How to Draft Patent Claims That Survive § 101 from the Start
If you are prosecuting AI or machine learning patents in the United States, the period between mid-2024 and early 2026 has been one of the most consequential stretches of USPTO activity since Alice Corp. v. CLS Bank International was decided in 2014. A new Director has taken over the USPTO with an explicitly pro-patent-eligibility agenda for AI inventions. Precedential decisions have recalibrated how examiners and the Patent Trial and Appeal Board approach AI claims. New examination memoranda have formalised prosecution tools that barely existed two years ago. And the Federal Circuit, in a landmark first-impression decision, has drawn a clear line between AI claims that are eligible and those that are not.
This article maps all of it — in chronological sequence, with the practical implications that matter for prosecution, portfolio strategy, and litigation posture.
Why AI Patent Eligibility Was Broken
To understand what changed, you have to understand what the problem was. Following Alice, the USPTO developed a § 101 examination framework — eventually codified in the 2019 Revised Guidance and MPEP 2106 — that theoretically accommodated AI inventions but, in practice, produced inconsistent and often overbroad rejections. Examiners characterising AI inventions at high levels of abstraction — “the claims are directed to the abstract idea of data analysis” — were routinely sustaining § 101 rejections against applications claiming genuinely novel machine learning architectures, training methodologies, and technical system improvements.
The result was twofold. AI applicants faced prosecution delays and scope-conceding amendments that a rigorous technical analysis did not justify. And the United States risked ceding patent protection for AI innovation to jurisdictions—notably Europe and China—with more accommodating eligibility frameworks. This is the institutional problem that the 2024–2025 changes were designed to address.
July 2024: The Three AI Patent Eligibility Examples
The first major move came in July 2024, when the USPTO published its 2024 Guidance Update on Patent Subject Matter Eligibility, specifically addressing artificial intelligence. Alongside a clarified framework for applying Step 2A of the Alice/Mayo analysis to AI inventions, the guidance introduced three new subject matter eligibility examples — Examples 47, 48, and 49 — that remain the most practically useful reference documents in AI patent prosecution today.
Example 47
Example 47 addressed an artificial neural network for anomaly detection in computer network traffic. The eligible claims in this example specified not only the ANN architecture but also post-detection steps — identifying malicious activity and executing remedial measures to block it. The USPTO found these claims eligible at Step 2A Prong Two because the complete claim, read as a whole, improved a specific technical field (network security) through a defined technological mechanism. The ineligible claims in the same example were broader, describing the anomaly detection function without tying it to a specific technical implementation or concrete follow-on action.
Example 48
involved an AI-based method for separating desired speech signals from background noise. The eligible claims specified that the separated audio components fed directly into a real-time speech recognition system to improve voice command accuracy in hands-free environments — a tangible, technically grounded outcome that the USPTO distinguished from mere data processing. The guidance was explicit: improving the accuracy and functionality of a real-time technical system is a practical application that satisfies Prong Two.
Example 49
addressed AI-assisted personalized medical treatment—a method for generating and administering a treatment plan using a trained machine learning model based on individual patient data. The eligible claims were grounded in a complete technical workflow with defined inputs, a specific model architecture, and actionable outputs tied to clinical administration. The ineligible claims described the same goal at a level of abstraction that could be read on any computerized treatment recommendation, without the technical specificity that distinguishes a patentable implementation.
Taken together, the three examples articulated a consistent principle: AI claims that specify the technical mechanism, anchor the machine learning function to a concrete improvement in a defined technical field, and include meaningful post-analysis steps—rather than merely describing a desired outcome—integrate the judicial exception into a practical application under Prong Two and are eligible. The examples also clarified three conditions the specification must satisfy: it must limit the AI concept to a particular field of use, provide a technical explanation of how the invention improves technology, and support non-abstract claim limitations reflecting that improvement.
These examples are, in practice, more persuasive in prosecution than most Federal Circuit opinions. Examiners are trained on them. Arguments mapping your claims to Examples 47–49 directly address the framework examiners use.
April 2025: The Federal Circuit Draws the Machine Learning Line
While the USPTO was moving toward greater accommodation of AI inventions, the Federal Circuit issued its first substantive ruling on machine learning and § 101, and it was not a green light.
In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), the court affirmed the invalidation of four patents directed to using machine learning models to optimize television broadcast schedules and generate network maps assigning programming to TV markets. The patents described applying established ML methods—gradient descent training, neural network architectures—to scheduling and viewership maximization data. They did not describe improvements to the ML models themselves.
Writing for the panel, Judge Dyk framed the question with precision that every AI patent drafter should memorise: “whether claims that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent eligible.” The court’s answer was no.
Several aspects of the holding deserve close attention. First, the court acknowledged that machine learning is “a burgeoning and increasingly important field” that “may lead to patent-eligible improvements in technology.” The holding is deliberately narrow: generic ML applied to a new domain is ineligible. ML that genuinely improves the model architecture, training methodology, or system-level performance is a different question — and the court explicitly left it open. Second, the court rejected the argument that performing human scheduling tasks with greater speed and efficiency through ML renders claims eligible. Efficiency gains from using computers, without technical improvement to the computer or the algorithm itself, do not satisfy § 101. Third, the Supreme Court declined in December 2025 to grant Recentive’s certiorari petition, leaving the Federal Circuit’s holding as the controlling authority.
For practitioners, Recentive Analytics does not close the door on ML patents. It closes the door on a specific type of claim—the “apply ML to domain X” claim—that describes a business application of generic technology rather than a technical improvement. The cases discussed in the following sections show what the door that remains open looks like.
