If a human being made a significant enough contribution to conceive of the invention, you can patent it, even if an AI system did much of the work. If no human conceived it, U.S. patent law currently offers nothing. There is no separate category of protection for inventions with no human inventor, and the AI system itself cannot be named.
This is general information about U.S. patent law, not legal advice. AI inventorship rules have changed twice in two years, and the specifics of your situation matter. Speak with a registered patent attorney before relying on any of it.
That’s the short version, and it’s the version most founders using generative design tools, drug discovery models, or code-generating systems actually need. The longer version is more interesting, because the rules moved in late 2025 in a way most published content hasn’t caught up to.
Only Humans Can Be Inventors
The Federal Circuit settled the threshold question in Thaler v. Vidal in 2022, holding that the Patent Act’s use of “individual” means a natural person and that an AI system named DABUS could not be listed as an inventor. The Supreme Court declined to review that decision in 2023.
So the AI isn’t an inventor. That part’s settled. The harder question stayed open: when a human leans on AI heavily, does that human still qualify?
The Standard Changed in November 2025
In February 2024 the USPTO issued inventorship guidance for AI-assisted inventions. It confirmed they aren’t categorically unpatentable, and it directed examiners to assess whether a natural person made a “significant contribution” using the Pannu factors, a test borrowed from joint inventorship law.
On November 28, 2025 the USPTO published revised guidance that pulled that approach back. The problem was structural: the Pannu factors were designed to sort out which of several humans count as joint inventors, and applying them where only one human was involved didn’t fit. Under the revised guidance, the Pannu factors apply only when you’re deciding whether multiple natural persons qualify as joint inventors. Where a single person worked with an AI tool, the question returns to the traditional conception standard.
Conception, as courts have long defined it, means having a definite and permanent idea of the complete and operative invention, clear enough in the inventor’s mind to be described to someone else with particularity.
Both versions of the guidance agree on the bottom line: AI cannot be named as an inventor on an application or an issued patent, and naming a non-natural person draws rejections under 35 U.S.C. 101 and 35 U.S.C. 115.
If you have pending applications that were prepared against the 2024 framework, that’s worth a conversation with your attorney. The analysis your file was built on is no longer the operative one.
Eligibility Is a Separate Question
People conflate two very different issues. Inventorship asks who invented. Subject matter eligibility asks whether the thing is the kind of thing patents cover at all.
The USPTO’s July 17, 2024 eligibility guidance update was blunt on this point: whether an invention was created with AI assistance is not a consideration in the Alice/Mayo analysis. How an invention was developed is irrelevant to eligibility.
What does matter is whether the claim recites an abstract idea and, if so, whether it integrates that idea into a practical application. For AI and machine learning claims, that generally turns on whether you’re claiming a specific technical improvement (to how a computer functions, or to another technology) rather than the idea of a result. The guidance added Examples 47 through 49, covering an artificial neural network used for anomaly detection, a speech signal separation method, and a personalized medical treatment model. Those examples are the most useful drafting reference available for anyone claiming an AI system.
The practical lesson: claims that describe what the model achieves struggle. Claims that describe a particular architecture, training approach, or technical mechanism that solves an identified problem fare better.
What Protects Output the AI Generated Alone
Say your model produced something genuinely novel with no human conception behind it. What’s left?
Copyright, mostly no. The D.C. Circuit affirmed in Thaler v. Perlmutter in March 2025 that human authorship is required under the Copyright Act. The Supreme Court denied review on March 2, 2026, leaving that holding in place. Works with no human author don’t get registered.
Trade secret, often yes. Trade secret protection doesn’t require an inventor or an author. It requires that the information have value from not being generally known and that you take reasonable steps to keep it secret. For model weights, training pipelines, and generated formulations you never intend to disclose, this is frequently the stronger option, and unlike a patent it doesn’t expire on a schedule.
Contract, always worth doing. Licensing terms, NDAs, and employment agreements govern who can use what regardless of whether an underlying patent exists.
The tradeoff is an old one and it hasn’t moved. A patent buys you a time-limited right to exclude in exchange for public disclosure. A trade secret buys you indefinite protection with no right to stop independent development or reverse engineering. If your AI-generated asset can be reverse-engineered from a shipped product, secrecy is thin cover.
What to Document Starting Now
Whatever you’re building, the record you keep today decides what you can claim later.
Write down what the humans did. Not “the team used the model,” but which person framed the problem, which one specified constraints, which one recognized that a particular output was the solution and why. Conception is a factual question and it gets litigated with documents.
Keep prompts, parameter choices, and iteration history. Note who reviewed and selected the outputs, and record the date a human first recognized the invention as an invention. If your engineering culture is “the model figured it out,” your paperwork will say a machine conceived it, and that’s the one answer the Patent Act won’t accept.
Frequently Asked Questions
What legal protections exist for AI-generated inventions in patent law?
A U.S. patent is available when a natural person conceived the invention, even if AI tools contributed substantially to developing it. There is no protection under patent law for an invention with no human inventor. Where no human conceived it, trade secret protection and contract terms are the realistic alternatives, since copyright also requires human authorship.
Can an AI system be listed as an inventor on a U.S. patent?
No. Thaler v. Vidal (Fed. Cir. 2022) held that an inventor must be a natural person, and USPTO guidance confirms that naming a non-natural person draws rejections under 35 U.S.C. 101 and 115.
Does using AI make my invention unpatentable?
No. AI-assisted inventions are patentable as long as a natural person qualifies as an inventor. The USPTO’s July 2024 eligibility guidance also states that how an invention was developed is not part of the eligibility analysis.
What changed in the USPTO’s November 2025 inventorship guidance?
The revised guidance withdrew the use of the Pannu factors for cases involving a single natural person. Those factors now apply only when determining whether multiple natural persons are joint inventors. For a single human working with an AI tool, the traditional conception standard governs.
What counts as a human contribution big enough to support inventorship?
Enough to establish conception: a definite and permanent idea of the complete and operative invention, described with sufficient particularity. Simply owning the AI system, supplying funding, or asking it a general question does not get you there.
Can I copyright what my AI generated?
Not without human authorship. The D.C. Circuit affirmed the human authorship requirement in Thaler v. Perlmutter in March 2025, and the Supreme Court denied certiorari on March 2, 2026.
Should I file a patent or keep it as a trade secret?
It depends on whether the asset can be reverse-engineered from what you ship. Model weights and internal pipelines are often better kept secret. Something visible in a product usually needs a patent, because secrecy won’t hold.
Get a Read on Your AI Portfolio
Inventorship and eligibility for AI-related inventions are moving faster than most published guidance reflects, and applications drafted against last year’s framework may need a second look. Goldstein Patent Law works with software, AI, and technology companies from offices in New York City and Red Bank, New Jersey, and with clients nationwide. Book a strategy call to review what you’re building and how to protect it.
