A machine-learning screening tool can modestly identify potentially disruptive AI patents before their impact appears — an AdaBoost model trained on patent metadata flags top-5% CD disruptors with 27% precision and 43% recall. The method is best used to generate candidates for expert review, not to certify future market disruption.

Main Finding

An interpretable ensemble framework guided by the CD disruption index can modestly screen ex ante for potentially disruptive AI patents (defined as the top 5% of CD scores among patents with ≥1 backward citation) using early patent metadata. An AdaBoost classifier, after feature selection (17 → 12 features), achieves accuracy = 0.914, precision = 0.271, recall = 0.431, F1 = 0.333 on the main specification—useful for candidate generation but not definitive classification of realized disruption.

Key Points

• Label construction: Disruptiveness is operationalized as being in the top 5% of the empirical CD distribution restricted to patents that cite at least one prior patent. This focuses the target on patents with identifiable prior art and reduces sensitivity to extreme CD values from poorly referenced patents.
• Data scope: AI patents from the IncoPat database. Time-split design: training/learning period = patents filed 2007–2021; identification (scoring) period = 2022–2024, approximating realistic forecasting conditions.
• Feature set: Multidimensional indicators across three domains:
  • Technological: backward citations, non-patent literature references, IPC breadth, inventor team size, document length, family citation activity.
  • Market-oriented: simple & extended family size, transfer/licensing/pledging flags, litigation, number of applicants.
  • Legal: claim counts and first-claim length (independent/dependent claims).
• Modeling approach: Ensemble learning with adaptive boosting (AdaBoost) benchmarked against other classifiers (e.g., Random Forest). Feature selection reduced predictors from 17 to 12.
• Best-performing model: AdaBoost with selected features; performance indicates modest predictive power—higher accuracy driven by class imbalance; precision (27.1%) implies many false positives, recall (43.1%) shows some ability to recover true disruptive cases.
• Interpretable drivers: Citation- and disclosure-related signals are the most predictive (family citation activity, number of backward citations, document length among top features).
• Intended use: A scalable screening/candidate-generation tool for monitoring and expert review within the patented segment of AI innovation—not a substitute for longitudinal validation or expert judgment.

Data & Methods

• Data source: IncoPat patent records filtered to the artificial intelligence domain.
• Label: Top 5% of CD disruption index distribution among patents with at least one backward citation.
• Train/test split: Time-split to mimic forecasting—models trained on 2007–2021, applied to patents filed 2022–2024.
• Predictor variables: 17 initial indicators capturing technological positioning, market signals, and legal scope; final model used 12 after feature selection.
• Modeling/benchmarks: Adaptive boosting (AdaBoost) as primary ensemble method; compared to Random Forest and other standard classifiers. Emphasis on interpretable feature importance rankings.
• Evaluation metrics reported: accuracy, precision, recall, F1 (main spec: accuracy 0.914, precision 0.271, recall 0.431, F1 0.333).

Implications for AI Economics

• Practical monitoring tool: Policymakers, R&D managers, and investors can use the framework to triage patent portfolios and surface candidate inventions that are more likely to redirect technological search—saving expert time by focusing review resources.
• Complementarity with other signals: Given modest precision and recall, the model is best used alongside market intelligence, technical expert assessment, and firm-level context (e.g., commercialization plans, open-source disclosure) rather than as a sole decision rule.
• Informing resource allocation: The approach provides an early-warning signal for potential regime-shifting innovations, which can inform strategic R&D investments, portfolio rebalancing, and selective support for follow-on work in AI.
• Cautions for economic interpretation:
  • Retrospective dependence: CD is inherently defined by future citation patterns; converting it to a predictive label preserves some of that retrospective logic but cannot escape uncertainties in future citation behavior.
  • Sample and domain limits: Results pertain to patented AI inventions recorded in IncoPat; generalization to non-patented innovation, other technology domains, or jurisdictions with different citation practices may be limited.
  • Biases in patenting and citation practices: Large firms, strategic citations, or field-specific norms can distort both training labels and predictive features—risking systematic biases in who/what is surfaced as “potentially disruptive.”
  • Trade-off between false positives and misses: With precision ~27%, many flagged patents will not realize disruptive CD outcomes, so flagged sets should be treated as candidates for targeted follow-up rather than guarantees.
• Research and policy directions:
  • Extend to additional features (text embeddings of claims/abstracts, assignee network ties, market adoption proxies) to improve signal power.
  • Cross-domain validation to test generalizability and robustness of feature importance.
  • Use as part of active monitoring systems that update models as forward citation histories accrue, creating a feedback loop between ex ante screening and ex post validation.
  • Consider policy uses (e.g., targeted funding, standard-setting vigilance) but incorporate safeguards against over-reliance on patent-based signals alone.

Short conclusion: The paper contributes an operational bridge between CD-style disruption measurement and supervised early-screening methods, producing a transparent, scalable tool for surfacing candidate disruptive AI patents—valuable for monitoring and prioritization but requiring expert curation and further refinement before being used for high-stakes allocation decisions.