Chinese listed firms that embed AI more visibly in disclosures form denser collaboration networks and account for a larger share of joint patents, implying AI boosts ecosystem innovation chiefly by improving partner matching and network centrality.

Main Finding

AI embeddedness within firms (measured from corporate texts using TF–IDF and BERT) is positively associated with stronger collaborative-network positions (higher density, centrality, and joint-patent ratios) and higher value co-creation performance. Crucially, AI raises innovation value primarily by improving cooperation structure and knowledge/resource matching in the innovation ecosystem — the effect is amplified when collaborative mechanisms (network density, centrality, betweenness, cooperation frequency) are stronger.

Key Points

• Sample: 480 Chinese A‑share listed firms, 2020–2025 (2,880 firm‑year observations); excluded ST firms, financial firms, firms listed <3 years, and observations with missing/mismatched data.
• AI embeddedness measure: semantic intensity from annual reports, ESG reports and announcements — TF–IDF × BERT semantic probability, length‑corrected.
• Collaborative mechanism index: weighted composite of network density, degree centrality, betweenness centrality, and annual cooperation frequency (NetworkX on joint patents and cooperation announcements).
• Value co‑creation performance: standardized, weighted combination of invention patent counts, joint patent ratio, patent citations, and textual intensity of new product revenue.
• Descriptive network differences by AI terciles:
  • Network density: low 0.026 → medium 0.041 → high 0.068
  • Mean degree centrality: 0.043 → 0.067 → 0.104
  • Joint patent ratio: 4.8% → 7.6% → 12.9%
• Panel regression (firm and year fixed effects; controls: firm size, leverage, R&D intensity, age, industry; winsorized 1/99):
  • AI embeddedness coefficient = 0.083 (positive)
  • Collaborative mechanism coefficient = 0.126 (positive)
  • Interaction AI × Collaborative mechanism = 0.041 (positive) — indicates complementarity
• Nonlinear/explainable ML: XGBoost + SHAP feature importance:
  • Top mean SHAP contributions: network centrality 0.186, AI textual intensity 0.159, cooperation announcement frequency 0.142, network density 0.126.
  • Network structure variables generally explain more than firm size, implying ecosystem position matters more than scale alone.

Data & Methods

• Data sources: annual reports, ESG reports, CNINFO announcements, joint patent records and public patent metadata for Chinese A‑share firms.
• Text pipeline: Python preprocessing (segmentation, stopword removal, keyword expansion), TF–IDF weighting, BERT sentence classification to filter true AI application mentions; combined into length‑adjusted AI embeddedness index.
• Network construction: joint patents and cooperation announcements used to build inter‑firm collaboration networks; NetworkX computed density, degree centrality, betweenness centrality; these feed into a weighted collaborative mechanism index.
• Econometrics:
  • Two‑way fixed effects panel models to estimate AI and collaborative mechanism effects on value co‑creation, including interaction term.
  • Controls: firm size, leverage, R&D intensity, age, industry dummies.
  • Robustness/heterogeneity: winsorization, sample exclusions as above.
• Explainable machine learning:
  • XGBoost predictive model for value co‑creation.
  • SHAP decomposition to quantify marginal contributions of predictors and capture potential nonlinearities and variable importance.

Implications for AI Economics

• Mechanism focus: AI raises innovation value not mainly by automating tasks within firms but by reshaping inter‑firm information flows, partner matching, and resource allocation across ecosystems. Economic models of AI should therefore include network position and data‑sharing externalities as key channels.
• Complementarity and returns: The positive interaction implies AI investments deliver larger returns when firms occupy central, dense network positions or when collaborative mechanisms are strong — suggesting increasing returns and potential concentration effects (firms already central may capture disproportionate benefits).
• Distributional & market structure effects: Because AI enhances partner‑matching and coordination, it may intensify winner‑take‑most dynamics in innovation ecosystems. Policymakers and competition economists should monitor data access, platform governance, and collaboration asymmetries that could amplify market power.
• Policy and governance: Effective value co‑creation requires attention to data openness, trust and IP rules, algorithmic transparency, and benefit‑sharing arrangements. Regulations promoting interoperable data standards, fair data sharing, and governance frameworks could affect how AI translates into ecosystem value.
• Measurement contribution: The paper provides a practical text‑based method (TF–IDF × BERT × length correction) to quantify firm‑level AI embeddedness from disclosures — useful for empirical AI economics research where direct measures of AI use are scarce.
• Research gaps & cautions:
  • Causality: The analysis is observational; reverse causality and selection (central firms investing more in AI) are possible. Causal identification (instruments, natural experiments, phased adoption) is needed to establish directional effects.
  • External validity: Sample is Chinese listed firms in selected sectors; results may differ across countries, institutional contexts, or smaller/private firms.
  • Unobserved heterogeneity: Data sharing policies, contractual terms, and informal ties are hard to observe but may drive results.
• Suggested follow‑ups: causal designs to isolate AI’s effect on network formation and innovation outcomes; cross‑country comparisons; microdata on data‑sharing contracts and platform matching algorithms; welfare analysis of concentration vs. innovation gains; labor and skill‑distribution consequences of ecosystem‑level AI adoption.