/*ST firms, financials, nonpositive employee cash payments) and 1% winsorization.
• Key variables:
  • Dependent variable (indis): net profit / cash paid to and for employees — interpreted as a firm-level indicator of profit vs. labor remuneration.
  • Main explanatory (AI): count of AI-related keywords in annual reports (NLP/text frequency approach).
  • Moderator: financing constraint measured by SA index.
  • Controls: size (log assets), leverage, ROE, firm age, largest shareholder share, CEO-chair duality, board size; industry, region, and year fixed effects.
• Empirical strategy:
  • Baseline panel regressions with fixed effects and clustered inference.
  • Nonlinear specification includes AI² to capture curvature.
  • Robustness checks: pre/post-2017 split, exclusion of municipalities, alternative dependent variable.
  • Endogeneity: 2SLS using lagged AI as instrument; Kleibergen–Paap and Stock–Yogo tests reported to support instrument strength.
• Data sources: CSMAR (financials), annual reports (text), China City Statistical Yearbook and local statistical bureaus (regional data).

Implications for AI Economics

• Technology-biased technical change: The findings are consistent with AI being capital-augmenting and skill‑biased. At early/medium adoption, AI substitutes low-skilled labor and increases capital returns, shifting income to capital and compressing wage share. At high adoption intensity, aggregate productivity and profits can be large enough for firms to raise employee compensation, producing the observed U-shape.
• Role of finance: Financing constraints shape firms’ technology choices and distributional outcomes. Credit-constrained firms may pursue labor-saving but low‑scale AI investments (wage compression), while better-financed firms can deploy AI at scales that increase profits and may raise wages — implying that access to finance mediates the distributional effects of AI.
• Policy and firm-level implications:
  • For policymakers: combine industrial/innovation support with distributional policies (retraining, active labor market programs, progressive taxation, incentives for firms to share productivity gains) to mitigate short-run wage compression.
  • For financial and industrial policy: targeted financing and subsidies for AI adoption in ways that encourage complementary investments in worker skills and higher‑value tasks could reduce adverse distributional effects.
  • For firms: governance and compensation design matter; larger, more profitable firms are more likely to sustain higher labor shares.
• Research implications:
  • Importance of measuring AI intensity and its nonlinear effects when studying broad macro- and firm-level distributional outcomes.
  • Need to consider heterogeneity by firm finance, ownership, industry, and worker skill composition.

Limitations (noted / implied)

• AI measure: keyword-frequency approach captures disclosed strategic attention to AI but may imperfectly measure actual AI investment or operational deployment.
• Sample: listed, non-financial Chinese firms — results may not generalize to SMEs, informal sector, other countries, or non-listed firms.
• Identification: lagged-AI IV helps but may not fully eliminate all endogeneity (e.g., persistent unobserved shocks correlated with both AI adoption and pay policy).
• Dependent variable (net profit / cash to employees) is an aggregate firm-level ratio and does not reveal worker-level heterogeneity (by skill, occupation, or wage percentile).

Suggestions for Follow-up Research

• Worker-level outcomes: link firm AI measures to wage distribution, employment transitions, and occupation-level impacts.
• Better causal designs: exploit exogenous shocks (policy rollouts, localized AI infrastructure changes) for stronger identification.
• Cross-country or SME-focused analyses to test generalizability.
• Investigate mechanisms in more detail: task content changes, complementarities between AI and high-skilled workers, and channels from profitability to compensation policy.