AI Patent Eligibility Under § 101: Key 2025-26 Rulings

Patent eligibility under 35 U.S.C. § 101 remains one of the most consequential thresholds — and strategic inflection points — for AI and machine learning innovators. A clear pattern has crystallized across 2025–2026 Federal Circuit and PTAB decisions: claims that merely apply established machine learning techniques to new tasks or data environments face significant eligibility headwinds, while claims that improve the operation, efficiency, or capabilities of the AI systems themselves are finding greater success at both the USPTO and on appeal. Understanding and operationalizing this distinction is important to building AI patent portfolios that withstand examination, PTAB review, and eventual litigation scrutiny.

Recentive Analytics v. Fox Corp.: The Federal Circuit Draws a Clear Boundary

The Federal Circuit's April 18, 2025 decision in Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), provides the clearest articulation of the current boundary. The court held that patents claiming the application of conventional machine learning techniques to new data environments — here, optimizing television programming schedules and generating network maps — were directed to an abstract idea and lacked an inventive concept under the Alice framework.

As practitioner analysis has noted, iterative model training, prediction generation, and resulting efficiency gains were viewed as inherent to machine learning itself rather than evidence of a technological improvement to the underlying ML technology. The court emphasized that simply using established ML methods in a particular field or for a new purpose, without disclosing improvements to the machine learning models or processes, does not confer eligibility.

This holding reinforces long-standing Federal Circuit precedent that the mere automation or acceleration of human or business processes through generic computing tools remains abstract.

In re McFadden: Reinforcing the Limits on High-Level Abstraction

In April 2026, the Federal Circuit (nonprecedential) affirmed a PTAB rejection in In re McFadden (Fed. Cir. Apr. 7, 2026). The claims involved high-level algorithms for calculating information distribution differences. The court found that instructions directing a standard computer to manipulate, transform, and compare data at a high level of generality "fare no better in the abstract idea analysis than generic software or computing components." Thus, and as legal commentary observed, enhancing abstract calculations on conventional hardware, without a specific technological improvement that defines how the implementation physically or logically alters computer operation, remains ineligible.

Together, Recentive and McFadden establish a firm floor: generic application or high-level abstraction will not suffice.

Ex Parte Desjardins: The PTAB Charts a Path Forward

While the Federal Circuit has defined the outer limits, the PTAB's precedential decision in Ex parte Desjardins, Appeal No. 2024-000567 (PTAB Sept. 26, 2025, designated precedential Nov. 4, 2025), offers a constructive roadmap. An Appeals Review Panel (authored by Director John A. Squires) vacated a § 101 rejection of claims directed to methods for training machine learning models sequentially across multiple tasks while mitigating catastrophic forgetting.

The claims used importance measures (approximations of posterior distributions over parameter values from prior tasks) and a penalty term in the objective function to protect performance on earlier tasks while optimizing for new ones. The specification highlighted concrete benefits: use of a single model instance across tasks, reduced storage requirements, lower system complexity, and preserved prior knowledge.

The Panel held that these limitations reflected an improvement to how the machine learning model itself operates, integrating the abstract idea (mathematical calculations) into a practical application under Alice Step 2A, Prong Two — analogous to Enfish, LLC v. Microsoft Corp.'s recognition that software improvements to computer functionality via logical structures and processes can be patent-eligible. The Panel stressed that §§ 102, 103, and 112 remain the proper tools for policing claim scope and that categorically excluding AI innovations risks undermining U.S. technological leadership.

The USPTO promptly incorporated Desjardins into the MPEP via a December 5, 2025 memorandum, directing examiners and PTAB panels to apply its reasoning to AI-related claims.

A Broader Empirical Review Confirms the Recalibration—and Its Limits

Recent PTAB decisions confirm that the recalibration is structural, not transient, while also showing that it operates entirely within the established two-step framework: the Board continues to apply Alice/Mayo through the lens of the USPTO's 2019 Revised Guidance, but, consistent with Desjardins, it now more readily credits limitations that improve the machine learning system itself. Independent empirical analysis of seven months of PTAB § 101 decisions captures the magnitude of the shift, with the reversal rate for examiner § 101 rejections roughly doubling under Director Squires — rising from the 8–12% range that prevailed through most of 2024 to a mean of approximately 22%, with a November 2025 spike to 29% before stabilizing near 20%.

