AI helped invent it: who counts as the inventor

Your engineers used an AI tool somewhere in developing the thing you are about to file on. Whose names go on the ai patent application?

The literal question people search, which is whether an AI can be the inventor, has a settled answer and it takes one line. The question that actually decides your filing is different and the rules on it moved in late 2025, which means a good deal of what is published about it is now wrong. Both answers are below, with the primary documents linked.

Can an AI be named as an inventor?

No, and the reason is statutory rather than philosophical. The United States Court of Appeals for the Federal Circuit held on 5 August 2022, in the Thaler litigation, that only a natural person can be an inventor. The current examination guidance repeats the consequence in operational terms: a rejection under 35 U.S.C. 101 and 115, or other appropriate action, should be made for all claims in any application that lists an AI system or other non natural person as an inventor or joint inventor.

There is a related trap for anyone filing internationally. A priority claim to a foreign application that names an AI tool as the sole inventor will not be accepted, so a filing strategy that routes through a jurisdiction with different rules does not import the result.

What happened when the same applications went around the world?

They were refused almost everywhere and granted once, which is the fact most people half remember and place incorrectly. Stephen Thaler filed applications naming his system DABUS as inventor in multiple jurisdictions, and the documented record of those proceedings now covers seven outcomes.

The South African grant is real and it is also the exception that proves the rule, since it came through a filing route whose examination does not test the point the other offices tested. Treat it as a curiosity in a table rather than a strategy.

What changed in November 2025, and why does it matter to you?

The framework that told you how to assess your own engineers' contribution was withdrawn and replaced. On 28 November 2025 the Patent and Trademark Office published Revised Inventorship Guidance for AI Assisted Inventions, at 90 FR 54636, and the revised guidance states that the February 13, 2024 version is rescinded in its entirety.

What went with it was the Pannu analysis. The 2024 guidance had applied the Pannu joint inventorship factors to AI assisted inventions, treating the human and the tool rather as collaborators to be weighed. The revised notice withdraws that approach and confines those factors to their original job: Pannu applies only when determining whether multiple natural persons qualify as joint inventors, and does not apply where one natural person is involved.

In its place, one standard for everything. The same legal standard for determining inventorship applies to all inventions, with no separate or modified standard for AI assisted ones, a position the office also states plainly on its own artificial intelligence resources page, which now lists the November 2025 guidance as current and marks the February 2024 version as archived and under review.

So what is the actual test now?

Whether a named human conceived the invention, under the ordinary conception standard that predates all of this. The notice is specific about what conception requires: a specific, settled idea, a particular solution to the problem at hand, not just a general goal or research plan.

The tool framing is explicit. AI systems may provide services and generate ideas, and they remain tools used by the human inventor who conceived the claimed invention. That sentence does a lot of work. It means the existence of an AI in your development process is not itself an inventorship event, in the way that using a microscope or a simulation package never was. It also means the burden sits where it always sat, on showing a human arrived at the specific solution.

Two points from the rescinded guidance are worth knowing even though it is withdrawn, because they describe the shape of the problem and nothing in the revision contradicts them. Merely recognising a problem, or having a general goal or research plan, does not rise to the level of conception. And maintaining intellectual domination over an AI system does not on its own make a person an inventor, so owning the model or paying for the licence earns nobody a place on the application.

One thing has changed in emphasis and you should know the difference. The 2024 guidance stated that a significant contribution could be shown by the way a person constructs a prompt, where the prompt was written to elicit a particular solution to a specific problem. The November 2025 notice does not address prompts at all. It sends the question back to conception as a fact intensive inquiry. So anyone telling you that clever prompting is a recognised route to inventorship is quoting a withdrawn document.

Who goes on the application when several people and a tool were involved?

The humans who contributed to conception, assessed among themselves, with the tool dropping out of the analysis entirely. This is the part the November 2025 change actually simplifies, and it is worth walking through because it is the common real situation.

The Pannu factors did not disappear. They were returned to their proper scope: deciding whether multiple natural persons qualify as joint inventors. So if three engineers worked on a problem and one of them used an AI tool heavily, you run the ordinary joint inventorship analysis across the three people. The tool is not a fourth candidate to be weighed, and the fact that one person used it is not by itself a reason to include or exclude them.

Take the concrete case. An engineer frames the problem and writes the prompts. A second engineer recognises that one of the generated candidates solves a failure mode the team had been stuck on, and specifies the modification that makes it work. A third runs the validation. Under the current standard the question for each is whether they contributed to conceiving the specific, settled solution that the claims cover. The second engineer's recognition and specification looks like conception. Running validation to somebody else's specification generally does not, and writing prompts toward a general goal, without arriving at the particular solution, generally does not either.

