Smith Researchers Develop More Accurate Way to Identify AI Patents, Release New Dataset

A newly published study in Research Policy substantially improves methods for identifying and tracking AI-related patents, offering clearer insight for researchers, policymakers, and industry analysts.

The paper — “Data and methods for identifying artificial intelligence‑related patents” — is authored by University of Maryland Robert H. Smith School of Business researchers  Waverly Ding, associate professor of management and organization, and Kunpeng Zhang, associate professor of information systems, along with Chao Min, Guolong Wang and Tianjun Wu of Nanjing University. The work presents a machine-learning approach that significantly improves the accuracy of identifying AI-related patents.

Patent data plays a central role in tracking technological progress. But the authors show that some of the most widely used resources — especially the U.S. Patent and Trademark Office’s Artificial Intelligence Patent Dataset (AIPD) — over-identify AI patents. Compared with a separate, human-reviewed benchmark, the AIPD mislabels nearly 19% of all patents issued between 1976 and 2023 as AI-related. 

To tackle this problem, the team developed a new method that combines Positive-Unlabeled (PU) learning with a Cooperative Patent Classification (CPC)-guided refinement process to clean messy training data. Rather than building larger or more complex models, they show that simply improving data quality yields major gains: models trained on their cleaned dataset boost accuracy by 18-21% compared with existing tools. This result underscores the paper’s core message — the real barrier to accurately identifying AI-related patents isn’t model size but the quality of the data those models rely on.

This improved precision has direct implications for how companies assess innovation and competitive strategy. As Ding explains, “With better precision in identifying AI patents, our novel AI patent data offer the research community a better data tool for understanding firms’ use and development of AI‑related innovations.”

It also matters for investors who rely on patent signals to evaluate emerging technologies. Ding notes, “As our data provide better precision in differentiating true versus pseudo‑AI patent innovations, investors — including VCs, PE investors, and equity investors — who are making financial investments based on firms’ innovation in the AI realm should have more accurate information for their investment decisions.”

New Tool for Tracking AI Innovation

The study also provides the broader research community with a valuable public resource. The authors have released AIPat, a free, open-access dataset that identifies which patents are genuinely related to AI, along with the underlying code and training data. Updated annually, AIPat offers a highly accurate reference point, making it easier for nontechnical researchers, smaller universities, and policy groups to study real AI-innovation trends. By making these tools openly available, the team hopes to support more trustworthy research and strengthen the wider network of scholars working in AI.

Beyond improving how AI-related patents are identified, the authors note that their approach can be applied to other areas where patent sorting is notoriously difficult — such as clean energy, biotech, and advanced materials. Because patent language is often dense and legalistic, it can be hard to tell which inventions truly belong to a specific field. Their CPC-guided refinement method offers a practical way to improve accuracy across many technology areas.

Although the techniques they use — such as PU learning, reducing label noise and modern text-analysis models — are already common in machine learning, the breakthrough lies in applying them to fix a widely relied-upon but flawed standard. In that sense, the study delivers a major infrastructure upgrade for researchers and policymakers who rely on patent data to understand real-world innovation.

Read the study, “Data and Methods for Identifying Artificial Intelligence-related Patents,” published in the November 2026 issue of Research Policy.