AI adoption among Chinese listed firms is linked to weaker corporate resilience — largely because it raises financing constraints, R&D intensity and management costs; the negative effect is strongest in the eastern region, non-manufacturing firms and heavily polluting industries.

Main Finding

Using 20,649 firm-year observations for Chinese A‑share firms (Shanghai & Shenzhen) from 2014–2023, the paper finds that greater corporate use of artificial intelligence (measured primarily as the log frequency of AI‑related words in annual reports) has a small but statistically significant negative effect on measured enterprise resilience. This negative effect is robust to endogeneity and robustness checks and operates (at least partly) through: higher financing constraints, greater R&D investment intensity, and higher management expense ratios. The inhibitory effect is stronger for firms in the eastern region, for non‑manufacturing industries, and for heavily polluting firms.

Key Points

• Hypothesized channels (and supported empirically):
  • Financing constraints: AI investments are high‑sunk, long‑tailed and increase information asymmetry → higher KZ index → weaker resilience.
  • R&D intensity: AI prompts concentrated R&D spending → reduced technological diversity and strategic flexibility → weaker resilience.
  • Management expense ratio: AI adoption raises talent, coordination, compliance costs → higher Mfee → organizational rigidity and attention crowding → weaker resilience.
• Main estimated coefficients (baseline):
  • Lnwords → Res: coefficient ≈ −0.0012 to −0.0013 (significant at 1%).
• Heterogeneity:
  • Stronger negative effect in eastern China vs other regions.
  • Stronger in non‑manufacturing firms than manufacturing.
  • Stronger among heavily polluting firms.
• Robustness and identification:
  • Results survive 1% winsorization, alternative AI measures (patents, MD&A keyword counts), instrumental variable approaches and propensity score matching (PSM) for endogeneity concerns.

Data & Methods

• Sample: A‑share listed companies in Shanghai and Shenzhen, 2014–2023; ST/PT and financial firms excluded; final N = 20,649 observations. Data sources: CSMAR and CNRD.
• Key variables:
  • Dependent — Enterprise resilience (Res): entropy‑weighted composite of 3‑year cumulative operating revenue growth (growth) and 1‑year monthly stock‑return volatility (volatility).
  • Main explanatory — AI adoption (Lnwords): log(1 + frequency of AI words in annual report). Alternative measures: log(1 + AI patents), log(1 + MD&A AI keywords).
  • Mediators: KZ index (financing constraints), RDintensity (R&D expenditure / operating revenue), Mfee (management expenses / operating revenue).
  • Controls: Size (ln assets), Lev, ListAge, ROA, Quick ratio, CEO‑chair duality, Growth (sales growth), plus year and city fixed effects.
• Empirical strategy:
  • Baseline: panel regressions with year and city fixed effects.
  • Mechanism tests: standard mediation regressions (AI → mediator; mediator → Res controlling for AI).
  • Endogeneity checks: instrumental variable methods (details referenced but not fully reproduced in text) and propensity score matching.
  • Robustness checks: alternative AI measures, different specifications; 1% winsorization applied to variables.

Implications for AI Economics

• Conceptual: Adds nuance to the predominantly optimistic literature on firm‑level AI benefits by documenting measurable downside effects on firm resilience arising from resource reallocation, higher financing friction, and increased management costs.
• For microeconomic modeling of AI adoption:
  • Models should incorporate non‑monotonic effects of AI on firm welfare — short‑term costs, financing frictions, and increased fixed/management costs can reduce shock‑absorption capacity even if productivity rises.
  • Consider endogenous financing constraints and liquidity as key margins affecting adoption paths and firm survival.
  • Introduce heterogeneous firm characteristics (region, industry, pollution intensity) to capture differential effects.
• For empirical work:
  • Be cautious using textual proxies for AI adoption — combine with patent, investment and capex data where possible; treat resilience as multi‑dimensional (growth vs volatility).
  • Use dynamic and causal identification strategies to separate short‑run disruption effects from longer‑run productivity gains.
• Policy and managerial implications:
  • Firms: stage AI adoption; protect liquidity buffers; diversify R&D portfolios; monitor management cost growth; invest in upskilling and stakeholder engagement to reduce social frictions.
  • Policymakers/Investors: design targeted financing instruments (subsidies, patient capital) for AI projects, strengthen governance/standards to lower compliance costs, and support regions/industries with higher vulnerability.
• Directions for future research:
  • Longer‑run follow‑up to detect whether negative resilience effects persist or reverse as AI investments mature.
  • Cross‑country comparisons to test generalizability beyond Chinese listed firms.
  • Microdata on AI capex, personnel allocation, and project outcomes to better separate investment intensity from effective implementation.
  • Welfare assessment: net effect of AI on firm survival, employment, and social welfare when accounting for resilience dynamics.

Limitations noted in the paper (implicit): resilience metric is constructed from revenue growth and stock volatility and may not capture all resilience dimensions; textual AI measures are imperfect proxies; findings pertain to listed Chinese firms and may not generalize to SMEs or other countries.