How to patent an AI or machine learning invention
A US patent application on an AI invention needs two things: claims to a specific technical contribution, such as a training method, a model architecture, a data pipeline or a concrete application, and a human inventor.
The USPTO does not accept an AI system as an inventor.
Since a precedential 2025 decision its examiners credit improvements to how a model itself works, while the Federal Circuit has held that claims applying a generic model to a new kind of data are not eligible.
Find the invention
An AI product usually holds several candidate inventions. Training: a new loss, curriculum, fine tuning or data selection method. Architecture: a change to the model structure or how its components connect.
Inference: serving, caching, quantization, routing or retrieval that makes it faster, cheaper or more accurate. Data pipeline: labeling, synthetic data, cleaning or feature extraction.
Application: a model wired into a physical system or a technical workflow in a specific way. Each can support its own claims, and your attorney decides which to lead with.
What the USPTO credits
In Ex parte Desjardins (2025), now cited in the examination manual, the USPTO vacated a Section 101 rejection of a method for training a model on a series of tasks while protecting what it learned on earlier ones.
The application explained how that improved the model itself: less storage, less complexity, performance kept on earlier tasks. The math alone is still an abstract idea; the improvement to how the model or the computer operates is what counts.
The USPTO’s July 2024 guidance update on AI added worked examples along the same lines.
What gets rejected
Using an off the shelf model on a new type of data with no change to the model.
In Recentive Analytics v. Fox (2025) the Federal Circuit held that claims applying generic machine learning to a new data environment, without disclosing an improvement to the model, were not eligible.
Claims that read as a mental process done faster, or as a result with no technical means of reaching it, run into the same problem.
Who the inventor is
Only people. The Federal Circuit held in Thaler v. Vidal (2022) that an inventor must be a natural person.
The USPTO’s revised inventorship guidance of November 2025 treats AI systems as tools: the same conception test applies whether or not AI was used, and an AI system is never named.
Using AI tools along the way is fine; the question is still which people conceived what is claimed. Name them, not everyone who touched the code.
How to write the disclosure
Describe the model concretely: inputs, outputs, architecture, the training objective, the data and how it is prepared, and the settings that matter.
Explain what was wrong with the earlier approach and measure the improvement in accuracy, latency, memory, compute or data needed. Describe alternatives: other architectures, other data, other domains where the method works.
For functional language such as a module that predicts a value, show the steps or the algorithm, because the examination manual requires computer implemented functions to be described in enough detail.
Patent or trade secret
A patent application is published about 18 months after its earliest filing date, in exchange for the right to exclude. Some AI advantages stay hidden anyway: training data, weights and tuning recipes a competitor cannot see from outside.
A technique a competitor could learn by using your product or reading your paper is a different case. Your attorney helps decide which parts to file and which to keep.
Timing
AI teams publish, demo and open source fast. In the US you have one year from your own public disclosure to file; in many other countries any public disclosure before filing ends the right to a patent.
A provisional filed before the paper, the launch or the release secures the date. Start with Tao AI for free, add the $300 prior art search report, then a provisional with our team at $2,500.
