In a rare procedural move, the United States Patent and Trademark Office’s (USPTO) director-convened Appeals Review Panel (ARP) recently vacated a § 101 rejection that had been introduced by the Patent Trial and Appeal Board (PTAB) at the appeal stage. The ARP applied the established Alice/Mayo framework, overturned the § 101 rejection and left the remaining § 103 rejections intact. This decision, when viewed alongside the USPTO’s August guidance to examiners, signals a shift in how the USPTO – under its current leadership – intends to approach subject matter eligibility for artificial intelligence (AI)- and software-related inventions. The emphasis appears to be moving toward careful, technology-focused analyses rather than broad categorical exclusions.
Background
The claims of the patent application in Ex parte Desjardins involved a machine learning application designed to train models on sequential tasks by identifying parameter importance and optimizing with a penalty term – allowing the model to learn new tasks while retaining performance on prior ones. During appeal, the PTAB had affirmed a § 103 rejection and added a new § 101 rejection. After rehearing was denied, the director convened the ARP sua sponte, which has occurred only twice since the ARP was established in 2023.
ARP Decision
Upon review, the ARP applied the established Alice/Mayo framework. At Step 2A, Prong One, the ARP acknowledged that the claims “recite” a mathematical concept (“computing … an approximation of a posterior distribution”) and therefore include an abstract idea.
At Step 2A, Prong Two, the ARP evaluated the claim as a whole and overturned the PTAB’s decision. The ARP found that the additional elements integrated the abstract idea into a practical application. Specifically, the claimed training method improves the operation of the model itself by enabling learning of a new task while preserving performance on a previously learned task, directly addressing the technical problem of “catastrophic forgetting.” The specification’s disclosure regarding reduced storage and lower system complexity reinforced that the improvement is technological, not aspirational. In vacating only the new § 101 rejection and leaving the § 103 rejection undisturbed, the ARP criticized the prior decision’s tendency to over‑generalize machine learning as mere “algorithms” running on generic computers. The decision also underscored that §§ 102, 103 and 112 – not § 101 – should remain the primary mechanisms for properly defining claim scope. This application of Step 2A supports the position that software can provide nonabstract technological advancements when the claims and supporting disclosure substantiate those improvements.
USPTO Memorandum
The ARP’s reasoning further aligns with the USPTO’s Aug. 4 memorandum to examiners on subject matter eligibility and provides practitioners with clear, actionable guidance. The memo instructs examiners to differentiate between claims that “recite” a judicial exception (e.g., naming a particular algorithm) from those that merely “involve” one. This is a critical distinction at Step 2A, Prong One. Additionally, the memo warns against broadening the “mental process” category to cover what cannot practically be done in the human mind and places renewed emphasis on evaluating the claim as a whole at Prong Two, including whether the specification demonstrates a technological improvement.
The ARP mirrored that approach; it acknowledged the mathematical concept at Prong One, then gave weight to the specification‑supported improvement in model operation at Prong Two. For applicants, this alignment between an adjudicative decision and examination guidance signals a more consistent, predictable path for AI and software claims. For further details and practical insights of the memo, see “Navigating Patent Eligibility in the Age of AI: Strategic Insights from the USPTO’s August 2025 Guidance” on JD Supra.
Moving Forward
Looking ahead, the use of the ARP sends a clear message that the USPTO intends to rein in broad, categorical applications of § 101 and redirect the focus of eligibility toward whether the claim, when viewed in its entirety, demonstrates a concrete technological improvement. In examination, applicants should anticipate more attention to the memo’s “recites versus involves” distinction, a narrower view of “mental processes” and a stronger emphasis on the claim‑as‑a‑whole analysis at Step 2A, Prong Two. In addition, as the ARP expressly noted, substantive discussions are likely to shift to §§ 102, 103 and 112, particularly in fast‑moving AI domains. For practitioners and clients, eligibility determinations are likely to favor claims that are specific and demonstrate technical impact.
Practice Takeaways
Given the current trajectory, inventors and companies can respond to these evolving standards by translating the policy signals into specific drafting and prosecution strategies. An effective specification will convey how the claimed technologies provide clear system-level or AI model improvements, such as stability during sequential learning, preserved task performance, reduced memory footprint or enhanced latency and throughput. When specific technical gains are clearly supported in the specification (e.g., with architecture diagrams, control flowcharts or objective metrics), the claimed subject matter is more likely to be recognized as a genuine technological advancement.
At Step 2A, Prong One, the distinction between reciting a bare algorithm and claiming the underlying computing architecture, computer logic or data structure becomes especially important. Claims that operationalize these improvements and produce measurable effects are more likely to succeed. A well-developed record that identifies both the technical problem and the technical solution not only strengthens the narrative for Step 2A, Prong Two, but also supports arguments under §§ 102, 103 and 112. For AI technologies, avoiding human-mind metaphors and instead focusing on specific machine operations and structures can be beneficial and aligns with the USPTO memo’s guidance on mental processes.
These principles are also particularly relevant for agentic AI – systems that exhibit advanced autonomy and real‑time, adaptive decision‑making by coordinating multiple models and tools to achieve goals with minimal human intervention. When autonomy and orchestration features are described (e.g., in terms of controller components, memory or knowledge mechanisms, or resource management techniques that deliver measurable technical effects), the claims can be framed as concrete improvements to computer technology rather than abstract ideas. The recent ARP decision suggests that such improvements will be given meaningful consideration under § 101. For further insights on protecting agentic AI inventions, see “Agentic AI: A Primer for Patent Practitioners,” in The Patent Lawyer.
Conclusion
The ARP’s decision, together with the USPTO’s August guidance, offers a meaningful procedural and policy signal for AI and software patent applicants. By emphasizing practical application and technological improvement, the USPTO appears to be fostering a more balanced and optimistic environment for subject matter eligibility. For practitioners, this reinforces the importance of drafting claims and specifications that clearly articulate concrete technical benefits. As the USPTO continues to refine its approach, the outlook for protecting AI innovations, particularly those grounded in real-world technological advancements, appears to be steadily improving.
[View source.]
