Data and methods for identifying artificial intelligence-related patents

This paper evaluates existing approaches to identifying artificial intelligence (AI)-related patents and introduces a novel, scalable framework for improving classification performance. Motivated by growing reliance on patent data in innovation research, we assess widely used methods, including patent class-based approaches and the USPTO’s Artificial Intelligence Patent Dataset (AIPD), with an independent, human-expert-annotated ground-truth dataset. We document substantial performance limitations in existing approaches, particularly in terms of precision and generalizability. To address these challenges, we develop a CPC-informed, iterative positive-unlabeled (PU) learning framework for constructing high-quality training data. Our approach integrates hierarchical patent classification with data-driven refinement procedures to reduce label noise and improve representativeness. Using this refined dataset, we train a range of machine learning, deep learning, and transformer-based models. Our results show that models trained within our framework significantly outperform existing methods, including AIPD, achieving improvements over AIPD in F1 scores of approximately 18–21% on the same benchmark dataset. These gains are primarily driven by enhanced precision without sacrificing recall, highlighting the central role of training data quality in classification performance. We further demonstrate the empirical value of improved AI patent identification through two applications, showing that the release of ChatGPT increased both the market valuation of AI patents and firms’ allocation of innovative effort toward AI technologies. To support future research, we release our training data, source code, and patent-level predictions, with ongoing updates to reflect the evolving nature of AI innovation.