AI inventorship: what the USPTO’s rules mean for R&D teams

Under current U.S. law, AI inventorship has a short answer: an AI system cannot be an inventor, but an invention made with AI can still be patented if a human being conceived what is claimed. The USPTO now treats AI like any other research tool, such as a mass spectrometer or a simulation package. The question is no longer how much the AI did. It is which people formed the specific, settled idea of the invention.

This article is for R&D leaders and in-house teams whose scientists use generative models, design software or machine-learning pipelines in daily work. I cover the case law, what happened to the USPTO’s 2024 guidance, how the current test works, how to document human contribution in lab notebooks and invention disclosures, what the duty of disclosure requires, and how Europe and the UK compare.

Key takeaways

• Only natural persons can be inventors on a U.S. patent. The Federal Circuit held this in Thaler v. Vidal (2022), and the Supreme Court declined review in 2023.
• The USPTO rescinded its February 2024 AI inventorship guidance in November 2025. The replacement says the same inventorship standard applies whether or not AI was used.
• Conception by a human is the test. AI is treated as a tool, like lab equipment or software.
• The multi-factor Pannu test now applies only to sort out joint inventorship among humans, not to weigh a human against an AI.
• Your best protection is a records habit: capture what each person decided, specified, selected and changed, not just what the model produced.
• The EPO and the UK Supreme Court also require a human inventor, and the EPO has rejected inventor designations that undercut themselves by crediting the machine.

Can AI be listed as an inventor on a U.S. patent?

No. The Patent Act defines an inventor as the “individual” or “individuals” who invented the subject matter, and the Federal Circuit held in 2022 that “individual” means a natural person. An application that names an AI system as inventor is incomplete. The Supreme Court denied review in April 2023, so this is settled law today.

The case was Thaler v. Vidal, No. 2021-2347 (Fed. Cir. Aug. 5, 2022). Stephen Thaler filed two applications, one for a “neural flame” device and one for a “fractal container,” and named only his AI system, DABUS, as inventor. The USPTO refused to process them. The district court and then the Federal Circuit agreed with the agency.

The court’s reasoning was statutory. The definition of “inventor” in 35 U.S.C. § 100(f) uses the word “individual,” and the court read that word in its ordinary sense: a human being. The court found the statutory language unambiguous, so the case turned on the words Congress used.

Just as important is what the court did not decide. It said it was “not confronted today with the question of whether inventions made by human beings with the assistance of AI” are patentable. That open question is the one most R&D teams actually face. Nobody in your lab is trying to name a model as inventor. They want to know whether using one puts the patent at risk.

What happened to the USPTO’s February 2024 AI inventorship guidance?

The USPTO withdrew it. In February 2024 the Office published guidance that applied the Pannu “significant contribution” factors to every human working with AI. On November 28, 2025, it rescinded that guidance in its entirety and replaced it with a shorter notice: one inventorship standard for all inventions, with AI treated as a tool.

The 2024 approach

The February 13, 2024 notice (89 FR 10043) asked whether each named person made a “significant contribution” to the invention, using the three factors from the Federal Circuit’s Pannu v. Iolab decision. Under those factors, each inventor must (1) contribute in some significant manner to conception or reduction to practice, (2) make a contribution that is not insignificant in quality measured against the full invention, and (3) do more than explain well-known concepts or the current state of the art.

The 2024 notice added five guiding principles. Among them: recognizing a problem is not conception, a carefully built prompt might be a significant contribution, and merely owning or overseeing an AI system is not enough. In practice, the framework pushed teams to ask whether the humans had done “enough” relative to the machine.

The November 2025 replacement

The Revised Inventorship Guidance for AI-Assisted Inventions (90 FR 54636, Nov. 28, 2025) changes the frame. It states that the same legal standard governs every invention. It describes AI systems, including generative AI, as “instruments used by human inventors,” comparable to laboratory equipment, software or research databases. The USPTO said the change implements Executive Order 14179 on AI policy.

As of this writing in October 2026, I have not found a later USPTO notice replacing the November 2025 guidance. One practical warning: the USPTO’s older AI inventorship FAQ page still reflects the 2024 guidance. If someone on your team pulls it up, make sure they know it predates the current notice.

How does the current USPTO test for AI inventorship work?

The test is conception, which the Federal Circuit calls the “touchstone of inventorship.” Conception is complete when a person has a specific, settled idea of a particular solution, not just a goal or a research plan. If one scientist conceived the claimed invention using AI as a tool, that scientist is the inventor. If several did, ordinary joint inventorship rules decide who is named.

