Often yes, if what you built is a real technical improvement: a new model architecture, a training method, a data-processing technique or a way to run models faster or cheaper on hardware. Taking a known model, training it on new data and applying it to a new field is usually not patentable on its own. Inventions made with the help of AI tools can be patented too, as long as a person conceived what’s claimed. For many software companies a trade secret protects the model better than a patent, so decide that question first.
Key takeaways
- The strongest AI patents claim technical improvements: architecture, training, data pipelines and inference efficiency.
- “Known model plus new data” is the weakest case. A federal appeals court rejected claims like that in 2025.
- The Patent Office became friendlier to AI claims in 2025, but courts still apply the same eligibility test, so a granted patent can still fall in litigation.
- Weights, training data and prompts that customers never see are often better protected as trade secrets.
- AI can’t be an inventor. Keep records of what your people contributed.
- This area is moving fast. Expect the rules to shift.
What parts of an AI system can be patented?
Patents protect inventions, not ideas or data. In an AI product, the patentable parts are usually where your engineers did something new to make the system work better: changes to the arrangement of layers or attention mechanisms, a new tokenization approach, a loss function or training method that reduces compute or memory, a pipeline that cleans or transforms data in a technical way, or a technique for running a model more efficiently on particular hardware. Systems that use a model’s output to control something physical, such as a robot, a medical device or a production line, also tend to do well. Even incremental improvements can qualify if they’re new, non-obvious and described in technical detail.
What doesn’t qualify on its own: a data set, a mathematical formula in the abstract, or the idea of using AI for a business task.
| What you built | Patent prospects | Often better protected by |
| New architecture or training method with measurable technical gains | Good, if described in detail | A patent, or a trade secret if it stays hidden |
| Faster or cheaper inference on specific hardware | Good | A patent |
| AI that controls a physical process or device | Good | A patent |
| A known model fine-tuned on your proprietary data | Weak | Trade secrets for the data and weights, plus contracts |
| A known model applied to a new industry or workflow | Weak, often ineligible | Speed, customer relationships, trade secrets and your brand |
| Prompts, system instructions and evaluation sets | Rarely | Trade secrets |
Why is “known model plus new data” so hard to patent?
Because of the patent eligibility test. Under the Supreme Court’s 2014 decision in Alice v. CLS Bank, a claim directed to an abstract idea, such as a mathematical method or a way of organizing information, isn’t patentable unless it adds something significantly more. In Recentive Analytics v. Fox (2025), the Federal Circuit applied that test to machine learning patents covering schedules and network maps for TV broadcasts and live events. It held that patents that do no more than claim the application of generic machine learning to a new data environment are ineligible, because they don’t disclose any improvement to the machine learning models themselves. The Supreme Court declined to review the case in December 2025.
The Patent Office has moved the other way. It published AI-specific eligibility examples in July 2024. In August 2025 it reminded examiners that the “mental process” category covers only steps a person could practically perform in their head, which makes complex model training harder to reject on that basis. And in Ex parte Desjardins (September 2025, designated precedential that November), the agency’s Appeals Review Panel held that claims to a method of training a machine learning model were eligible, because the application described concrete technical improvements, including reduced storage and better learning across multiple tasks, and the claims reflected them.
The result: it’s easier to get an AI patent issued than it was two years ago, but the courts haven’t changed the test. Write the application for both audiences by tying each claim to a specific technical improvement that the description explains.
Should I patent my model or keep it a trade secret?
Start with whether competitors can see how it works. If your model runs behind an API or inside your own product, no one outside the company can inspect the architecture, weights or training data. A patent would require you to publish how the invention works, and you’d have trouble detecting whether a competitor’s hidden model infringes. A trade secret can last as long as you keep it secret, costs nothing to register and covers things patents can’t, like data and weights. You do have to take reasonable steps to protect it: access controls, confidentiality agreements and invention assignment agreements with everyone who touches it.
A patent makes more sense when the invention will be visible: a model that ships on devices, a method that can be worked out from your outputs or published papers, or code you plan to open-source. It also makes sense when you expect competitors to develop the same thing independently, since trade secret law doesn’t stop independent development or reverse engineering. Many AI companies do both: patent the parts that will become visible and keep the weights, data and training recipes secret. If you plan to release model weights or code, see Open Source.
What makes a strong AI patent application?
The best applications explain structure, not just results. Describe how data flows through each component and how the new parts produce the output, rather than listing the components on their own. Tie each new element to the technical improvement it delivers, such as less memory, lower latency or better accuracy on a defined task. Describe the hardware the model runs on where it matters. Include alternatives, such as other architectures, activation functions or training regimes that would also work, so a competitor can’t design around a single embodiment. The more you can explain how and why the model works, the broader the claims you can usually support.
Timing matters as much as content. File before you publish a paper, post to arXiv, demo to people who haven’t signed an NDA or release weights. The US gives you one year after your own disclosure, but most other countries give you none. A detailed provisional application can hold your filing date for 12 months while development continues. See Startup Patents: Are They Worth It? for provisionals, costs and timelines.
AI tools are useful for organizing your notes into an invention disclosure for your patent attorney. Asking one to draft the application usually isn’t. Counsel has to rework the claims and description anyway, which often costs more than starting from a clear disclosure. Use a tool whose terms don’t let the provider train on or share what you enter, because an unfiled invention is confidential.
Can I patent an invention my team made using AI?
Yes, if a person conceived it. Only people can be inventors, and an application that names an AI system as an inventor will be rejected. Under Patent Office guidance issued in November 2025, AI is treated like any other research tool, such as lab equipment or software, and the usual question applies: did a person form a definite and permanent idea of the complete invention? If your team uses AI to generate candidates, write code or run experiments, keep records of what the people did: the problem they framed, the outputs they selected, and how they modified and verified them.
Ownership works the same way as for any invention. The inventors own it until they assign it in writing, so make sure every founder, employee and contractor has signed a CIIAA. See Who Owns Startup IP?
Is a patent worth it for an AI startup?
It depends on where your advantage lies. Companies building models, infrastructure, chips or AI for physical systems often have inventions worth patenting, and investors in those sectors may expect a filing strategy. Companies building applications on top of someone else’s model usually don’t. Their edge is product, distribution and data, and a patent on the workflow is likely to be rejected or invalidated. Remember too that a US patent typically takes two to three years to issue, and AI techniques can be obsolete by then, so focus on improvements you expect to matter for years.
For more, see How Do I Protect My Startup Idea?, Startup Intellectual Property and How Should a Startup Prioritize IP Protection on a Limited Budget?
Westaway doesn’t prosecute patents, but we work with patent counsel on ownership, assignments and strategy. If you’d like help deciding whether to patent, contact us.
