Artificial Intelligence and the Reconfiguration of Innovation Collaboration: Evidence from Innovation Networks in China 

Technical Summary: Artificial Intelligence and the Reconfiguration of Innovation Collaboration

Problem Statement
While Artificial Intelligence (AI) is widely recognized as a general-purpose technology driving innovation, its specific impact on the structure of inter-city innovation collaboration remains theoretically ambiguous. Existing literature offers conflicting views: AI tools may reduce information search costs and improve partner matching, potentially fostering collaboration; conversely, they may enhance individual innovation capacity and knowledge spillovers, thereby reducing the necessity for formal partnerships. This paper investigates whether the development of AI technology reshapes formal, legally recorded inter-city innovation networks. Specifically, it asks whether AI development promotes or inhibits collaboration measured through joint patent applications among Chinese cities, a setting where alternative, informal channels of knowledge access are becoming increasingly available.

Methodology
The study employs a comprehensive dataset covering 284 prefecture-level and above cities in China from 2010 to 2022.

• Data Sources: The analysis utilizes patent data from the China National Intellectual Property Administration (CNIPA) and business registration data from the State Administration for Industry and Commerce. City-level economic controls are drawn from the China City Statistical Yearbook.
• Dependent Variable: Inter-city innovation collaboration is measured by the frequency of joint patent applications between city pairs. The dataset is constructed as an undirected, non-duplicated matrix of city pairs (upper-triangular), excluding diagonal elements, resulting in over 40,000 observations.
• Independent Variable: AI development is quantified using the cumulative stock of AI-related patent applications (based on the Key Digital Technology Patent Classification System) over a two-to-six-year lag window to account for technological iteration and commercialization lags.
• Empirical Strategy: The baseline specification is a regression model with city-pair fixed effects and year fixed effects. To address endogeneity concerns, the authors employ:
  1. Heckman Two-Stage Model: To correct for sample selection bias, using distance and cross-province indicators as exclusion restrictions.
  2. Instrumental Variable (IV) Approach: Using the interaction between a city's 2005 computer/software industry employment ratio and its cumulative AI patent stock as an instrument.
• Mechanism Testing: The study tests two primary channels:
  1. Cost Effect: Examining how AI-driven specialization and centralization increase coordination costs, measured via interaction terms with technological, AI-level, and infrastructure similarity.
  2. Substitution Effect: Analyzing whether AI-enhanced knowledge spillovers (measured by patent citations and localization indices) crowd out formal collaboration, particularly in regions with stringent intellectual property (IP) protection.

Key Results

• Primary Finding: Contrary to the hypothesis that AI facilitates collaboration by lowering search costs, the results indicate that AI development significantly reduces formal inter-city innovation collaboration. This negative association holds across various specifications and robustness checks.
• Mechanisms:
  • Cost Channel: AI development intensifies urban specialization and centralization. As cities diverge in their technological paths and AI capabilities, cognitive distances widen, and coordination costs rise. The negative effect is mitigated when cities have high technological, AI, and infrastructure similarity, supporting the cost-channel hypothesis.
  • Substitution Channel: AI expands the efficiency and reach of knowledge spillovers, allowing cities to access external knowledge through informal channels (e.g., patent mining, literature search) rather than formal joint patenting. This substitution effect is stronger in regions with high IP protection, where the cost of formal collaboration is higher.
• Heterogeneity: The inhibitory effect is not uniform:
  • Collaboration Quality: The suppression is more pronounced for high-quality collaborations (invention patents) compared to lower-quality ones (utility models).
  • Geography: The negative effect is strongest for cross-provincial partnerships and collaborations between eastern and western cities, where technological disparities and coordination costs are highest. It is weakest within the central region.
  • Actor Type: The effect is most severe for inter-firm collaboration, likely due to intensified competitive pressures and reluctance to share core technologies, whereas university and research institute collaborations are less affected.

Key Contributions

1. Meso-Level Perspective: The study shifts the focus from firm-level outcomes (growth, productivity) to the meso-level structure of innovation networks, revealing a reconfiguration where stronger autonomous innovation capacity may coexist with weaker formal inter-city ties.
2. Integrated Framework: It synthesizes classical concepts—the "burden of knowledge," "absorptive capacity," and "cognitive distance"—into a unified framework explaining how a single technological shock (AI) can simultaneously enhance local innovation while disrupting regional collaboration architectures.
3. Empirical Evidence on Substitution: The paper provides empirical evidence that AI can act as a substitute for formal collaboration by enhancing the efficiency of informal knowledge access, challenging the assumption that technological progress linearly generates positive network externalities.

Significance and Policy Implications
The paper argues that the finding is counterintuitive: technological progress enhances individual efficiency while weakening incentives for inter-agent collaboration. This suggests that innovation activities may become increasingly fragmented, potentially impeding breakthroughs that require integrating knowledge from multiple domains.

The authors caution against a purely technology-centric view of AI policy. Instead, they propose that regional innovation policies must account for the resilience and connectivity of innovation networks. Specifically:

• Institutional Frictions: Policy should address the institutional frictions that make cross-regional collaboration costly, such as complex IP arrangements and benefit-sharing mechanisms.
• Differentiated Strategies: Policies should be tailored to regional contexts. For eastern cities and inter-firm collaborations, the focus should be on reducing transaction and knowledge-management costs that make collaboration less attractive when informal access is available. For central and western cities, the priority is strengthening absorptive capacity to prevent further marginalization in high-quality innovation networks.

Limitations
The authors explicitly acknowledge that the study measures only formal, patent-based collaboration. It does not capture non-patent forms of collaboration prevalent in the AI era, such as open-source communities, model-sharing platforms, algorithm competitions, or informal knowledge exchange. Therefore, the findings reflect a reconfiguration of formal ties rather than a total reduction in all innovation collaboration. Future research is needed to integrate these non-patent channels to understand the full scope of AI's impact on innovation networks.