AI invention has clustered in well-funded institutions: government backing and corporate R&D are the strongest drivers of concentration, while disruption from new AI patents has declined. This growing intellectual-monopoly pattern raises concerns about reduced technological turnover and uneven control of AI advancement.

Main Finding

Using a large USPTO AI-patent dataset (1976–2020) and a new pairwise disruption index (PDI), the authors show that AI inventions have become both less disruptive and increasingly concentrated in well‑resourced institutions. Macro‑level forces—especially government support and organizational R&D capacity—are the dominant drivers of this concentration; micro‑level factors (team size, inventor human capital) matter but to a lesser degree. Government funding has the largest effect overall, and organizational R&D capacity has grown in importance since ~2010.

Key Points

• Trend: A long‑run decline in technological disruption in AI and rising concentration of AI patents in a small set of well‑resourced organizations (consistent with intellectual monopoly capitalism).
• Measurement: Introduces a pairwise disruption index (PDI), a variation on CD/disruption metrics, to quantify how disruptive individual patents are.
• Drivers tested: Three dimensions drawn from literature — (1) government support/governance and policy, (2) organizational R&D capacity/knowledge reservoir, (3) inventor human capital and team structure.
• Relative effects:
  • Macro forces (government support, organizational R&D) have stronger effects on concentration than micro forces.
  • Government support is the single most substantial predictor of concentration.
  • Organizational R&D capacity’s influence has become increasingly dominant in recent years (post‑2010).
  • Micro factors such as larger R&D team size and more experienced inventors contribute to concentration but less so than macro factors.
• Robustness: Results are reported as robust to alternative citation thresholds and specifications (authors report robustness checks).

Data & Methods

• Core dataset: USPTO Artificial Intelligence Patent Dataset (Giczy et al. style), originally ~668,808 AI patents/PGPubs identified via ML across 1976–2020, covering eight AI technology families (e.g., NLP, AI hardware, computer vision).
• Analysis sample: Focus on granted patents with substantive impact — retained 367,441 patents published 1976–2018 (citations measured up to five years after publication; citations for 2018 tracked to 2019–2023).
• Supplementary USPTO data: government interest (federal agency funding occurrences), patent assignee (organizational owners), inventor records, and patent citation flows.
• Disruption metric: Pairwise Disruption Index (PDI) — a novel, network‑based disruption measure derived from citation relationships (conceptually related to the CD index).
• Empirical strategy: Statistical models (panel/pooled regressions and temporal analyses) estimating how government support, organizational R&D capacity, and inventor/team attributes predict patent disruption and concentration; temporal decomposition to assess changes in effect sizes over decades.
• Limitations noted: USPTO patent focus excludes much open‑source / non‑patented AI work (e.g., model releases on GitHub), and is US‑centric; patents are an imperfect but meaningful proxy for high‑impact, application‑oriented AI inventions.

Implications for AI Economics

• Innovation dynamics: The observed decline in disruption suggests slowing of foundational, paradigm‑shifting AI breakthroughs (at least as captured by patented inventions), with more incremental advancement accumulating within incumbents.
• Market structure and welfare:
  • Rising concentration amplifies intellectual monopoly effects: incumbents accumulate proprietary data, human capital, and R&D capacity, raising entry barriers and potentially reducing competition, diffusion, and aggregate social returns to innovation.
  • Distributional concerns: Concentration concentrates rents (and power) among a narrow set of firms/governments, with implications for labor markets (wage/power imbalances), access to technology, and political influence.
• Role of policy:
  • Government funding is a double‑edged sword: it can catalyze high‑impact work but also help entrench dominant actors if funding and procurement disproportionately favor large incumbents.
  • Policy levers to consider: more dispersed public funding (grants to startups, universities, smaller labs), conditions on public funding to encourage openness/interoperability, stronger antitrust enforcement tailored to intangible‑asset markets, and policies promoting talent mobility and knowledge diffusion.
• Research and measurement:
  • PDI offers a tool for tracking disruptive vs. incremental innovation in AI over time and across subdomains; can inform real‑time monitoring of innovation concentration.
  • Future work should integrate non‑patent innovation channels (open‑source models, preprints, code repositories) and extend analysis beyond U.S. patents to capture global AI dynamics.
• Practical takeaway for economists and policymakers: to preserve a vigorous, disruptive AI innovation ecosystem and mitigate monopoly harms, interventions should target the concentration mechanisms identified here—funding allocation, firm R&D accumulation, and institutional barriers to talent and knowledge diffusion.