AI Regulation Tracker / Agency practice change
USPTO Eases Section 101 Eligibility for AI and Machine-Learning Inventions
In the second half of 2025 the United States Patent and Trademark Office reset how it examines AI and machine-learning claims under 35 U.S.C. 101. An August 4, 2025 examiner memo told examiners to stop sweeping AI and hardware-specific limitations into the abstract-idea mental process bucket, and the Ex parte Desjardins decision, designated precedential on November 4, 2025, held that claims to training a machine-learning model can be patent-eligible. This is agency examination posture, not a change to the statute or to Alice.
For a decade after Alice Corp. v. CLS Bank, software and AI applicants have run into the same wall at the USPTO. An examiner reads a claim, decides that what it describes is really just thinking, drops it into the "mental process" grouping of abstract ideas, and rejects it under 35 U.S.C. 101. For machine-learning inventions that pattern was especially frustrating, because a lot of what a model does looks like classification or prediction, which an examiner can wave off as something a person could do in their head. In the back half of 2025 the USPTO moved to close that gap. It did not change the statute and it did not overrule Alice. It changed how its own examiners are supposed to apply the existing test.
The August 2025 examiner memo
The first piece is a five-page memorandum dated August 4, 2025, from Charles Kim, Deputy Commissioner for Patents, addressed to the technology centers that handle software, AI, and business-method applications. Its subject line is exact: "Reminders on evaluating subject matter eligibility of claims under 35 U.S.C. 101." The memo walks through Step 2A of the eligibility analysis and hits the point that matters most for AI applicants: examiners should stop over-reading the mental process grouping.
The memo puts it directly. "Examiners are reminded not to expand this grouping in a manner that encompasses claim limitations that cannot practically be performed in the human mind." And on AI specifically it states, "Claim limitations that encompass AI in a way that cannot be practically performed in the human mind do not fall within this grouping." In plain terms, if a claim limitation is doing something a human brain is not equipped to do, an examiner should not call it a mental process and reject it on that basis.
The memo also draws a clean line that patent counsel should memorize. Using the USPTO's own worked examples, it says that a limitation reciting "training the neural network in a first stage using the first training set" does not, by itself, recite an abstract idea, because it does not set forth any mathematical relationship, calculation, or formula. Contrast that with a limitation that expressly names the math, for instance one reciting a training algorithm that "includes a backpropagation algorithm and a gradient descent algorithm." That one does recite an abstract idea, because it calls out specific mathematical calculations by name. The drafting lesson is obvious once you see it.
The ceiling: this is a reminder, not a new rule
I want to be precise, because the memo itself is precise. It says, word for word, that it "is not intended to announce any new USPTO practice or procedure and is meant to be consistent with existing USPTO guidance." Nobody got a new legal standard on August 4, 2025. The statute, 35 U.S.C. 101, still reads the same way. Alice and Mayo are still the governing Supreme Court cases. What the memo does is tell examiners how to apply guidance that was already on the books, including the July 2024 AI eligibility update. That distinction matters when you are arguing a rejection: you are holding the office to its own stated practice, not citing a new rule.
Ex parte Desjardins and why it carries more weight
The heavier instrument is Ex parte Desjardins. It is a decision of the USPTO's Appeals Review Panel, issued September 26, 2025 and authored by Director John Squires, that the Director designated precedential on November 4, 2025. The application claimed a method for training a machine-learning model in a way that lets the model learn new tasks in sequence while protecting what it already learned, using less storage in the process. The panel treated that as a technical improvement to the model itself and held the claims eligible under Section 101, analyzing them under the Enfish line rather than treating them as a bare abstract idea.
The reason this outranks the memo is procedural. A precedential Appeals Review Panel decision binds future PTAB panels and signals to examiners how the office will decide close calls. It is the office putting its name behind the proposition that improvements to an AI or machine-learning model can be enough for eligibility, even when a claim recites something that looks abstract on its face. For a patent examiner or an applicant staring at a Section 101 rejection, a precedential decision on point is a far stronger card than a reminder memo.
