On the question of whether these patents are valid, AI patents are more likely to lose than comparable non-AI patents. The losses are driven disproportionately by subject matter eligibility rules rather than by the usual novelty, obviousness, or disclosure tests. Put another way, when courts invalidate AI patents, they are usually saying the invention is the wrong kind of subject matter, not that it is already known, obvious, or inadequately described.

Semet’s study finds that pattern holding across multiple AI definitions, case factors, and both software and non-software patents. Thus, it cannot be dismissed as a software problem.

On infringement, the picture is different but equally troubling. AI patents that reach a decision on the merits are found infringed in far fewer cases than non-AI patents, especially in areas such as language processing and machine learning. In these fields, AI systems are opaque and spread across different actors, making it hard to show that an accused system uses the patented features or to pin down a direct infringer.

Roughly 15% of patents litigated in U.S. district courts between 2000 and 2025 involve AI under the USPTO’s AI classification system, and nearly a third of recent cases include at least one AI patent. The litigation is concentrated in a handful of venues, so a small cadre of judges is effectively writing the law of AI patents, while AI patents are attacked at the Patent Trial and Appeal Board far more often than non-AI patents. That is a built-in disadvantage in how our patent system treats AI.

All of this plays out against a global backdrop that should sharpen Washington’s focus. The Stanford AI Index reports that China now grants far more AI patents per year than the U.S. and holds a dominant share of global AI patent volume, even as American models retain a narrow performance edge and U.S. firms far outspend China on private investment. It is hard to square talk of American AI leadership with a patent system that invalidates AI inventions at elevated rates and makes them harder to enforce at home.

Section 101’s court-created exceptions for “abstract ideas, laws of nature, and natural phenomena,” developed in cases like Alice and Mayo, now serve to invalidate AI patents at the pleading stage, before courts reach claim construction or discovery. Congress wrote sections 102, 103, and 112 to judge inventiveness and disclosure, but those provisions are pushed to the sidelines, and AI inventions are labeled “abstract” instead of being tested on whether they represent a real advance.

TRUMP’S AI EXEMPTION ISN’T AN OVERSIGHT GAP. IT’S A NATIONAL SECURITY MASTERSTROKE

Fortunately, the institutions that can fix this problem have AI on their agenda. The Federal Circuit can clarify how eligibility applies to machine-learning claims, separating disembodied math from applied systems with technological results. Congress can move the Patent Eligibility Restoration Act to rein in section 101’s judicially created exceptions and direct judges toward the statutory tests that ask whether an invention is new and properly disclosed. The USPTO, which has already issued AI-specific eligibility guidance under Executive Order 14110, can tighten AI disclosure standards so that overbroad claims are screened at the USPTO instead of left to district courts.

While these steps cannot guarantee victory, they put AI inventions on the same footing as other technologies, judged on their technical merits. If Washington leaves 101 jurisprudence in its current legal limbo while China races ahead in AI patenting, it will be our own courts that undermine America’s AI dominance. The White House, Congress, and the USPTO must move now, before our patent system buries the very innovation edge it is supposed to protect.

James Edwards is executive director of Conservatives for Property Rights.