Imagine you have built a model that dramatically reduces memory overhead during continual learning by selectively preserving parameter weights from prior tasks. The technical performance gain is real and measurable. But when you file the patent application, the examiner rejects the claims as ineligible for being directed to an abstract mathematical concept. The specification never explained what the system was actually solving at the hardware level or why the architecture choices were technically necessary. The invention was real. The drafting was the problem.
That scenario plays out more often than it should. The past twelve months have produced a clearer roadmap for practitioners willing to study what the USPTO and courts are actually rewarding. Drawing on key 2025 developments—and early 2026 insights—this article outlines actionable strategies to strengthen machine learning (ML) patent applications from the start.
The Specification Is Where Applications Win or Lose
The most durable guidance for ML patent drafting remains straightforward: the specification must give a person of ordinary skill in the art enough architectural, training, and data-handling detail to reproduce the invention without undue experimentation. The Intellectual Property Owners (IPO)'s AI Patenting Handbook V3.0, published in December 2025, stresses that practitioners must disclose sufficient algorithmic, architectural, and training detail to support broad ML claims, and that broad claims filed without this foundation have consistently struggled at both the PTAB and the Federal Circuit.
The EPO's 2025 examination guidelines, as analyzed by patent law experts, retained language from Board of Appeal decisions warning that an AI patent disclosure is insufficient when the mathematical methods and training datasets are disclosed in insufficient detail for the skilled person to reproduce the technical effect over the whole scope of the claim. The EPO characterized such deficient disclosures as "an invitation to a research programme," specifically in the context of insufficiency under Article 83 EPC. While not binding on the USPTO, practitioners filing globally should treat it as an early signal about the direction of travel. For additional context, the EPO Guidelines Preview by industry commentators (Feb. 2025) overviews these updates, confirming alignment with computer-implemented invention (CII) rules and emphasizing reproducibility for ML innovations.
Claim Around the Technology, Not Around the Result
The IPO's AI Patenting Handbook frames this in terms of anchoring claims in how concrete technical improvements—such as reduced latency, improved resource utilization, or enhanced model robustness—are achieved, rather than in generic references to "machine learning." This distinction matters enormously under the Alice/Mayo framework, and the Federal Circuit clarified exactly why in April 2025.
In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), the court addressed a question of first impression: whether claims that do no more than apply established methods of machine learning to a new data environment are patent eligible. The answer was no. As legal commentators summarized, the court found that Recentive's patents relied on conventional machine learning techniques and generic computing equipment, and that the claims failed to "delineate steps through which the machine learning technology achieves an improvement." Applying ML to event scheduling and network map creation was not enough. The court held that iterative training and dynamic model adjustment are incidental to the nature of machine learning itself and do not constitute a technological improvement.
The practical takeaway, as industry analysts noted, is that eligibility hinges on disclosing concrete, technical improvements regardless of whether claims are directed to training processes or the application of AI outputs. Even training claims fail if the techniques used are routine or described in abstract terms. Early 2026 analyses, such as patent firm reviews, emphasize the need for "model-specific" improvements to avoid pitfalls, while other legal publications describe Recentive as a "precedential case of first impression" that could chill overly broad AI claims. Similarly, IP commentary highlights how iterative training alone isn't inventive, underscoring the importance of specificity.
What Ex Parte Desjardins Changed
The most consequential development for ML patent prosecution in 2025 was the Appeals Review Panel (ARP) decision in Ex Parte Desjardins, Appeal No. 2024-000567 (PTAB Sept. 26, 2025), designated precedential on November 4, 2025. The application, filed by Google, involved a continual learning method addressing catastrophic forgetting—the loss of previously learned information when a model is trained on new tasks.
The ARP vacated a §101 rejection and found that the PTAB had evaluated the claims at too high a level of generality. As legal experts explained, the ARP credited the claimed training techniques for reducing storage requirements, lowering system complexity, and enabling effective continual learning, finding that these improvements integrated any abstract mathematical concept into a practical application. The ARP also reaffirmed that §§102, 103, and 112 are the traditional and appropriate tools for limiting patent protection, and that §101 should not be used to categorically exclude AI innovations.
As patent blogs reported, Director Squires' precedential designation formally embedded this reasoning into USPTO policy. The USPTO then incorporated Desjardins into the MPEP through a December 5, 2025 memorandum, highlighting the case as an example of claims that solve a specific technological problem through concrete training techniques that improve performance and reduce system complexity. The USPTO's memo formalizes the changes based on Desjardins into MPEP §§2106.04(d) and 2106.05, providing examiners with examples of eligible AI claims focused on improved continual learning. IP industry analysis discusses how this, alongside Director Squires' AIPLA remarks, signals broader eligibility but maintains tensions with Federal Circuit scrutiny.
Comparing Recentive and Desjardins illustrates key drafting lessons: In Recentive, generic application of ML to new data failed for lacking delineated improvements; in Desjardins, specific techniques mitigating catastrophic forgetting succeeded by demonstrating practical integration (e.g., reduced storage). Where possible, practitioners should quantify such gains—e.g., "reducing memory usage by 40% via selective parameter preservation"—as one way to bridge this gap.