Subject Matter Eligibility Declarations
The third piece is procedural and easy to miss. The USPTO formalized the use of Subject Matter Eligibility Declarations, filed as declarations under Rule 132 (37 CFR 1.132) and governed by the evidentiary declaration practice in MPEP 716. These let an applicant respond to a Section 101 rejection by putting objective evidence into the record, for example evidence that a claimed technique produces a concrete technical improvement, rather than arguing eligibility on attorney argument alone. It gives AI and machine-learning applicants a documented, evidence-backed way to build the eligibility record instead of trading conclusory statements with the examiner.
What this means for US patent counsel
If you draft or prosecute AI and machine-learning inventions, this is a practical opening, and you should adjust accordingly. Draft claims around the technical improvement the invention actually delivers, and describe that improvement clearly in the specification, because the improvements path is now the office's stated route to eligibility. Be deliberate about naming specific mathematical algorithms in the claims. Reciting backpropagation or gradient descent by name can drag a limitation into abstract-idea territory, while claiming the training operation at a functional level may not. When you hit a mental process rejection, push back with the August 2025 memo language and cite Desjardins, and consider a Subject Matter Eligibility Declaration to put objective proof of the technical improvement on the record.
Keep the ceiling in view when you counsel a client. This is examination practice at the USPTO, not a change in the law. The Federal Circuit and the Supreme Court are not bound by an examiner memo or by an Appeals Review Panel decision, and a patent that issues more easily under this posture still has to survive validity challenges in litigation under the same Alice framework. The posture makes prosecution easier. It does not make an AI patent bulletproof.
Questions professionals are asking
Did the USPTO change the law on AI patent eligibility?
No. The August 4, 2025 memo states that it "is not intended to announce any new USPTO practice or procedure." The statute, 35 U.S.C. 101, and the Supreme Court's Alice and Mayo decisions are unchanged. What changed is how USPTO examiners are told to apply the existing eligibility test to AI and machine-learning claims.
What does the August 2025 memo actually tell examiners?
It reminds examiners not to over-read the "mental process" grouping of abstract ideas. It states that examiners should not expand that grouping to cover claim limitations that cannot practically be performed in the human mind, and that AI limitations which cannot practically be performed in the human mind do not fall within the grouping. It also distinguishes a plain neural-network training limitation, which does not recite an abstract idea, from one that names specific math such as backpropagation and gradient descent, which does.
What is Ex parte Desjardins and why does it matter?
Ex parte Desjardins is an Appeals Review Panel decision authored by USPTO Director John Squires, issued September 26, 2025 and designated precedential on November 4, 2025. It held that claims directed to training a machine-learning model can be patent-eligible under Section 101 when they reflect a technical improvement to the model. As a precedential decision it binds future PTAB panels, which gives it more weight than the reminder memo.
What is a Subject Matter Eligibility Declaration?
It is a voluntary declaration an applicant can file under Rule 132 (37 CFR 1.132), following the evidentiary declaration practice in MPEP 716, to put objective evidence of eligibility on the record in response to a Section 101 rejection. It lets an applicant support eligibility with proof of a technical improvement rather than attorney argument alone.
Should US patent counsel change how they draft AI claims?
Yes, in practical ways. Draft claims around the technical improvement the invention delivers and describe that improvement in the specification, be deliberate about whether to name specific mathematical algorithms in the claims, and use the memo language, Desjardins, and a Subject Matter Eligibility Declaration when responding to mental process rejections. Remember this is examination posture, so a patent that issues more easily still faces the same Alice framework in litigation.
RELATED BRIEFINGS
Browse the full AI Regulation News tracker
Informational analysis for working professionals, not legal advice. Patent eligibility is fact-specific and examination posture can change. Confirm how the USPTO guidance, Ex parte Desjardins, and Rule 132 practice apply to a specific application with qualified patent counsel.