Rising AI capacity in Chinese cities correlates with a fall in formal inter-city patent collaborations as specialization and richer spillovers raise coordination costs and substitute for joint projects; the decline is largest for high-quality, cross-provincial and eastern-region partnerships.

Main Finding

AI development in China (2010–2022) is associated with a statistically significant reduction in formal inter‑city innovation collaboration (measured by joint patent applications). Two mechanisms drive this effect: (1) a cost effect—AI increases urban specialization and centralization, raising coordination and matching costs for cross-city partnerships; and (2) a substitution effect—AI expands efficient knowledge spillovers and lowers the marginal value of formal, legally recorded collaborations. The negative impact is stronger for high‑quality collaborations, cross‑provincial ties, eastern cities, and inter‑firm partnerships.

Key Points

• Scope and outcome: Uses joint patent applications among 284 Chinese prefecture-level (and above) cities over 2010–2022 to capture formal, codified inter‑city collaboration.
• Core result: Greater AI development at the city level correlates with fewer formal collaborative patent ties between cities.
• Mechanisms:
  • Cost channel: AI fosters deeper specialization and agglomeration in hub cities, increasing cognitive distance, information asymmetries, negotiation and monitoring costs, and reducing incentives to form external partnerships.
  • Substitution channel: AI reduces search/prediction costs and broadens spillovers (e.g., through smarter literature/patent mining), enabling cities to access and recombine external codified knowledge without formal joint projects.
• Heterogeneity: The dampening effect is more pronounced for:
  • High‑quality collaborations (e.g., more valuable patents),
  • Cross‑provincial city pairs (administrative borders intensify frictions),
  • Eastern (more advanced) cities,
  • Inter‑firm collaborations (versus university or public‑sector ties).
• Conceptual framing: Integrates "burden of knowledge", absorptive capacity, and cognitive distance to explain how a single technological shock can boost local autonomy while weakening formal network ties.
• Limitation: Analysis is limited to patent‑recorded collaborations and cannot observe informal, platform‑based, or open‑source collaboration channels that AI may also enable.

Data & Methods

• Data: City‑level patent records covering joint patent applications among 284 Chinese cities from 2010 to 2022.
• Outcome measure: Formal inter‑city collaboration operationalized as joint patents filed by agents located in different cities.
• AI measure: City‑level development of AI technology (paper uses a city AI development indicator; specific construction details are in the full paper).
• Empirical approach: Panel analysis of city‑pair collaboration networks over time, with mechanism tests and heterogeneity analysis to probe cost and substitution channels (the paper explicitly acknowledges and tests for specialization/centralization and enhanced spillovers as mediators).
• Robustness/limitations: Results pertain to patent‑based collaborations; non‑patent collaboration (remote development, open‑source, unrecorded consulting) is not observable and thus outside the empirical scope.

Implications for AI Economics

• Rethinking network externalities: AI can simultaneously raise individual innovation productivity and weaken formal collaborative networks, implying non‑monotonic effects of technological progress on aggregate innovation that models should capture.
• Regional inequality and policy: AI‑driven centralization can exacerbate core–periphery divides. Policies should support absorptive capacity and connectivity in peripheral regions to prevent fragmentation of innovation systems.
• Trade‑offs in innovation strategy: Firms and regions may substitute informal, low‑cost spillovers for formal partnerships—this affects how we evaluate social returns to AI and the design of incentives for joint ventures, cross‑regional consortia, and shared infrastructure.
• Policy interventions to preserve collaboration value:
  • Strengthen cross‑regional coordination (subsidies or tax incentives for cross‑provincial joint R&D).
  • Invest in absorptive capacity (training, mobility programs, joint labs) in lagging cities to make formal collaboration feasible and valuable.
  • Support mechanisms that lower coordination costs (standardized data/process platforms, trusted IP frameworks, mediated matchmaking).
  • Balance open knowledge diffusion and incentives for formal collaboration (e.g., selective funding for collaborative projects addressing interdisciplinary grand challenges).
• Modeling and measurement: Empirical and theoretical work in AI economics should incorporate network structure, tacit vs. codified knowledge channels, and the potential for AI to alter the returns to different collaboration modalities.

Overall, the paper highlights a potentially unintended consequence of AI: improved access to codified knowledge and stronger local capabilities can reduce incentives for formal cross‑regional collaboration, with implications for innovation policy, regional development, and how economists model technological change and network effects.