In September 2025 the Office’s Appeals Review Panel vacated a § 101 rejection of a DeepMind machine-learning claim, and the decision became precedential in November. The claim recited math. It was eligible because it reflected an improvement to how the model itself works. The applicant still faced an obviousness rejection.
Before you read
A claim recites a mathematical calculation for training a model. Under Desjardins, is it ineligible for that reason?
Not necessarily. Claim 1 recited math, but the claim as a whole reflected an improvement to the model’s operation, so it was “not directed to an abstract idea.”
Part 1 of 7
A Board panel had affirmed an obviousness rejection of a DeepMind application on training machine-learning models. Then, on its own initiative, it entered a new rejection: the claims were ineligible abstract ideas under § 101. An Appeals Review Panel led by the Director was convened to review that step. It vacated the new ground, calling the reasoning “overbroad” and “troubling.”
“Categorically excluding AI innovations from patent protection in the United States jeopardizes America’s leadership in this critical emerging technology.”
Part 2 of 7
Claim 1 covers a computer-implemented method for training a model that already learned one task to learn a second, different task while protecting performance on the first. It does this by weighting how important each parameter was to the first task.
Eligibility runs through the Supreme Court’s Alice framework: laws of nature, natural phenomena and abstract ideas are not patentable. The Office applies it in steps.
Part 3 of 7
A statement in the specification isn’t enough on its own; the claim has to reflect the improvement. Here it did. The specification described a model that can learn new tasks in succession while protecting what it learned before, using less storage because one set of parameters serves several tasks. The claim’s limitation carried that through.
The specification’s stated benefit
The model can effectively learn new tasks in succession whilst protecting knowledge about previous tasks.
Claim 1’s limitation
“adjust the first values of the plurality of parameters to optimize performance of the machine learning model on the second machine learning task while protecting performance of the machine learning model on the first machine learning task”
What the panel concluded
an improvement to how the machine learning model itself operates
Source
The holding: “although independent claim 1 may recite an abstract idea, it is not directed to an abstract idea.” The same analysis covered claims 18 and 19 and every dependent claim.
Part 4 of 7
The panel criticized reasoning that treats any machine learning as an unpatentable algorithm and the rest as generic computer parts: “Examiners and panels should not evaluate claims at such a high level of generality.” It also said where the real limits are.
Dismissing a claim as nothing more than machine learning plus generic hardware.
“The traditional and appropriate tools to limit patent protection to its proper scope. These statutory provisions should be the focus of examination.”
The claims still stood rejected as obvious. The applicant won the eligibility issue without getting a patent from this decision.
Part 5 of 7
Desjardins is one of two precedential § 101 decisions in ex parte appeals on the USPTO list, with Mewherter (2013).
16 entries on the USPTO list, Oct 2026
Tap a segment to see the decisions.
Source
Part 6 of 7
Choose your side.
Part 7 of 7
Desjardins binds the Office within its scope, not the courts. A claim the Office finds eligible can still be held ineligible in litigation, and a court’s § 101 ruling can end a pending IPR, as it did in Hulu v. Piranha. An IPR can’t hear § 101 at all.
It decides
It doesn’t decide
The takeaway
Report the eligibility win, then the next obstacle between the client and useful claims.
Precedential decisions bind the Board within their designated scope. Informative decisions show the Office’s reasoning but do not bind the Board. Court opinions are cited for what they decided; later history was checked on the date shown above.