Yes, you can patent AI in the United States, but not every use of AI qualifies. Patents are available for specific technical advances, such as a new way to train a model, an architecture that uses less memory, or a system in which model outputs control another technology. Applying a generic, off-the-shelf model to a new business problem is much harder to protect, and the Federal Circuit has held claims of that kind ineligible.
This guide covers U.S. law as of November 2026. It is written for founders, engineers, and in-house teams deciding whether and how to pursue AI patents.
What AI inventions can be patented?
An AI invention has to meet the same requirements as any other invention. It must be eligible subject matter (35 U.S.C. § 101), new (§ 102), non-obvious (§ 103), and described well enough that a skilled person could make and use it (§ 112). Eligibility draws the most attention, but it is only the first hurdle. Published papers and open-source code are prior art, so a claim that clears § 101 can still fail under §§ 102 or 103.
The question that usually decides eligibility is whether the invention improves the technology itself or only uses the technology to do something else. These examples show how that tends to play out. They are general patterns, not guarantees.
| Stronger candidates | Weaker candidates |
|---|---|
| A training technique that makes the model itself work better, such as reducing storage needs or preserving earlier learning | Applying conventional machine learning to a new field or type of data |
| A new model architecture with a specific, explained technical advantage | Doing a task people already do, only faster or at larger scale |
| Inference optimization that reduces latency, memory, or compute on identified hardware | Training a standard model on domain-specific data, with nothing more |
| Retrieval, orchestration, memory, or guardrail mechanisms in LLM and agentic systems that solve a stated technical problem | Sending a prompt to an off-the-shelf model and presenting the answer |
| Model outputs that directly control a machine, process, or other technical system | Claims defined by a desired result, without saying how it is achieved |
The Section 101 split: USPTO versus the courts
Eligibility is assessed under the two-step Alice/Mayo framework. Is the claim directed to an abstract idea, and if so, does it add something that transforms it into an eligible invention? For AI, the important recent development is that the USPTO has become more receptive to these claims, while the Federal Circuit has continued to invalidate claims that rely on generic machine learning.
The courts: generic machine learning is not enough
In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), the court held ineligible machine learning patents directed to event scheduling and network maps. Its holding was specific:
“[P]atents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101.”
The court also rejected the arguments that a new field of use, or doing a human task with greater speed and efficiency, makes a claim eligible. It treated iterative training and dynamic updating as incident to the very nature of machine learning.
In July 2026, a nonprecedential decision applied similar reasoning to deep learning image analysis. In Dental Monitoring SAS v. Align Technology, Inc., No. 2024-2270 (Fed. Cir. July 7, 2026), the court affirmed summary judgment that claims analyzing dental images with a deep learning device were ineligible. A requirement to train on more than 1,000 labeled images did not save them. The decision does not bind future panels, but it shows the risk of relying on training data alone.
The USPTO: real improvements to machine learning are eligible
In Ex parte Desjardins, Appeal No. 2024-000567, a USPTO Appeals Review Panel found eligible a claimed method of training machine learning models that addressed “catastrophic forgetting,” preserving performance on earlier tasks while reducing storage needs. The panel relied on Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016), and cautioned against evaluating claims at too high a level of generality. It also said that §§ 102, 103, and 112 are the tools for policing novelty, obviousness, and claim scope.
The decision was designated precedential in November 2025, and the USPTO incorporated it into MPEP § 2106 in December 2025.
The USPTO also now invites Subject Matter Eligibility Declarations (SMEDs), voluntary declarations under 37 C.F.R. § 1.132 that put objective evidence of a technological improvement in the record. Its best practices were updated on April 30, 2026. A declaration can support the application, but it cannot supply what the application should have disclosed in the first place.
Why you need to satisfy both
USPTO guidance decides whether you get the patent. Court decisions decide whether it survives when you enforce it. USPTO guidance does not bind federal courts, and future USPTO leadership can change it.
The two lines are not necessarily in conflict. Recentive expressly addressed claims that did not disclose improvements to the models, and Desjardins involved claims that did. The safest position is on the improvement side of that line, with the improvement written into both the specification and the claims.
For more detail, see how to patent machine learning after Recentive and our in-depth look at Section 101 for AI.
How to write an AI patent application that holds up
Eligibility and disclosure are largely decided by what you write at filing, because new matter cannot be added later. A strong AI application will usually:
- Name the technical problem. Describe the limitation in existing systems, such as memory growth, retraining cost, latency, or model drift, not just the business problem.
- Explain how the invention solves it. Describe the mechanism: the architecture, data structures, loss function, retrieval logic, or sequence of operations. Do not rely on phrases like “a trained model determines.”
- Connect the mechanism to a result. State the improvement and, where you can, include measurements such as benchmarks or comparisons with a baseline. These can also support a later declaration.
- Put the improvement in the claims. The MPEP looks for an improvement that is described in the specification and reflected in the claims. A great specification will not rescue a claim that leaves out what makes the invention work.
- Describe the training and the data with enough detail. Explain how inputs are selected and processed, how training works, and what the model outputs. Under Amgen Inc. v. Sanofi, 598 U.S. 594 (2023), broad claims must be enabled across their full scope.
- Include alternatives and fallback positions. Describe variations and combinations of features so you have support for narrower claims if the broad ones are rejected. See how multiple dependent claims preserve those combinations.
