Project on Quality Assessment for AI-based Patent Drafting
About the Project
The expanding role of AI in patent drafting, prior-art search, and patent assessment has accelerated the need for an objective, evidence-grounded framework for evaluating patent quality. Patent offices face mounting pressure on examination capacity, and courts must apply doctrinal standards (inventive step, sufficiency of disclosure, the person skilled in the art) to filings that are increasingly AI-assisted. IP-intensive organizations require validated instruments for benchmarking portfolios, and developers of patent-assessment software require a defensible foundation against which their tools can be evaluated. Policy actors, in turn, require an empirical basis for assessing how the patent system is evolving. The empirical patent-quality literature offers relevant components, but no shared, application-ready framework currently exists.
Foundation in prior research
This methodological gap was identified in the researchers’ recent working paper, AI, Innovation and the Future of the Patent System (Thumm & Glänte, 2026), based on over 25 expert interviews across the United States and Europe. The study examined how AI is reshaping the patent system from the inside and identified the question of whether AI-assisted patents meet established quality thresholds as a central open question for further research. The proposed pilot is a focused empirical follow-on to that work.
Approach
Step 1: Theoretical framework. The first step develops, from established sources, a baseline definition of what constitutes a good patent and a corresponding framework for measuring patent quality. The framework draws on the empirical patent-quality literature, doctrinal materials on patentability criteria, examination guidelines at the EPO and USPTO, and the qualitative findings of the prior working paper. It is presented as an evidence-grounded working hypothesis, open to refinement in subsequent steps and through wider stakeholder engagement.
Step 2: Empirical application. The framework developed in Step 1 is then applied to a curated patent dataset including AI-generated material and associated quality assessments, provided by an external stakeholder. This step produces initial empirical results under the framework and identifies refinements that purely theoretical work cannot anticipate.
Step 3: Validation methodology. The third step develops the methodology for a subsequent expert-evaluation phase. This includes the design of a Turing-test protocol for distinguishing AI- and humandrafted patents, the identification of relevant evaluator stakeholder categories and the calibration of evaluation questions appropriate to each. The methodology is documented in a form that allows a subsequent empirical phase to be undertaken at appropriate scale and with appropriate resources. Execution of the empirical phase lies outside the scope of this pilot.
Outputs
The principal output is a working paper presenting the framework, the empirical results from Step 2 and the validation methodology developed in Step 3, accompanied by a stakeholder brief on the doctrinal and policy implications. The deliverable is positioned as a focused interim contribution that opens the path to larger-scale empirical work.
Project Team
Dr. Bowman Heiden, Project Lead
Dr. Nikolaus Thumm, Research Fellow
Gabriel Glänte, Research Fellow
The Berkeley Policy Institute (formerly BRG Institute) Project on Quality Assessment for AI-based Patent Drafting is supported in part by Tradespace. Berkeley Policy Institute maintains complete editorial independence.
Updated: August 5, 2026