IP Law Daily, PATENT—P.T.A.B.: USPTO Director designates as precedential ARP decision vacating § 101 rejection for AI-training claims, (Nov 5, 2025)

By Saurabh Kashyap, B.A., M.A., LL.B., LL.M.

Director Squires’ panel vacated a PTAB § 101 rejection, holding that claims improving the operation of a machine-learning model constitute patent-eligible subject matter.

USPTO Director John A. Squires, on November 4, 2025, designated as precedential an Appeals Review Panel (ARP) decision vacating the PTAB’s new ground of rejection under 35 U.S.C. § 101 of a patent application, finding that claims directed to continual machine-learning training were not “abstract ideas” but integrated into a practical technological application. While the panel left standing the examiner’s § 103 rejection for obviousness, the decision—now precedential—establishes a significant benchmark for assessing the eligibility of artificial-intelligence inventions. (In re Desjardins, Appeal No. 2024-000567 (P.T.A.B. Sept. 26, 2025) (precedential designation Nov. 4, 2025)).

Background. The appellants, all researchers affiliated with DeepMind Technologies Limited, were represented by Fish & Richardson P.C. before the USPTO's Technology Center 2100. DeepMind, a pioneer in artificial intelligence research, sought protection for methods of efficiently training a single neural network model to perform multiple tasks without performance degradation. The USPTO served as appellee through its internal PTAB review structure.

The subject application, U.S. Application No. 16/319,040, titled "Training Machine Learning Models," claims a computer-implemented method for sequentially training a neural network model on different tasks while preserving knowledge from earlier ones. Representative claim 1 recites computing "an approximation of a posterior distribution" over model parameters, assigning each a measure of importance, and training on new data to optimize performance "while protecting performance on the first task." The specification (¶ 21) explains that the approach mitigates catastrophic forgetting, reduces storage needs by maintaining a single parameter set, and lowers system complexity.

On March 4, 2025, a PTAB panel affirmed the examiner’s § 103 rejection of claims 1-6 and 8-20 and, sua sponte, introduced a new ground of rejection under § 101, deeming the claims directed to mathematical concepts. DeepMind filed a Request for Rehearing on May 5, 2025, which the Board denied on July 14. The applicants then sought review by the ARP under 35 U.S.C. § 6(b).

Director’s analysis. The ARP, led by Director Squires, Acting Commissioner Valencia Martin Wallace, and Vice Chief Judge Michael W. Kim, confined its review to Alice step one (MPEP § 2106, Step 2A). Under Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 573 U.S. 208 (2014), the first inquiry asks whether a claim is "directed to" a patent-ineligible concept. The panel agreed that computing a posterior distribution was a mathematical operation and, therefore, an abstract idea under Step 2A Prong One. However, under Prong Two, the ARP found that the claims integrated that idea into a practical application that improved computer functionality itself.

Citing Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016), and McRO, Inc. v. Bandai Namco Games Am. Inc., 837 F.3d 1299 (Fed. Cir. 2016), the panel observed that software innovations may be patent-eligible when they enhance computer operation rather than apply an abstract process. The claims, it concluded, addressed a specific technical problem—preserving model performance across sequential learning—by modifying the training mechanism itself. That improvement constituted “an advancement in the functioning of the machine-learning model,” not merely an abstract mathematical result.

Caution against categorical exclusions. The ARP criticized the original PTAB panel’s reasoning as overly general and inconsistent with Federal Circuit precedent. “Categorically excluding AI innovations from patent protection jeopardizes America’s leadership in this critical emerging technology,” the decision warned, emphasizing that examiners should not equate all machine-learning algorithms with unpatentable “mathematical formulas.” The panel reiterated that §§ 102, 103, and 112 remain the proper statutory tools for limiting scope and validity, while § 101 should not be expanded to bar technological progress.

Outcome. Thus, the ARP vacated the PTAB’s new ground of rejection under § 101 but left intact the existing § 103 obviousness rejection. The decision reaffirmed that claims improving machine-learning efficiency and preserving task performance integrate abstract elements into a patent-eligible practical application.

The Case is Appeal No. 2024-000567.

Judge: Squires, J.,

Attorneys: (Fish & Richardson P.C.) for DeepMind Technologies Limited.

Companies: DeepMind Technologies Limited

Cases: AINews Patent TechnologyInternet USPTO