Artificial intelligence inventions are patentable. Thousands issue every year across model architectures, training methods, inference optimization, and applied systems.
They are also rejected under 35 U.S.C. § 101 more often than most technology areas, which has produced a persistent myth that AI cannot be patented. The rejection rate is real. The conclusion drawn from it is not.
What separates the applications that issue from the ones that die is almost entirely how they are drafted.
The Test
Eligibility is governed by the two-step framework from Alice Corp. v. CLS Bank (2014) and Mayo v. Prometheus (2012):
Step one: Is the claim directed to an abstract idea, law of nature, or natural phenomenon? Mathematical concepts, mental processes, and methods of organizing human activity are the recurring categories. Most machine learning claims touch a mathematical concept somewhere.
Step two: If so, does the claim contain an inventive concept that transforms it into significantly more than the abstract idea itself?
The USPTO's 2019 Revised Patent Subject Matter Eligibility Guidance refined this into a more workable analysis. In practice, examination turns on one question: does the claim recite a specific technical improvement, or does it recite a result?
The Distinction That Decides It
Consider two ways to claim the same underlying work.
Result-oriented: "A method for detecting fraudulent transactions comprising receiving transaction data, applying a machine learning model to the transaction data, and outputting a fraud determination."
This is an abstract idea (analyzing data to reach a conclusion) with a generic computer applied to it. This claim loses. It would have lost in 2015 and it loses now.
Improvement-oriented: "A method for training a fraud detection model comprising generating synthetic minority-class transactions using a conditional generative adversarial network conditioned on merchant category and time-of-day features, weighting the synthetic samples by discriminator confidence, and training a gradient-boosted classifier on the augmented dataset, wherein the weighting reduces false positive rate on imbalanced production traffic."
This claims a specific technical mechanism that addresses a specific technical problem, class imbalance in fraud data, with a described technical effect. This has a real path.
The invention behind both claims may be identical. The difference is entirely in the drafting.
The most common failure is not that the invention was ineligible. It is that the application described what the system accomplishes instead of how it works.
What Makes an AI Application Survive
Claim the architecture, not the outcome. Specific model structures, loss functions, sampling strategies, quantization schemes, attention mechanisms, and data pipeline steps are technical subject matter. "Determining a recommendation" is not.
Describe a technical problem in technical terms. Not "improves customer satisfaction." Rather: reduces inference latency, cuts memory footprint during training, improves convergence on sparse data, decreases labeled-data requirements, reduces model drift under distribution shift. Federal Circuit cases such as Enfish and McRO turned on improvements to computer functionality itself.
Include the technical effect and quantify it. If the method cuts training time by 40 percent or reduces parameter count by an order of magnitude, that belongs in the specification. It supports both eligibility and non-obviousness.
Give real implementation detail. Layer configurations, feature engineering, preprocessing steps, hyperparameter relationships, hardware interactions. Sparse specifications produce both § 101 rejections and § 112 enablement problems, and you cannot add detail later.
Avoid generic "apply it" framing. Applying a known model to a new dataset in a conventional way is the paradigm case of ineligibility.
Where AI Patents Are Working
Filings issue steadily in:
- Model architecture and training methods, where the technical subject matter is inherent
- Inference optimization, including quantization, pruning, distillation, and hardware-aware scheduling
- Applied systems with physical grounding, such as AI-driven medical devices, autonomous navigation, industrial control, and robotics, where the claim ties to a physical process
- Data pipeline and infrastructure inventions, including distributed training, federated learning, and privacy-preserving techniques
- Domain-specific systems where the improvement is architectural rather than merely applicational
Harder ground: pure business-logic applications, recommendation engines claimed at a high level, and anything that reads as "conventional model, new industry."
Two Practical Considerations
Inventorship must be human. The Federal Circuit held in Thaler v. Vidal (2022) that an inventor under the Patent Act must be a natural person. AI-assisted inventing is fine, and the USPTO issued guidance in 2024 confirming that inventions made with AI assistance are patentable so long as a natural person made a significant contribution to conception. An AI system cannot itself be named as an inventor.
Consider whether you should patent at all. Patents publish. A training methodology that is undetectable in a deployed product may be better protected as a trade secret, since nobody can reverse engineer it from your API and a patent would hand competitors the recipe. Model architectures visible in a shipped product, or that you intend to license, are better patented. Many AI companies should run both strategies simultaneously, and the analysis is covered in patent vs trade secret.
The Bottom Line
AI inventions are patentable. Eligibility is a drafting discipline, not a coin flip. The applications that fail usually failed at the specification stage, months before an examiner ever saw them, by describing a business result instead of a technical mechanism.
If your invention involves a genuine technical improvement, and most real AI work does, the job is making sure the application says so in the language the statute and the case law respond to.
To discuss protection for an AI or machine learning invention, see our AI patents practice or schedule a consultation.
This article is general educational information, not legal advice, and reading it does not create an attorney-client relationship. Patent law is fact specific and deadlines are unforgiving. For advice on your situation, schedule a consultation.