AI-assisted patent drafting and inventorship: where the law stands

The first question is whether an artificial intelligence can be named as an inventor. The second is whether an AI may be used to write the application. The law answers the first with a near-universal 'no' and leaves the second almost entirely to the applicant's judgement — with the applicant carrying the consequences.

Between 2018 and 2024 a series of test cases sought to name an AI system, DABUS, as inventor. The answers were consistent. The EPO's Legal Board of Appeal held in J 8/20 that an inventor within the meaning of the European Patent Convention must be a person with legal capacity, and that a machine cannot be designated. The United States Court of Appeals for the Federal Circuit reached the same result under the Patent Act in Thaler v. Vidal (2022), and the UK Supreme Court in Thaler v Comptroller-General (2023) confirmed that an inventor must be a natural person under the Patents Act 1977.

French law reaches the same point from Article L611-6 CPI: the right to the patent belongs to the inventor or their successor in title, and the inventor is identified as a person. There is no procedure at INPI for designating a machine.

When a human uses an AI tool and something patentable results, the question becomes which humans made an inventive contribution. The USPTO's inventorship guidance for AI-assisted inventions, first issued in February 2024, applies the ordinary test: each named inventor must have contributed significantly to the conception of the invention. Merely presenting a problem to a system, or recognising and appreciating its output, is not enough; designing the experiment, selecting among outputs on the basis of technical insight, or building the system specifically to solve the problem can be.

European practice does not have a separate AI-inventorship test, but the same logic applies through the concept of the inventor as the person who conceived the invention. The practical rule for a research team is to document, contemporaneously, who decided what — which prompts, which selections, which experiments — so that inventorship can be established and, if challenged, proved.

No patent office prohibits the use of software, including generative AI, in preparing an application. What the offices regulate is the result and the conduct of the person who signs. Three duties do not move.

• Accuracy: every statement in the specification and every cited reference is the applicant's own. The USPTO's February 2024 guidance on the use of AI tools reminds practitioners that the duty of candor and the certification under 37 C.F.R. § 11.18 apply to AI-generated content; courts have sanctioned lawyers for filing fabricated citations produced by language models.
• Sufficiency and plausibility: an AI-written example that was never run is a prophetic example. It may be acceptable in the United States if not written in the past tense, but before the EPO an effect must be at least encompassed and embodied by the application as filed (G 2/21) and a purely invented data set will not carry an inventive-step argument.
• Confidentiality: sending an unfiled invention to an AI service whose terms allow training on user content is a disclosure risk and, for employee inventions, a possible breach of the employer's confidentiality rules. Read the processing terms before the first prompt.

Language models are strong at the mechanical and structural parts of drafting: turning a technical brief into a well-formed field, summary and detailed description; enumerating alternatives and ranges systematically; producing claim ladders from a broad independent claim down to narrow fallbacks; converting a claim set between office formats — the EPO's two-part form under Rule 43(1) EPC, USPTO formalities, PCT layout; and generating first-pass prior-art maps from claim language.

They are weak where the work is judgement under uncertainty: deciding how broad the independent claim should be in view of a specific prior-art landscape, choosing which technical effect to anchor the inventive step on, and knowing which example must be run before filing rather than described. Those decisions should be made by a person, ideally with the AI's alternatives on the table.

Whoever the vendor, an AI drafting tool used for real filings should offer, at a minimum: no training on your content, with the processor's terms in writing; verbatim quotation of every prior-art passage it relies on, so that a hallucinated citation cannot survive a check; multiple independent generations with a vote or a review rather than a single sample; an explicit human sign-off step before any document leaves for an office; and an export in the office's native format so that counsel can review the exact text that will be filed.

The text produced by a drafting tool is a work product, not an invention; copyright in AI-generated text is unsettled in most jurisdictions, which is one more reason the contract with your provider should assign whatever rights exist in the output to you and should say that the provider claims no inventorship or ownership in the subject-matter. IPZilla's terms do exactly this; read the equivalent clause in any tool you use.