Whether an artificial-intelligence invention can be patented in the United States turns first on a single statutory sentence and the case law layered on top of it. 35 U.S.C. 101 states the categories of patentable subject matter: "Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title." Machine-learning methods comfortably read on the "process" and "machine" categories on their face. The problem is that the Supreme Court has carved out three judicial exceptions that sit outside those categories, and AI claims tend to brush up against all of them.

The USPTO instructs examiners to evaluate eligibility through the framework described in MPEP 2106, which the agency calls the Alice/Mayo test after the two Supreme Court decisions that built it: Mayo Collaborative Services v. Prometheus Labs., Inc., 566 U.S. 66 (2012), and Alice Corp. Pty. Ltd. v. CLS Bank Int'l, 573 U.S. 208 (2014). MPEP 2106 frames the inquiry in its own words.

"The Supreme Court in Mayo laid out a framework for determining whether an applicant is seeking to patent a judicial exception itself, or a patent-eligible application of the judicial exception."— USPTO, MPEP 2106, source

The three judicial exceptions matter because machine learning is, at its core, mathematics. MPEP 2106 lists the exceptions as "abstract ideas, laws of nature and natural phenomena (including products of nature)." A neural network is a sequence of mathematical operations; a training procedure is an algorithm; a model's output is the result of a calculation. Each of those can be characterized as an abstract idea — specifically, as a mathematical concept or a "mental process" — and that characterization is where many AI claims run into trouble.

The two-step structure examiners apply

The MPEP recasts the Mayo framework as a flowchart with named steps. Step 1 asks whether the claim falls within one of the four statutory categories (process, machine, manufacture, composition of matter). A claimed method or system almost always clears this. Step 2A then proceeds in two prongs: Prong One asks whether the claim recites a judicial exception, and Prong Two asks whether the claim nonetheless integrates the exception into a practical application. If the claim recites, say, a mathematical formula but applies it to improve the functioning of a computer or to solve a specific technical problem, the integration prong can carry it to eligibility without further analysis.

If a claim recites an exception and does not integrate it into a practical application, the examiner moves to Step 2B, which asks whether the claim adds an inventive concept — something significantly more than the exception itself. Under MPEP 2106, generic computer components performing generic functions, or instructions to simply "apply" the exception, do not supply that inventive concept. This is why a claim reciting "a processor configured to apply a machine-learning model to classify data" is vulnerable: stripped of the model's mathematics, what remains is a generic computer doing generic things.

For AI applicants, the practical lesson embedded in the framework is about specificity and technical effect. The USPTO's subject-matter-eligibility guidance, which the MPEP incorporates, directs examiners to look for an improvement to the functioning of a computer or to another technology. A claim that ties its machine-learning step to a concrete technical improvement — reducing memory usage, enabling a hardware capability, improving a sensor-driven control system — supplies the kind of practical application Step 2A Prong Two is looking for. A claim that describes the model in the abstract and leaves the result as data on a screen does not.

What this means for how AI claims are drafted

The eligibility framework explains a drafting pattern visible across granted AI patents: claims are anchored to structure and to a technical context rather than left as bare algorithms. The same discipline that MPEP 2106 imposes on examiners shapes how practitioners write. An independent claim that recites the specific architecture of a network, the particular data transformation, and the technical outcome is reciting more than a mathematical concept; it is reciting an application of one. By contrast, a claim that could be performed in the human mind, or with pen and paper, is the paradigm of the "mental process" abstract idea the courts have refused to patent.

None of this is unique to AI — the Alice/Mayo test governs eligibility across software, diagnostics, and business methods alike. But AI concentrates the difficulty because the invention so often is the math. The statute's four categories are broad and the judicial exceptions are narrow in theory, yet in practice the line between "the abstract idea itself" and "a patent-eligible application of the abstract idea" is exactly where AI prosecution is fought. MPEP 2106 is the document that tells examiners, and therefore tells applicants, how that line is drawn. Reading it is the starting point for understanding why two AI patents covering similar underlying mathematics can land on opposite sides of 35 U.S.C. 101 depending entirely on how their claims are framed.

The takeaway is that eligibility is a claim-language question, not a technology question. The same machine-learning method can be ineligible when claimed as an abstract calculation and eligible when claimed as a specific technical application that integrates that calculation into a practical, structurally grounded use. The statute supplies the categories; Mayo and Alice supply the exceptions; MPEP 2106 supplies the two-step procedure the Office uses to decide which side of the line a given AI claim sits on.