A few details in the 2025 notice matter for R&D teams.

• The USPTO presumes the named people are the inventors. Examiners generally accept the inventors listed on the application data sheet or oath. Inventorship fights usually surface later, in litigation, due diligence or an ownership dispute.
• Naming an AI triggers rejection. An application that names an AI system as inventor or joint inventor should be rejected under 35 U.S.C. §§ 101 and 115 for all claims.
• Foreign priority needs a shared human inventor. A U.S. application can’t claim priority to a foreign application that names an AI tool as its sole inventor. If a foreign filing lists both a person and a machine, the U.S. filing should list only the person or people.
• Design and plant patents follow the same logic. The notice applies the conception standard to design patents and the existing plant patent rules to plant varieties developed with AI.

What conception looks like when AI is in the loop

“AI as a tool” does not mean anyone who touched the tool is an inventor. It also doesn’t mean the output belongs to whoever pressed “run.” The question is who held the specific idea that ended up in the claims. Two examples show how that plays out.

Small molecules. A medicinal chemist uses a generative model to propose kinase inhibitor scaffolds. The model returns 4,000 candidates. She defines the binding constraints, rejects most of the output on her own chemical judgment, modifies one scaffold by adding a fluorine at a position the model never proposed to block a metabolic soft spot, and specifies the synthesis route. Her claimed compound and its rationale came from her decisions. The model narrowed her search, much as a docking program or a compound library would.

Now change the facts. A project manager types “find a better inhibitor,” picks the top-ranked structure, and sends it for synthesis with no changes. A general goal is not conception. That is the scenario where inventorship becomes hard to defend, and it is also the scenario where your records matter most.

Photonics. An optical engineer runs an inverse-design algorithm to optimize a silicon waveguide bend. If she chose the figure of merit, constrained the fabrication rules, and then recognized and kept a non-obvious sub-wavelength grating feature the optimizer produced, the record should show each of those choices. If the claim covers that feature, the file should explain why she understood its function, not just that it appeared in the output.

One nuance a chemist will recognize. In unpredictable arts like chemistry and biology, the law has long allowed that conception and reduction to practice can happen together, because you may not have a settled idea until the experiment works. The USPTO acknowledged this doctrine in its 2024 notice. When a model proposes and a human validates, your notebook should capture the moment the human understood what worked and why. That is where the conception argument often lives. If you work in these fields, my notes on patent work for computational chemistry and AI drug discovery and on protecting small molecules and new chemical forms go further.

How should R&D teams document human contribution?

Record decisions, not just outputs. For AI-assisted work, a good record shows the problem each person framed, the constraints and inputs they chose, what the model returned, what the humans selected, rejected and changed, and the reasoning behind each step. Dated, attributed entries made at the time carry far more weight than a reconstruction written when the patent is filed.

A lab notebook and invention disclosure checklist

1. Name the problem and the person. Write down who defined the technical problem and the success criteria, such as target potency, loss budget or thermal stability.
2. Log the tool and version. Record the model or software, its version, and whether it was trained or fine-tuned in-house, and by whom. Custom training choices can be part of the human story.
3. Save the inputs. Keep prompts, constraint files, training-data selections and parameter settings with dates. Treat them like experimental conditions.
4. Save the raw outputs. Archive what the model actually produced, so it is clear later which features came from the model and which came from people.
5. Record selection and rejection reasons. Note why a scientist kept candidate 37 and discarded the rest. Selection based on technical judgment is evidence of a settled idea.
6. Record every human modification. Document changes to structures, geometries, sequences or process steps, and who made them.
7. Record validation and the “why.” When an experiment confirms the result, write down what the team now understands about how it works.
8. Attribute each claim element. In the invention disclosure form, ask contributors to map their input to specific features of the invention. Your patent counsel will later map those features to the claims.
9. Keep it contemporaneous and witnessed. Use dated electronic notebook entries with audit trails, and have a colleague review key entries.

Update your invention disclosure form to add a short AI section: which tools were used, for which steps, and what the humans contributed at each one. That takes ten minutes per disclosure and saves hours when an investor’s counsel asks about it. Inventorship and chain of title are standard diligence items; I cover them in what investors ask about your patents.

Also check your tool terms. Some AI service agreements address ownership of outputs and use of your inputs. That is a contract question, separate from inventorship, but it belongs in the same review, especially for confidential compound data or unpublished device designs.

