SSRN: An Empirical Study of Artificial Intelligent Patent Litigation by Amy Semet
Artificial intelligence (AI) is now at the heart of modern innovation, but we know surprisingly little about how AI patents perform when they are challenged in U.S. district courts. This Article offers the first large-scale empirical study that connects the U.S. Patent and Trademark Office’s (USPTO) AI patent database with an originally constructed database of district court outcomes. Using several different methods to identify AI inventions, the study builds tiered measures that classify AI inventions at increasingly levels of confidence. It then examines how these patents fare in court on issues such as invalidity and infringement within the subset of cases decided on the merits.
The results show a clear and consistent pattern. Conditional on reaching a contested merits outcome, AI-related patents are significantly more likely to be invalidated than similar non-AI patents, a result that holds across multiple ways of defining AI as well as numerous control variables. When broken down by legal doctrine, the higher invalidation rate is driven mainly by section 101 subject-matter eligibility. Further, AI inventions are less likely to a statistically significant degree to be found obvious under by section 103. At the same time, AI patents are less likely to yield findings of infringement conditional on a merits-only sample of cases, suggesting they face a double hurdle in litigation: AI patents are easier to invalidate and harder for the patentee to hold an alleged infringer accountable for any infringement.