AI spending boosts firms' innovation resilience up to a point but too much investment undermines it; the effect works through learning, coordination and R&D capacity and is strongest in state-owned, large and high-tech Chinese firms operating in dynamic environments.

Main Finding

The paper finds an inverted U-shaped relationship between firm-level AI investment and innovation resilience: AI initially strengthens firms’ innovation resilience but beyond a threshold its effect reverses and becomes detrimental. This nonlinear effect operates through three mediators—learning & absorption capacity, coordination & integration capacity, and technological innovation capacity—and is positively moderated by environmental dynamism. The AI → innovation-resilience effect is stronger for state-owned, high‑tech, and large firms. Higher innovation resilience is also linked to higher‑quality corporate development.

Key Points

1. Nonlinear effect
2. AI raises innovation resilience at low-to-moderate adoption levels but, past a critical point, additional AI investment reduces resilience (inverted U-shape).

3. Drivers of decline include algorithmic exclusion, organizational rigidity, increased management complexity, data/security risks, and resource dispersion.

4. Mediating mechanisms (all supported empirically)

5. Learning & absorption capacity: AI improves knowledge acquisition, assimilation, and transformation up to a point; excessive AI can cause information overload and dilute learning.
6. Coordination & integration capacity: AI can better coordinate internal/external resources initially; beyond saturation it creates mismatches with innovation needs and reduces flexibility.

7. Technological innovation capacity: AI enhances perception, ideation, and human–machine collaborative innovation until overreliance or resource misallocation undermines innovation capability.

8. Moderation by environment

9. Environmental dynamism (uncertainty) positively moderates the AI–innovation resilience link — the enabling effect of AI is stronger under more dynamic external conditions.

10. Heterogeneity

11. Effects are more pronounced in state-owned enterprises (SOEs), firms in high-tech industries, and large-scale firms.

12. Policy/business relevance

13. There exists an “optimal range” for AI deployment; firms should build complementary capabilities to capture AI benefits and avoid “technology application pitfalls.”

Data & Methods

• Sample: Chinese A‑share listed firms (Shanghai & Shenzhen), 2010–2024.
• Final dataset: 23,213 firm-year observations after excluding financial firms, ST/∗ST firms, and severe missing data; continuous vars trimmed at 1% to reduce outliers.
• Data sources: Innovation-resilience indicators from CNIPA; AI and firm-level covariates from CSMAR.
• Empirical model: Two-way fixed-effects panel regressions (firm and year effects).
• Dependent variable — Innovation Resilience (EIR):
  • Constructed from three dimensions: innovation input (R&D investment, researcher count), innovation output (patent applications), and innovation efficiency (patents / ln(R&D)).
  • Indicators combined using the entropy-weight method: EIR = ω1·RI + ω2·RE + ω3·PA + ω4·IE.
• Independent variable — AI:
  • Measured as (annual software investment + hardware investment) / total assets (relative AI investment intensity).
• Mediators:
  • Learning & Absorption (Ab): R&D expenditure intensity = total R&D expenditure / operating revenue.
  • Coordination & Integration (Co): negative coefficient of variation (CV) of R&D, capital, and advertising expenditures (lower CV → better coordination; sign adjusted so higher values indicate stronger coordination).
  • Technological Innovation (In): sum of Z-scores of R&D intensity and technical personnel ratio (technical staff / total employees).
• Moderator:
  • Environmental dynamism/uncertainty measured following prior literature (industry-adjusted metric; paper references [58] for exact construction).
• Additional analyses: Heterogeneity tests by ownership, industry (high-tech vs. others), and firm size; robustness checks reported in paper (data trimming, exclusions).

Implications for AI Economics

1. Marginal returns and policy design

2. AI exhibits diminishing and eventually negative marginal returns for innovation resilience. Policy instruments (subsidies, tax credits, public procurement) should encourage AI adoption up to the socially productive margin and emphasize complementary investments rather than pure scale-up.

3. Complementary capabilities matter

4. Economic gains from AI depend critically on firm capabilities (learning, coordination, technological skills). Subsidies or programs that couple AI adoption with training, organizational redesign, and R&D support will be more effective than technology grants alone.

5. Targeting and heterogeneity

6. One-size-fits-all AI promotion risks overinvesting in firms less able to integrate AI productively. Tailored support is warranted: smaller firms and non-SOEs may need additional assistance to realize benefits; large, SOE, and high-tech firms show stronger baseline gains but can also hit the diminishing-return threshold.

7. Regulation and governance

8. Negative effects at high AI intensity point to concerns around algorithmic exclusion, data security, and managerial complexity. Regulatory frameworks should address data governance, transparency, and workforce retraining to reduce the costs that cause the downturn in resilience.

9. Measurement and evaluation

10. For policymakers and researchers, measuring AI as investment intensity (software + hardware per assets) and tracking mediating capabilities allows more precise evaluation of AI’s economic effects than looking at AI adoption as a binary variable.

11. Research directions for AI economics

12. Quantify the threshold where AI’s marginal effect turns negative and study how that threshold shifts with complementary investments.
13. Evaluate welfare implications: do private optimal AI levels align with social optima, considering externalities (labor market, competition, data privacy)?
14. Extend analysis to different institutional contexts outside China to test generalizability.

If you want, I can extract the paper’s regression coefficients (e.g., estimated peak AI intensity) and robustness results if you provide the tables or full text of the results section.