Study finds AI patents often struggle with the "abstract idea" test, not inventiveness.
Associate Professor Amy Semet (University at Buffalo School of Law) published the first large-scale empirical study linking the USPTO's AI Patent Dataset to district-court outcomes, covering nearly 80,000 patents litigated between 2000 and 2025. Among cases decided on the merits, AI patents were invalidated at 74.2% versus 47.9% for non-AI patents. A statistically significant AI-specific effect (roughly +5 points) survives full controls, multiple confidence thresholds, and court/year/technology fixed effects.
The USPTO classification underlying the dataset covers patents issued from 1976–2023. The litigated AI set therefore skews toward planning, hardware, and knowledge-representation technologies that pre-date the post-2022 commercialization wave of modern neural-network and foundation-model systems.
Some observations from the data:
- The invalidity gap is driven mainly by patent eligibility under §101. Before the Supreme Court's 2014 Alice decision, AI patents were not especially vulnerable under §101. After Alice, eligibility accounted for 53.5% of AI invalidations versus 42.0% for non-AI. AI patents are also resolved on the pleadings far more often (37.1% vs. 9.5% for invalidity determinations).
- AI patents are significantly less likely to be invalidated for obviousness under §103. As Prof. Semet explains, patents screened out early as ineligible never reach the fact-intensive Graham inquiry. Many AI inventions' technical inventiveness is thus never fully tested in litigation.
- AI patents also show lower infringement success (24.0% found infringed versus 34.7% for non-AI). The invalidity penalty persists even within software patents, while the infringement penalty concentrates in non-software fields where distributed or opaque systems are harder to observe and attribute.
One encouraging note: when AI patents do reach the prior-art inquiry, they tend to fare better than non-AI patents. The primary barrier appears to be eligibility characterization rather than lack of inventiveness.
The USPTO has signaled movement in this space. The precedential Ex parte Desjardins decision (2025) recognized certain machine-learning improvements as eligible technical advances. Whether litigation patterns shift for the next wave of generative-AI patents — many drafted under more recent guidance — remains an open question for courts and Congress.
Read the full paper on LinkedIn.