Do you have to disclose AI use to the USPTO?

There is no USPTO rule requiring a general statement that AI was used. But the duty of candor still applies. Under 37 CFR 1.56, everyone substantively involved in an application must disclose information they know is material to patentability, and must not mislead the Office. Facts about who actually conceived the invention can be material.

The rule is 37 CFR § 1.56, often called Rule 56. It applies to each named inventor, each attorney or agent who prepares or prosecutes the application, and others substantively involved who are associated with the inventor or assignee. The November 2025 notice does not create a new AI disclosure duty. It does not mention Rule 56 at all.

In practice, three points matter:

• Don’t name the wrong people. Naming someone who did not contribute to conception, or leaving off someone who did, is an inventorship error. Honest errors can be corrected under 35 U.S.C. § 256, but a deliberate misstatement is a different problem.
• Watch your prior art. If a model output, an internal report or a public dataset is material prior art, the source being an AI does not change your disclosure obligations.
• Keep the specification accurate. If the application describes how the invention was made, describe it truthfully. Don’t overstate or understate the human role.

Timing also matters. AI tools make it easy to generate and share many candidates quickly, and an early public disclosure can start deadlines. For spin-outs and early-stage companies, I discuss filing timing in patent strategy for startups and university spin-outs.

How do Europe and the UK treat AI inventorship?

Both require a human inventor. The EPO’s Legal Board of Appeal held in J 8/20 that an inventor designated on a European application must be a person with legal capacity, and the UK Supreme Court held in December 2023 that an inventor under the Patents Act 1977 must be a natural person. Neither ruling bars patents on inventions made with AI assistance.

At the EPO, the DABUS applications were refused, and the EPO announced in its July 2022 notice on decision J 8/20 that “an inventor designated in a patent application must be a human being.” A newer decision adds a lesson for formal paperwork. In T 528/25 (February 5, 2026), a Board accepted that designating an inventor is in principle possible for inventions developed using AI. But it refused designations naming Dr. Thaler alongside statements that DABUS conceived the invention, because those statements made the designation unclear.

In the UK, the Supreme Court decided Thaler v Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49 on December 20, 2023. It said the appeal did not concern the broader question of whether machine-generated technical advances should be patentable.

For a global filing program, the lesson is consistency. Name the same humans for the same contribution everywhere, and keep the U.S. application data sheet, the PCT request and any European designation aligned. A foreign filing that credits a machine can create problems for a U.S. priority claim.

Frequently asked questions

Can an invention made with AI be patented?

Yes, if at least one natural person conceived the claimed invention. The USPTO’s November 2025 guidance treats AI as a tool, similar to lab equipment or software, and applies the same inventorship standard as for any other invention. The patent must still meet every other requirement, including novelty, non-obviousness, written description and enablement.

Is the USPTO’s February 2024 AI inventorship guidance still in effect?

No. The USPTO rescinded the February 2024 guidance in its entirety in a notice published November 28, 2025. The replacement drops the approach of applying the Pannu significant-contribution factors to every human using AI. Pannu now matters only for deciding joint inventorship among several people.

Does writing a prompt make someone an inventor?

Not by itself. Under the current guidance, the question is whether the person conceived the specific invention that is claimed. A generic prompt stating a goal is not conception. A prompt that encodes a specific technical solution, followed by informed selection and modification of the output, can be strong evidence of conception. Records of those steps help.

Do I need to tell the USPTO that I used AI?

There is no general rule requiring an AI-use statement. The duty of candor under 37 CFR 1.56 still applies, so you must disclose known material information and must not misstate inventorship. The November 2025 guidance creates no new AI-specific disclosure duty. Your patent counsel should know how AI was used so the inventors are named correctly.

What if we named the wrong inventors on an AI-assisted patent?

Inventorship errors can often be fixed. For issued patents, 35 U.S.C. § 256 lets the USPTO, or a court, correct the named inventors, and a correctable error does not by itself invalidate the patent. Fixing it early, while the records are fresh and the people are available, is easier than fixing it during litigation or a financing.

Talk it through

If your team is building AI into its research workflow and wants a disclosure form and notebook practice that will hold up, I’m happy to look at what you have.

This article is general information, not legal advice, and reading it does not create an attorney-client relationship. Patent law changes and every situation differs. Please talk to a patent attorney about yours.

Message Alexander on LinkedIn