Is it legal to patent an AI-generated invention?

The Short Answer: Yes, But Only With a Human Inventor

As of August 2026, the answer to whether you can patent an AI-generated invention is a qualified yes. You can obtain a patent on an invention that was created with substantial help from artificial intelligence, but you cannot obtain one if the AI system itself is named as the inventor, and you cannot get one at all in most jurisdictions if no human made a genuine inventive contribution. Every major patent office that has ruled on this question — including the United States Patent and Trademark Office (USPTO), the UK Intellectual Property Office, the European Patent Office, and courts in Australia and elsewhere — has rejected applications listing an AI system as the sole inventor.

Also worth reading: What should be on an AI patent drafting security checklist before sending invention disclosures to a generative AI tool? · Who is the inventor on a patent when an agentic AI system contributes to the invention in 2026? · How can I check if my invention is patentable and if a patent is valid before filing?

The test case that defined this area is DABUS, an AI system created by Stephen Thaler. On 17 September 2019, Thaler filed patent applications for a "food container" and related devices, naming DABUS as the inventor with himself as the assignee. Those applications were refused in the US, UK, Europe, and multiple other jurisdictions because their laws require inventors to be natural persons. Courts upheld those refusals through successive appeals. The consistent judicial reasoning is that an inventor must be a person, both because patent statutes say so and because inventorship carries legal duties — such as the declaration of inventive contribution — that only humans can perform.

So the practical rule is straightforward: if your invention came out of an AI tool but a human identified the problem, directed the AI, recognized the inventive result, and can explain how it works, you can likely patent it by naming the human(s) as inventors. If the machine did everything autonomously with no meaningful human contribution, the invention currently falls into a protection gap in nearly every country.

Why the Law Draws This Line

Patent systems around the world are built on the premise that an inventor is a natural person. In the United States, the Patent Act refers to inventors as individuals, and in Thaler v. Vidal (2022), the Federal Circuit held flatly that "an 'inventor' must be a natural person." The USPTO reinforced this position in February 2024 when it issued formal guidance on inventorship for AI-assisted inventions, followed by additional guidance updates in 2025 clarifying how examiners should evaluate human contribution.

The reasoning goes beyond statutory text. A named inventor must make specific declarations under penalty of law, including affirming they believe themselves to be the original inventor. An AI system cannot swear declarations, cannot be sued for inequitable conduct, cannot assign rights in a legally enforceable way, and cannot hold moral entitlement to a patent. Legislatures could rewrite these rules — some scholars have proposed creating a new sui generis right for autonomous AI creations — but as of mid-2026, no major jurisdiction has done so. Thaler's arguments that denying AI inventorship would discourage innovation were considered and rejected; courts reasoned that the patent bargain (disclosure in exchange for exclusivity) functions fine when humans remain in the loop.

There is also a policy concern about disclosure quality. Patents require enough detail for a person skilled in the art to reproduce the invention. If an AI produced a result nobody fully understands, the application may fail the enablement requirement regardless of who is named as inventor. This is an underappreciated risk: even human-named AI-assisted patents can be invalidated later if the specification reads like unexplained model output rather than a genuine technical teaching.

How the USPTO Decides Who Counts as an Inventor

The February 2024 USPTO guidance, titled "Inventorship Guidance for AI-Assisted Inventions," does not create a bright-line percentage test. Instead it applies existing case law on joint inventorship: each named inventor must contribute to the conception of the invention, and their contribution must not be insignificant when measured against the full invention. Applying AI tools during research or development does not automatically disqualify someone from being an inventor, just as using a microscope, a simulation package, or a lab robot does not.

The guidance lays out several factors examiners and practitioners weigh. Did the human pose the problem that the AI solved? Did they design the prompt, the training data selection, or the architecture? Did they recognize that the AI output was inventive — a moment courts treat as analogous to conception? Could they reduce the output to practice, verify it experimentally, and explain why it works? A researcher who iterates with a generative model, tests dozens of candidate molecules, identifies the one that works, and articulates the mechanism has plainly contributed to conception. Someone who pastes raw model output into a filing without understanding it has not.

The practical consequence is that companies should document the human contribution trail contemporaneously. Lab notebooks, prompt logs, iteration records, and internal memos identifying why a particular AI output was selected all become evidence of inventorship. Patent prosecutors increasingly build this record deliberately, because examiner rejections based on improper inventorship — or later invalidity challenges — turn on exactly this documentation. Firms that treat AI outputs as black boxes and file whatever comes out are accumulating real prosecution and litigation risk.

Comparison: Human-Invented vs. AI-Assisted vs. Fully Autonomous

The middle column is where almost every real-world case now sits. Modern R&D uses AI the way earlier generations used computational chemistry or CAD: as a powerful tool operated by people who remain responsible for the inventive act. The right-hand column remains largely theoretical outside of Thaler-style test cases, precisely because genuinely autonomous invention without any human direction is rare, and because applicants know such filings will fail.