The August 2025 Examiner Memo
Before Desjardins, Deputy Commissioner Charles Kim issued a memorandum on August 4, 2025, directed to Technology Centers 2100, 2600, and 3600. The memo clarified subject matter eligibility examination (Step 2A) by limiting the mental process category to what can practically be performed in the human mind, stating expressly that claim limitations encompassing AI in ways that cannot be practically performed in the human mind do not fall within that grouping. The memo also instructed examiners to analyze claims as a whole under Prong Two rather than treating additional elements in isolation, and established that a subject matter eligibility rejection should be made only when it is more likely than not that a claim is ineligible.
As legal observers noted, the Kim memo provides practitioners with authoritative language to rebut overly expansive §101 rejections, and its guidance is useful both for drafting patent applications and preparing responses to examiner rejections. Legal news overviews note the memo's potential to ease AI rejections by emphasizing holistic claim review, offering strategies for applicants in ongoing prosecutions.
The SMED as a Prosecution Tool
One underused lever that received formal attention in December 2025 is the Subject Matter Eligibility Declaration under 37 C.F.R. §1.132. On December 4, 2025, Director Squires issued two memoranda expressly encouraging applicants to use expert declarations to overcome §101 rejections when specifications lack sufficient technical improvement language.
As IP analysts reviewed, the USPTO confirmed that improvements to model performance, memory, data structures, and system architecture can provide the "something more" required under Alice, and that applicants may submit focused SMEDs under Rule 132. Examiners must weigh such declarations under a preponderance standard. The guidance also clarified that a specification need not explicitly label an invention as an "improvement," provided the improvement would be apparent to a person of ordinary skill in the art. For example, a declaration could evidence reduced latency in an ML model by detailing benchmark comparisons, as suggested in the memos.
The IP analysis noted the SMED cannot cure disclosure gaps in the original specification. It can substantiate how a person of ordinary skill in the art would read a specification, but it cannot supply information that was required to be present at filing. The practical implication: the decision whether a SMED may be helpful should be evaluated at the drafting stage, not after prosecution has run into trouble. To streamline prosecution, consider using SMEDs proactively in continuations or when addressing potential §101 issues in complex ML inventions.
Training Claims Deserve Separate Attention
Practitioners sometimes treat training claims as secondary to inference claims, but they raise distinct drafting challenges. The IPO's AI Patenting Handbook devotes dedicated discussion to training-related claims and §112 issues, noting common written description and definiteness challenges when applications fail to disclose sufficient detail about the training phase. Specialized ML modules described in functional terms without corresponding algorithmic structure can trigger §112(f) concerns, requiring the specification to provide algorithmic structure, input and output descriptions, and handling of exceptions that the claim language implies.
In recent trends, training claims often survived §101 scrutiny (as in Desjardins) but faced heightened §112 challenges if functional without alternatives like specific loss functions or hyperparameters. To broaden scope while mitigating risks, disclose variations—e.g., different optimizers or data augmentation techniques—and quantify improvements over baselines. Vague specs also invite post-grant invalidations, such as in IPRs at the PTAB, where §112 enables challengers to argue lack of enablement for broad claims.
Likely Eligible ML Claims (e.g., Desjardins-like):
- Specific architecture: "A system for continual learning comprising a neural network with selective parameter freezing to mitigate catastrophic forgetting with memory usage reduced by X% compared to a baseline."
- Quantified improvement: "Training the model using replay buffers and elastic weight consolidation having increased accuracy retention of Y% across tasks."
- Practical integration: "Integrating the trained model into edge devices for real-time inference to provide reduced computational complexity."
Likely Ineligible ML Claims (e.g., Recentive-like):
- Generic application: "A method of applying machine learning to schedule events using iterative training on data."
- Abstract result: "Dynamically adjusting a model for better predictions in a new domain."
- Routine techniques: "Using conventional neural networks and generic hardware to process data."
A Practical Checklist Before Filing
Drawing on the current guidance landscape, the pre-filing questions that are important to consider:
- Does the specification describe the model architecture with enough specificity that a skilled practitioner could implement it?
- Does the specification distinguish the training phase from the inference phase, and provide an inventive contribution for one or both phases?
- Does the specification quantify performance improvements over conventional approaches?
- Does the specification explain why the claimed architecture choices were technically necessary to achieve the invention's benefits, including supplementing any mathematical formulations with descriptions of their practical application and improvements to technology?
- Are the dependent claims drafted with specific technical features that can serve as fallback positions if the independent claims face §101 challenges?
The combination of Recentive, Desjardins, the August 2025 examiner memo, and the December 2025 SMED guidance gives practitioners more tools and clearer standards than they have had in years. Looking ahead, these shifts may reduce subject matter eligibility rejections for well-drafted AI patents, but Federal Circuit cases like Recentive remind us to prioritize specificity. For global filers, the EPO's stable guidelines complement the USPTO's evolution, promoting harmonized disclosures. The practitioners who use these tools at the drafting stage may spend less time defending claims they should not have had to defend in the first place—and better withstand post-grant challenges.