Analysis of AI patents (2002–2021) reveals four distinct classes of AI inventions and the topical pathways through which scientific research is incorporated into technology; the mapped science-to-technology links offer targeted insights for corporate R&D strategy and innovation policy.

Main Finding

AI technologies differ in how scientific knowledge enters and shapes them depending on their structural role in the technology ecosystem. Technologies that act as brokers (connecting disparate CPC classes) tend to draw on scientific literature selectively and contextually to meet specific technological needs, whereas low-brokerage (more peripheral or specialized) technologies follow more science-driven development paths. Methodological and performance-enhancing AI innovations (core technique improvements) serve as primary conduits for transferring scientific knowledge into technological applications.

Key Points

• The paper develops a two-stage, semantic science→technology exploration framework tailored to AI:
  • (1) classify AI patents into four categories using centrality measures from a CPC co‑occurrence network to identify core vs. brokerage roles; and
  • (2) perform semantic topic extraction on patents and their cited scientific articles using BERTopic, with topic labels generated via generative AI.
• Data cover USPTO patents (PATSTAT Spring 2024) identified as AI-related using an updated WIPO hybrid IPC/CPC+keyword strategy: 321,683 AI patents (2002–2021), segmented into four 5-year periods (2002–06: 25,246; 2007–11: 39,181; 2012–16: 85,543; 2017–21: 171,713).
• Scientific publications cited by those patents were matched from Web of Science (WoS) using PATSTAT citation strings.
• Main empirical patterns:
  • Brokerage technologies selectively appropriate scientific results in context-specific ways (they recombine knowledge across domains rather than relying on deep, continuous scientific lines).
  • Low-brokerage technologies show stronger, more direct science→technology linkages (science provides foundational inputs).
  • Methodological advances and performance-oriented improvements within AI (e.g., algorithmic, optimization, training techniques) are central nodes in the flow from science to technology.
• The study emphasizes heterogeneity within a single domain (AI): not all AI subfields or patent types depend on science in the same way.

Data & Methods

• Data sources:
  • PATSTAT (Spring 2024) for USPTO patent applications (2002–2021).
  • Web of Science for scientific publications cited in those patents.
• AI patent identification:
  • Hybrid WIPO-derived search combining CPC/IPC codes and keywords, updated to reflect classification revisions; validated through expert consultation and benchmarks.
• Patent taxonomy:
  • Built a CPC co‑occurrence network from patent classification co‑mentions and computed centrality measures to assign each technology/patent to one of four structural categories (distinguishing core vs. broker roles; exact centrality metrics reported in paper).
• Semantic analysis:
  • Extracted abstracts of patents and cited papers.
  • Applied BERTopic (topic modeling suited to short texts / contextual embeddings) to discover thematic structures.
  • Used generative AI to produce human-interpretable topic labels for discovered topics.
• Temporal analysis:
  • Tracked topic prevalence and science→technology citation patterns across four five-year periods (2002–2021).
• Limitations noted by authors:
  • Patent-to-paper citation is an imperfect proxy for knowledge flow (context/purpose of citation varies).
  • WoS access/license restrictions limit public sharing of raw bibliometric records.
  • Analysis is limited to USPTO patent filings; other jurisdictions and non-patent knowledge flows may differ.
  • Some methodological choices (classification thresholds, centrality metrics, generative labeling) may influence results.

Implications for AI Economics

• For firms / R&D strategy:
  • Investing in methodological and performance-enhancing AI research yields high leverage for technological downstream impact — these areas act as multipliers for science→technology transfer.
  • Firms occupying brokerage roles should prioritize absorptive capacity and targeted scientific scouting: they benefit more from selective, contextual scientific inputs than from broad investments in foundational science.
  • Firms focused on low-brokerage technologies should strengthen sustained ties to relevant scientific fields; deeper science linkages correlate with more foundational, long-term technological value.
• For policymakers:
  • Policies that support methodological AI research (core algorithmic and performance improvements) can produce wide spillovers across application domains.
  • Supporting interdisciplinary research and boundary-crossing collaborations can catalyze broker technologies that accelerate diffusion across sectors.
  • Tailored science funding and translation programs are warranted: some subfields need deep, continuous science investment, others benefit more from targeted translational support and cross-domain convening.
• For measurement and evaluation in AI economics:
  • Combining network-based structural classification (core vs. broker roles) with semantic topic extraction provides richer indicators of how science feeds technology than raw citation counts alone.
  • Patent-to-paper citation semantics matter — semantic/qualitative analyses help identify whether citations indicate foundational reliance or selective, instrumental use.
• Market / industrial organization effects:
  • Because methodological improvements propagate broadly, firms controlling core-method IP may obtain disproportionate economic rents and shape industry trajectories.
  • Brokerage technologies may enable faster cross-sector diffusion of AI, changing competitive dynamics and increasing returns to firms that can integrate diverse technical components.

Practical caveats: the study’s conclusions rely on patent–paper citations and USPTO coverage; real-world knowledge flows also include tacit exchange, open-source contributions, and conference dissemination (important in AI), which may not be fully captured.