Practical Steps Before You File

First, conduct an honest inventorship analysis before drafting. List everyone who contributed to conception — problem identification, solution design, output recognition, reduction to practice — and exclude anyone whose role was purely administrative. Overstating inventorship is itself a defect; omitting a true joint inventor can render the patent unenforceable. When AI was involved, document specifically what the human contributed versus what the model generated.

Second, preserve the AI interaction record. Save prompts, model versions, dates, seeds or parameters where available, and the sequence of iterations. This serves two purposes: it evidences human conception, and it addresses the growing concern flagged in recent commentary from sources like The National Law Review that disclosing confidential invention details to third-party generative AI tools can create prosecution risk. Some AI providers' terms grant them broad rights over inputs and outputs, which can raise prior-art and ownership questions. Enterprises increasingly restrict which models may see unpublished invention data, or use self-hosted models for sensitive work.

Third, ensure the specification demonstrates human mastery. Examiners and future litigants will probe whether the applicants actually understand the invention. Include experimental validation, comparative data against known solutions, and explanations of mechanism. Fourth, consider jurisdictional strategy: standards differ slightly between the USPTO, EPO, and other offices regarding how much human involvement suffices, so a global filing plan should account for the strictest regime among target countries. Finally, budget realistic time — AI-related inventorship questions add weeks to prosecution, and examiner interviews on this topic are becoming routine.

Common Mistakes That Sink AI-Related Applications

The most damaging mistake is naming an AI system as an inventor, even partially. It triggers near-certain rejection and invites scrutiny of the entire application. The second is the opposite error: listing a senior executive or funding scientist as an inventor to look legitimate when they contributed nothing. Courts have voided patents for misjoinder and nonjoinder of inventors, and inequitable conduct findings can make the whole family unenforceable.

A third mistake is treating AI output as a finished invention. Raw generative output frequently contains plausible-sounding but technically wrong content. Filing without independent verification risks failing the enablement and written-description requirements, and worse, publishing incorrect technical assertions that competitors can exploit. Fourth, many applicants ignore confidentiality discipline: feeding an unpublished core invention into a public chatbot may constitute a public disclosure in some analyses, starting a one-year clock in the US and destroying novelty immediately in absolute-novelty jurisdictions like Europe. Fifth, teams sometimes assume software-implemented AI inventions face no subject-matter eligibility issues. They do. Most countries limit what software can be patented, there is no single agreed legal definition of artificial intelligence, and abstract-idea rejections remain common for AI applications that claim little more than applying a model to data. Claims should be drafted around concrete technical improvements, not the algorithm in the abstract.

Finally, some organizations overcorrect and refuse to use AI in R&D at all, fearing contamination of inventorship. That is unnecessary. Used properly, AI is simply a tool, and the USPTO guidance explicitly confirms AI assistance does not preclude patentability.

Costs, Timing, and Where Professional Review Fits

Filing costs are largely unchanged by the AI dimension. A US non-provisional application with professional drafting typically runs $8,000–$15,000 for moderate complexity, plus USPTO fees ranging from roughly $300 to $1,800 depending on entity size, and prosecution often adds $5,000–$20,000 over three to five years. What AI involvement changes is the labor mix: expect added hours for inventorship documentation, claim-drafting around eligibility issues, and possible examiner interviews. Budget an extra 10–25% on prosecution fees for AI-heavy filings compared with conventional mechanical cases.

Timing matters more than usual. Because AI fields move fast, prior art accumulates quickly, and provisional filings within months of conceiving an AI-assisted invention are prudent. Reviewing drafts with AI-assisted patent-analysis tools before filing has become standard practice at many firms — used to check claim scope, spot weak support, and run prior-art sweeps — though practitioners caution that outputs must be verified by qualified attorneys, since hallucinated citations and missed references remain documented failure modes. Disclosure of sensitive material to consumer-grade AI tools should be governed by written company policy, not left to individual engineers.

When to Act and What the Future Holds

Act now if you have an AI-assisted invention ready for filing: the current rules reward early, well-documented filings, and waiting exposes you to competitor prior art. Act now also if your organization lacks an AI-use policy for R&D — the gap between informal experimentation and defensible inventorship records widens daily.

Legislatively, change is possible but slow. Proposals to create special rights for autonomous AI creations have been debated in the UK, US, and at WIPO, which publishes guidance for small and medium-sized enterprises navigating AI and IP. As of August 2026 none has been enacted in a major economy, and courts show no appetite to expand inventorship by interpretation. The pragmatic consensus among practitioners is that the human-in-the-loop model will persist for years. Companies that build disciplined documentation habits today will find the transition painless if rules evolve; those that file sloppy AI-assisted applications will discover the weaknesses during litigation, when it is far too late to fix them.