A separate review of 28 recent appeal decisions addressing § 101 rejections of AI/ML-enabled and other computer-implemented inventions (e.g., neural networks, ensemble models, predictive analytics, and diagnostic algorithms) (citations below) corroborates that figure: the Board affirmed in roughly 21 of 28 decisions (75%), with six full reversals and one mixed outcome, yielding a reversal rate of approximately 21–25%. These decisions do not redefine the abstract-idea categories or announce new tests; they turn almost entirely on the familiar pivot point of Step 2A, Prong Two — whether the claim integrates a recited judicial exception into a practical application, typically by reflecting an improvement to a technology or technical field. The Board consistently infers the focus of the "claimed advance" from the specification rather than crediting attorney characterizations of the claims.

Applicants succeeded primarily by tying claimed limitations to a concrete, specification-supported improvement in computer or technical functionality — precisely the Desjardins path. In Ex parte Kumar, Appeal No. 2025-003688 (PTAB May 13, 2026), the Board reversed because a "configuring a finite rank deep kernel" step provided a practical application, crediting the specification's identification of deep kernel learning as computationally expensive and the claim as a specific manner of improving that process rather than a mere result (citing Enfish and SAP America). In Ex parte Kimura, Appeal No. 2026-000830 (PTAB May 18, 2026), the Board was persuaded that a Logical Neural Network structure for action pruning in reinforcement learning reflected a genuine improvement — that is, allowing AI systems to "effectively learn new tasks in succession whilst protecting knowledge about previous tasks" (over a dissent). In Ex parte Schneidewend, Appeal No. 2026-000916 (PTAB June 16, 2026), the Board reversed where continuously updated ECG learning templates "improve[d] the functioning of the technology itself" by more accurately identifying heartbeats in irregular signals while reducing processing demands (analogizing to CardioNet). The Board reached the same result for dynamically retrained scheduling models (Ex parte Rajagopalan, Appeal No. 2026-001313 (PTAB June 18, 2026)), malware-detection models with a tuned "activation range" that manages the overall false-positive rate of conjoined detection models (Ex parte Rao, Appeal No. 2026-000295 (PTAB May 6, 2026)), and mining-facility simulation monitoring, where it found Prong One dispositive and the claims not directed to an abstract idea at all (Ex parte Gonzalez, Appeal No. 2026-000922 (PTAB June 2, 2026), reversing § 101 while entering new § 103 grounds).

Affirmances, by contrast, tracked Recentive's caution. The Board upheld rejections where claims were directed to an abstract idea — most often a mathematical concept, a mental process, or a method of organizing human activity — with no additional elements supplying an inventive concept or technical improvement beyond invoking generic computing components. In Ex parte Kobayashi, Appeal No. 2026-000535 (PTAB May 19, 2026), the Board affirmed because automated determination of "explanatory variables" recited a mathematical concept, rejecting the argument that a recited formula escaped abstraction and observing that "[a] mathematical relationship may be expressed in words or using mathematical symbols." In Ex parte Dagan, Appeal No. 2026-000561 (PTAB May 14, 2026), the Board agreed that a visual facet search claim, considered as a whole, "applies the abstract idea using generically-recited computers and computing components." And in Ex parte Mathew, Appeal No. 2026-000401 (PTAB May 15, 2026), the Board agreed with the examiner that claims directed to predicting future account events based on past transfers described "certain methods of organizing human activity," finding that "[a]t its core, Appellant's argument is that processing less data reduces processing time," which "does not reflect an improvement in computer technology or functionality" — a similar echo to Recentive. Notably, technical or quantitative subject matter did not guarantee eligibility: in Ex parte Narayan, Appeal No. 2026-000299 (PTAB May 15, 2026), a divided panel affirmed against claims reciting a dual-exponent data-storage format, treating the claimed factoring as a mental/mathematical concept despite a dissent contending the claims described an improved data structure deserving of protection (and that any prior-art concern belonged under §§ 102/103, not § 101).

