Guide, September 2026
Machine Learning Patent Eligibility in 2026: AI Claims Under 35 U.S.C. 101
A practitioner's explainer on how machine learning claims are examined for eligibility, and how to draft toward the technical improvement rather than the math.
The statute and the exceptions
35 U.S.C. 101 allows a patent for any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement of one. Courts have layered on exceptions for abstract ideas, laws of nature, and natural phenomena. Mathematical concepts and mental processes sit inside the abstract idea exception, which is why machine learning claims meet this question so often.
The Alice/Mayo two-step
Step one asks whether the claim is directed to a judicial exception. Step two asks whether the additional elements amount to significantly more than the exception. The USPTO implements this as its examiner-facing analysis in MPEP 2106, and points to MPEP 2103 through 2106.07 as the current eligibility guidance.
The 2024 Patent Subject Matter Eligibility Guidance Update, effective July 17, 2024, added examples aimed at artificial intelligence claims. It remains the worked-example reference.
The August 2025 reminders to examiners
On August 4, 2025 the USPTO issued a memorandum titled “Reminders on evaluating subject matter eligibility of claims under 35 U.S.C. 101”, addressed to examiners in Technology Centers 2100, 2600, and 3600. It expressly states it is not intended to announce any new USPTO practice or procedure, so read it as a calibration document rather than a change in law.
It covers four things. The Step 2A Prong One difference between a claim that recites a judicial exception and one that merely involves it. The limits of the mental-process category, which does not reach limitations that cannot practically be performed in the human mind. The improvements consideration at Step 2A Prong Two. And a caution against oversimplifying a claim into an “apply it” characterization.
It also reminds examiners that a rejection should issue only where there is a more than 50 percent probability that the claim is ineligible.
For a practitioner that last point is the usable one. If a rejection reads as a coin flip resolved against the applicant, the memorandum is the right thing to cite in the response.
Drafting toward the technical improvement
- Name the technical problem in the specification, then name what improves. An improvement stated in engineering terms survives step two better than one stated in business terms.
- Claim the implementation. Which data, which transformation, which architecture, what happens at inference, what the output controls.
- Avoid claiming the math. A claim whose novelty sits entirely in a formula, with a generic processor bolted on, is the pattern the exception was written for.
- Keep the specification ahead of the claims. Implementation detail you did not write down is not available later to argue that the claim is more than the abstract idea.
Where this fits for a company
Eligibility is a drafting problem before it is a prosecution problem. Most of what determines the outcome is fixed at the moment the specification is written. If you are earlier than that, the guide for AI startups covers what to file and when.
Sources
Common questions
Is machine learning patent eligible under 35 U.S.C. 101?
Machine learning inventions are eligible in principle. Section 101 permits any new and useful process, machine, manufacture, or composition of matter, subject to judicially created exceptions for abstract ideas, laws of nature, and natural phenomena. The work is in showing that a claim is directed to a specific technical implementation rather than to the underlying mathematics.
What is the Alice/Mayo two-step test?
Step one asks whether the claim is directed to a judicial exception, such as an abstract idea, a mathematical concept, or a mental process. If it is, step two asks whether the claim recites additional elements amounting to significantly more than the exception itself. The USPTO implements this framework in MPEP 2106 as its eligibility analysis.
What did the USPTO's August 2025 eligibility memorandum say?
On August 4, 2025 the USPTO issued 'Reminders on evaluating subject matter eligibility of claims under 35 U.S.C. 101' to examiners in Technology Centers 2100, 2600, and 3600. It expressly states it is not intended to announce any new USPTO practice or procedure. It covers the difference between a claim that recites a judicial exception and one that merely involves it, the limits of the mental-process category, the improvements consideration at Step 2A Prong Two, and a caution against oversimplifying a claim into an 'apply it' characterization.
When is a machine learning claim treated as an abstract mental process?
The August 2025 memorandum reminds examiners that the mental-process grouping does not reach claim limitations that cannot practically be performed in the human mind. That matters for AI claims, where steps described in plain language can look like human judgment while in fact describing operations no person performs.
How do you show a technical improvement in an AI patent claim?
Tie the claim to a concrete technical problem and say what improves. Faster convergence, lower memory at inference, better handling of a specific data defect, a result a prior architecture could not reach. Then support that in the specification with enough implementation detail that the improvement is credible rather than asserted.
Where does the USPTO's eligibility guidance live in the MPEP?
The USPTO points to MPEP 2103 through 2106.07 as its current eligibility guidance, with MPEP 2106 carrying the two-step analysis. The 2024 Patent Subject Matter Eligibility Guidance Update, which included examples on artificial intelligence, is the other document examiners work from.
This is a practitioner's explainer, not legal advice for any particular application. For fees, see the patent prosecution cost guide.