How to claim an AI invention
AI systems are usually split across companies. One may train the model, another may host it, and a third may use it. A method claim generally requires one actor to perform, or be responsible for, every step, so a single claim that spans those layers may leave you with no direct infringer. Most AI portfolios should therefore separate the claim types:
| Claim type | What it covers | Who it tends to reach |
|---|---|---|
| Training method | The steps used to create or update the model | The developer that runs the training |
| Trained model or storage medium | A model with defined structural features | Those who make, use, or sell the model |
| Inference method or system | Running the model in a defined way | Platforms and product companies that serve it |
For systems built around LLMs, see how to patent agentic AI technology. The first drafting question is who you would want to enforce against. Our article on divided infringement and AI patent claims covers this in depth, with sample claim patterns.
Who is the inventor when AI helped?
Only people can be inventors. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), held that the Patent Act requires inventors to be natural persons, and the Supreme Court declined review in 2023.
Using AI tools does not prevent you from getting a patent. The USPTO’s revised inventorship guidance, 90 Fed. Reg. 54636 (Nov. 28, 2025), rescinded the February 2024 guidance in its entirety and changed the analysis in several ways:
- The same legal standard applies to every invention, whether or not AI was used. There is no separate test for AI-assisted inventions.
- AI systems are treated as instruments used by human inventors, analogous to laboratory equipment.
- Conception is the touchstone. The question is whether a person formed a definite and permanent idea of the complete invention.
- The Pannu factors apply only when deciding whether multiple people are joint inventors.
In practice, keep records of who conceived what, including how people directed, evaluated, and changed any AI-generated output. How courts will apply conception to heavily AI-generated inventions remains an open question.
When to file
AI teams tend to publish, demo, and ship early, so timing mistakes are common and often cannot be fixed.
- File before you disclose if foreign rights matter. The U.S. generally allows a one-year grace period after your own disclosure, but many other countries do not. A paper, public repository, launch, or non-confidential demo can end those rights.
- Use a provisional application for an early date, not as a shortcut. You have 12 months to file the non-provisional, and it only gets the earlier date for what the provisional actually supports. See our guide to moving from provisional to non-provisional.
- Remember that applications usually become public. Most U.S. applications publish about 18 months after the earliest filing date, which matters when you are choosing between a patent and secrecy.
Patent or trade secret?
A patent requires you to disclose the invention, and a trade secret requires you to keep it secret. The decision usually depends on whether a competitor’s use could be detected:
- Patents tend to suit features visible in a product or API, inventions that competitors could independently develop or reverse engineer, and assets that matter for investors or licensing.
- Trade secrets tend to suit internal training techniques that are hard to detect from the outside, proprietary datasets, weights kept behind an API, and tuning details that change quickly.
Open source raises a similar choice. See patent protection vs. open source for startups. Many companies use both patents and trade secrets, patenting the visible architecture and keeping internal know-how confidential. Make sure that withholding details does not leave the application short of what § 112 requires.
A practical path to filing
- Inventory what is new. List the technical work that differs from standard practice, ignoring for now the features that only differ commercially.
- Check your disclosure dates. Identify any papers, repositories, demos, or launches, whether past or planned.
- Search the prior art. Include academic papers and open-source projects as well as patents.
- Test for a technical improvement. For each candidate, state the technical problem, the mechanism, and the measurable result. If you can only state a business benefit, reconsider it.
- Decide who you want to reach. Choose the likely infringers first, then design claim groups around them.
- Document the invention. Capture architectures, pseudocode, alternatives, experimental results, and who conceived each part.
- File, then build the portfolio. File a provisional or non-provisional application, and keep identifying new filings as the technology develops.
Frequently asked questions
Is a patent allowed by the USPTO safe from a Section 101 challenge?
No. USPTO guidance and decisions such as Ex parte Desjardins govern examination, but they do not bind federal courts. A granted patent can still be challenged as ineligible in litigation, so an AI application should be written to hold up under Federal Circuit decisions such as Recentive as well as under USPTO guidance.
Can you patent an AI algorithm?
Not as an abstract mathematical idea. You can seek protection for a specific way of implementing an algorithm that produces a technical improvement, such as a training technique that reduces memory use or an inference method that lowers latency on particular hardware.
Can you patent an application built on someone else’s AI model?
Sometimes. Using a third-party model does not rule out a patent, but sending it a prompt and using the answer is unlikely to qualify. The stronger subject matter is usually the technical system around the model, such as how it retrieves context, coordinates tools, validates outputs, or controls another system.
Does publishing a paper or open-sourcing code stop you from patenting AI?
It can. The United States generally gives inventors a one-year grace period after their own public disclosure, but many other countries do not. If rights outside the United States matter, file before the paper, repository, or demo goes public.
Can an invention developed with generative AI tools be patented?
Yes, if a human inventor conceived it. The USPTO’s November 2025 guidance treats AI systems as tools used by human inventors and applies the same conception standard to all inventions. An AI system cannot be named as an inventor or joint inventor.
How long does an AI patent last?
AI patents follow the same rule as other U.S. utility patents. The term generally ends 20 years after the earliest non-provisional filing date to which the patent claims benefit, subject to possible adjustments and to payment of maintenance fees. Time spent in a provisional application does not count against that term.