The mixed decision in Ex parte Landgraf, Appeal No. 2026-000428 (PTAB May 12, 2026) illustrates the dividing line within a single case. The Board sustained the rejection of claims 69–72, whose recitation of a handheld monitoring device with an ECG sensor and audio sensor amounted to "nothing more than mere presolution activity of data gathering," while reversing as to claims 2, 5, 53, 60, and 61 because it was persuaded the examiner erred at Step 2A, Prong One — the specification and a supporting declaration showed that determining a heart's ejection fraction in real time from ECG and audio data was not something that could be performed by the human mind. The common thread across the entire body of decisions is unmistakable and confirms the Desjardins framework: the inventions that prevailed focused on improving how the machine learning system operated, not merely deploying ML to perform an existing task more efficiently.

What This Means for Drafting, Prosecution, and Portfolio Strategy

These decisions translate into concrete priorities for AI innovators and their counsel:

Claim Drafting Priorities. Claims should explicitly recite technical improvements to the AI/ML system itself — specific model architectures or modifications, mechanisms for reducing computational or storage requirements, techniques for handling data drift or catastrophic forgetting, enhanced training efficiency or convergence, improved explainability, reliability, or robustness. As practitioner commentary following Recentive has noted, eligibility hinges not just on the subject matter but on how concretely it is presented in the claim. The Desjardins claims succeeded because they tied a particular mechanism (an importance-weighted penalty term protecting prior task performance) to concrete, measurable benefits, and the reversals above succeeded on the same logic — a specific LNN inference structure, continuously updated ECG templates, a tuned activation range. Generic "apply ML to X" language remains vulnerable under Recentive, as the affirmance in Ex parte Mathew, Appeal No. 2026-000401 (PTAB May 15, 2026) confirms.

Ensure Strong Specification Support. Robust disclosures — architectural diagrams, flowcharts, pseudocode, training algorithms with sufficient detail, performance benchmarks comparing the claimed approach to baselines, and embodiments demonstrating the technical mechanism and its effects — create the evidentiary foundation both for overcoming examiner rejections and for defending validity later. Desjardins and the reversals above rewarded clear linkage between the claimed technique and improvements to model operation; Recentive, McFadden, and the affirmances penalized its absence.

Prosecution Tools. The Subject Matter Eligibility Declaration (SMED) procedure emphasized in December 2025 USPTO guidance provides a powerful, underutilized vehicle. A standalone Rule 132 declaration can supply objective evidence — performance metrics, technical implementation details, or expert testimony on why the claimed improvement is not practically performable by the human mind or a generic computer — that examiners must consider before maintaining a § 101 rejection. Ex parte Landgraf, Appeal No. 2026-000428 (PTAB May 12, 2026) illustrates the point: a declaration establishing that real-time ejection-fraction determination could not be performed mentally was decisive in securing reversal on the relevant claims.

Portfolio Strategy. Continuation applications remain valuable to preserve flexibility as standards evolve. Trade secrets may be preferable for proprietary datasets, detailed training methodologies, or implementation specifics that are difficult to reverse-engineer from a published patent. Hybrid approaches — patenting core architectural or methodological improvements while protecting supporting data and know-how — are increasingly common.

The Central Strategic Tension. A meaningful gap persists between what the USPTO (particularly post-Desjardins and under current Director guidance) is likely to allow and what the Federal Circuit will sustain if the patent reaches litigation. The recurrence of dissents in close PTAB cases (e.g., Ex parte Kimura, Appeal No. 2026-000830 (PTAB May 18, 2026) and Ex parte Narayan, Appeal No. 2026-000299 (PTAB May 15, 2026)) signals that even within the Board the practical-application inquiry remains contested. Practitioners who build portfolios that satisfy the demands of both forums — detailed technical improvement disclosure in the specification, claims that recite specific mechanisms rather than results or business outcomes, and strategic use of continuations and SMEDs — are best positioned to maximize value and minimize risk.

The practitioners and innovators who internalize both Recentive's caution and Desjardins' opportunity — and draft specifications and claims accordingly — will be more likely to receive durable protection for AI innovations.

Navigating Patent Eligibility for AI and Machine Learning Innovations

Practical Lessons from Recent Federal Circuit and PTAB Decisions Under 35 U.S.C. § 101

Recentive Analytics v. Fox Corp.: The Federal Circuit Draws a Clear Boundary

In re McFadden: Reinforcing the Limits on High-Level Abstraction

Ex Parte Desjardins: The PTAB Charts a Path Forward

A Broader Empirical Review Confirms the Recalibration—and Its Limits

What This Means for Drafting, Prosecution, and Portfolio Strategy