Note what that means for how teams work. The decision turns on who arrived at the specific solution, which is rarely the same question as who did the most work or who ran the tool. That is also why the records described below are about decisions rather than about activity logs.

What should you record, and when?

The specific idea, the choices you made, and the judgement you applied, all dated as it happens. Reconstructing conception afterwards is exactly what an opponent attacks, and a document created the week before litigation reads like one.

The settled idea, with a date. The moment somebody can state the particular solution rather than the goal. One paragraph, dated, is worth more than a folder of exports. This is the element the standard actually asks about.

What was asked of the tool and why it was framed that way. Keep the prompts, and keep the reasoning behind them. Even though the current notice does not treat prompt construction as its own route to inventorship, the reasoning behind a prompt is evidence about what the human understood and intended, which is what conception turns on.

Model and data selection. Which system, which version, which training or input data, and why those were chosen for this problem. Version matters practically: a model that answered differently last March is a fact you may need to establish.

How outputs were evaluated and refined. What was rejected, on what grounds, and what was changed. This is often where the human contribution genuinely lives, and it is the part least likely to be captured anywhere unless somebody writes it down.

Mechanically, the cheapest version of all this is a dated running note per project, kept by whoever is doing the work, in a system with version history. We argued the same point about provenance for creative assets in what sellers control over AI training and their own content: records made at the time beat reconstructions made later, and that principle does not care which branch of intellectual property you are in.

Do you have to tell the office you used AI?

Inventorship and disclosure are two different questions, and the November 2025 notice is about the first. It applies the same inventorship standard whether or not AI systems were used and creates no separate standard for an AI assisted invention, which is the opposite of a special AI reporting rule for inventorship purposes.

The use of AI tools in dealings with the office is handled separately. The Patent and Trademark Office published distinct guidance on the use of artificial intelligence based tools in practice before it, at 89 FR 25609 in April 2024, which addresses conduct in proceedings rather than who counts as an inventor. If your attorney is drafting with AI assistance, that is the document that governs the question, not the inventorship guidance.

For a company the practical reading is this. You are not required to volunteer a narrative about which tools your engineers used in order to establish inventorship. You are required to name the right humans, and to be able to support that if challenged. Those two facts together are why the recording habit matters more than any disclosure decision: nobody asks you for the record at filing time, and everybody asks for it later.

What is the risk if you get the names wrong?

Worse than a refusal, because it surfaces late and it can take the whole patent with it. A rejection during examination is a letter you respond to. An inventorship defect found years later, by someone with a commercial reason to look, is an attack on the enforceability of a patent you have been relying on, licensing, or asserting.

That asymmetry is why this is a filing discipline question rather than a legal theory question. The upside of getting it right is nothing visible. The downside of getting it wrong arrives at the worst possible moment, usually in the middle of a dispute you started.

Two practical consequences for a small company. First, do not add names for diplomatic reasons, and do not drop them for simplicity. Inventorship is a question of fact about conception, not an acknowledgement of effort or seniority. Second, when work passes through a contractor or an agency, settle in the contract who records what and who gets named, before the work begins. The same early paperwork discipline we recommended for commissioned creative work in who owns the work when a freelancer used AI applies here with much higher stakes attached.

Does any of this apply if you are not filing patents?

Parts of it, and the parts that transfer are the ones about records. Most small businesses never file a patent. Many of them do make something distinctive with an AI tool in the process, and then need to establish later what they made, when, and how.

The intellectual property regimes answer differently and the question of who did what is common to all of them. Copyright turns on human authorship, trademarks turn on use in commerce and distinctiveness, and patents turn on conception. We covered the first two for ordinary commercial situations in what you can protect when AI made your logo and in running a trademark search before you print. Being able to say what a human decided, and when, is the common requirement.

If you want a short version of the discipline for a business with no legal department: write down what you decided and when, say publicly how you use AI, and keep the two consistent. Our own AI policy is that second part made concrete. None of this requires a lawyer until something is worth protecting, and by then the records either exist or they do not.

What to watch next

Three things, all of them checkable against primary sources rather than commentary. The archived 2024 guidance is marked under review on the office's own resources page, so further movement is possible and the examples attached to it may be reissued. Examination practice under a fact intensive conception standard will generate decisions that tell you more than the notice does about where the line sits. And the jurisdictional picture keeps adding rows, with Switzerland the most recent in June 2025.

What is stable is the part worth building habits around. Only humans can be inventors, the standard is conception, and conception is proved with contemporaneous records. That has survived two rounds of guidance and a decade of litigation, and it is the one piece of this subject that has never been in doubt.