USPTO Sharpens the Line on AI Prior Art

The United States Patent and Trademark Office has issued supplemental guidance on generative AI and patent examination, bringing two practical questions into the same frame: how examiners should treat material that appears to have been generated by AI when it is cited as prior art, and how applicants should explain human contribution in inventions developed with AI assistance.

The message is not that AI-generated material is automatically irrelevant. Nor is it that anything found online is reliable enough to defeat novelty or support an obviousness rejection. The centre of gravity moves back to evidence: whether the disclosure was publicly accessible, whether the technical teaching is enabling, whether an apparent combination is more than synthetic text assembly, and whether the named human inventors can explain their contribution to the claimed invention.

AI-generated disclosures are not just another web page

Generative AI does not merely increase the volume of possible prior art. It changes the reliability problem. A conventional article, patent publication or product manual usually raises familiar questions about date, accessibility and correspondence with claim features. AI-generated material can look complete while containing fabricated parameters, invented citations, inconsistent embodiments or technical combinations that were assembled by language rather than engineering logic.

The guidance brings the issue back to patent evidence. Examiners may still rely on publicly available material. But where an applicant identifies signs of hallucination, non-enabling disclosure or data stitching, the question should not stop at whether the page was searchable. For novelty, the issue is whether a single reference clearly and completely discloses every claim limitation. For obviousness, the issue is whether a skilled person would have had a reasoned basis to combine credible technical teachings, not whether a machine-generated paragraph happens to place several features side by side.

Rejections will need a more verifiable footing

United States examination practice often gives printed publications a degree of practical reliability, but that reliability can be challenged. The guidance matters because it pushes examiners to articulate why a suspect AI-generated source remains usable as prior art, which passages are being relied on, and what factual basis supports an obviousness combination.

Applicants should respond with the same discipline. A bare assertion that a source was “AI-generated” is unlikely to carry much weight. A stronger response points to technical defects: a chemical structure inconsistent with the stated property, an impossible mechanical assembly, a missing algorithmic input, experimental data that fails basic dimensional checks, citations that cannot be verified, or embodiments within the same page that contradict each other. The more the objection is tied to technical facts, the harder it is for the rejection to rest on the source without further explanation.

For AI-assisted inventions, the focus remains human conception

The second strand concerns inventions developed with AI tools. USPTO’s approach does not treat AI use as a defect in itself. The key point is that inventorship remains anchored in human conception. An AI system may serve as a research tool, modelling environment, drafting aid or source of candidate solutions. The inventor, however, must be a natural person who made a concrete contribution to the conception of the claimed invention.

That should change how applicants prepare records. It is no longer enough for an internal file to say that a research team “used an AI tool”. A better record maps human contribution to claim features: who framed the technical problem, who set model inputs and constraints, who selected and modified outputs, who verified feasibility through experiments or engineering tests, and who made the substantive choices reflected in the final claims. The number of prompts is less important than the ability to show that human inventors possessed the final technical solution.

Disclosure should become an auditable record

The guidance should not be read as requiring every use of AI to be reported automatically. A more practical reading is that AI use becomes relevant when it bears materially on patentability, inventorship or the accuracy of submissions. If a key claimed embodiment came from model output and no named inventor can explain the inventive contribution, the issue is sensitive. The same is true where AI-assisted drafting introduces unverified alternatives or unsupported technical effects into the application.

Companies can make their records clearer without making them unwieldy. A workable format has three layers: identify whether AI was used for search, solution generation, drafting support or data analysis; describe the natural person’s conception, selection, modification or verification of the core technical features; and confirm that the submission was reviewed for citations, data and technical effects. Whether such material should be submitted to the USPTO depends on the case, counsel’s judgment and disclosure obligations. Internally, prompts, key outputs, version histories, validation results and claim amendments should be preserved. AI may speed up invention work, but it also exposes record gaps. What survives examination is still provable human technical judgment.