September 2025: Ex Parte Desjardins and the Institutional Signal
In September 2025, USPTO Director John Squires—who had taken office just days earlier—participated in an Appeals Review Panel decision that sent the clearest institutional message yet about the USPTO’s direction on AI patent eligibility.
In Ex Parte Desjardins, the invention at issue was a method of training machine learning models on sequential tasks without suffering “catastrophic forgetting”—the well-known phenomenon where an ML model trained on a new task loses accuracy on previously learned tasks. The original PTAB panel had entered a new ground of § 101 rejection. The ARP, composed of Director Squires, Acting Commissioner Valencia Martin Wallace, and PTAB Judge Michael Kim, vacated that rejection in a decision that was immediately designated precedential.
The ARP’s reasoning went beyond the technical facts of the case. The panel stated directly that “categorically excluding AI innovations from patent protection in the United States jeopardizes America’s leadership in this critical emerging technology.” It further specified that §§ 102, 103, and 112 — not § 101 — are the appropriate tools for limiting patent protection to its proper scope. Applied to the Desjardins claims specifically, the ARP held that a machine learning method that reduces storage requirements and prevents catastrophic forgetting during continual learning constitutes a patent-eligible technological improvement, not an abstract idea.
The decision’s significance cannot be overstated. An ARP decision including the USPTO Director carries significant weight across the examining corps and the PTAB. The message to examiners was unmistakeable: overbroad § 101 rejections that categorically exclude AI innovations are not acceptable practice. §§ 102 and 103 are where the patentability of AI claims should primarily be tested. Section 101 is not a catch-all rejection mechanism for technology the examiner finds hard to evaluate.
December 2025: Three Memoranda, New SMED Framework, Revised Inventorship Guidance
In December 2025, the USPTO issued a sequence of formal guidance documents that together constitute the most comprehensive recalibration of AI patent examination practice since 2019.
The SMED Memoranda
Three memoranda addressed Subject Matter Eligibility Declarations — a prosecution tool that existed informally but had never been formalised with examination-side obligations. The first memorandum, directed to examiners, established that when an applicant submits a timely, properly formatted SMED, the examiner must assess it under a preponderance of the evidence standard and explain in the next office action how the evidence affects the § 101 analysis. An examiner can no longer simply ignore objective technical evidence submitted by the applicant. The second memorandum, directed to practitioners, provided best practices for SMED submission — including filing as a standalone document focused solely on § 101, submitting before prosecution closes, and grounding the declaration in objective technical evidence such as performance benchmarks, expert declarations, or trade publications documenting the technical problem and the invention’s departure from it. The third memorandum updated the MPEP at §§ 2106.04 and 2106.05 to incorporate Ex Parte Desjardins as a precedential example and to clarify the Prong Two analysis for AI inventions.
For prosecution strategy, the SMED framework represents a meaningful shift. Where previously a § 101 response was purely a legal argument, it can now incorporate technical evidence — inventor declarations, performance data, expert testimony — that creates a factual record supporting eligibility. However, practitioners should be aware that SMED content becomes prosecution history and can be used for claim construction and estoppel arguments in later litigation. Draft SMED declarations with the same care you would apply to expert reports in district court.
The Revised AI Inventorship Guidance
Published in the Federal Register on November 28, 2025, the revised inventorship guidance rescinded the February 2024 Biden-era guidance and replaced it with a framework that explicitly implements Executive Order 14179, “Removing Barriers to American Leadership in Artificial Intelligence.” Several points from the new guidance require direct attention.
The Pannu factors—the three-part test for determining whether a natural person qualifies as an inventor — no longer apply when evaluating inventorship in AI-assisted development contexts. The guidance explains that Pannu governs joint inventorship disputes between multiple natural persons, and since AI systems are not persons and cannot be joint inventors, the Pannu analysis is simply inapplicable. Only natural persons can be inventors, regardless of how extensively AI tools were used in the inventive process.
The practical standard—which has not changed substantively from prior law—is that a human must have conceived the claimed invention. The guidance frames AI tools as analogous to laboratory equipment or software: they assist in the process but do not supply the required human conception. A practitioner who uses an AI system for modelling, simulation, code generation, or prior art analysis does not compromise the human inventor’s status, provided that the human directed, selected, and refined the AI-generated outputs to arrive at the specific claimed solution.
Documentation matters here. The guidance implicitly supports maintaining clear inventor records that show how the human inventor shaped and directed the development process—what the problem was, how the human defined the parameters of the solution, and how the human evaluated and selected among AI-generated alternatives. These records become important if inventorship is later challenged.
Director Squires' Three Pillars: A Framework Worth Understanding
At the 2025 AIPLA Annual Meeting, Director Squires articulated what he called the “Three Pillars” of patent eligibility—a framework he indicated would guide forthcoming formal guidance updates.
The first pillar draws on 35 U.S.C. § 100(b)’s definition of “process,” which includes “new use of a known process, machine, manufacture, composition of matter, or material.” Squires argued that applying an AI system as a tool to achieve a genuinely new use — even of a known ML methodology — has a statutory basis for eligibility that prior examination practice undervalued.
The second pillar derives from Enfish v. Microsoft and its progeny, recognizing that improvements to computer architecture, data structures, and system functionality—including improvements to how AI systems process or represent information—constitute patent-eligible subject matter. AI-specific architectural innovations, novel training methodologies, or model structures that demonstrably improve how the underlying technology operates fall within this pillar.
The third pillar addresses the “something more” standard from Alice and Mayo—interpreting it as encompassing claims directed to changes in system architecture that modify the flow of information, rather than merely targeting an end result. Under this reading, the emphasis is on what the invention does to the system, not what the system does for the business.
Whether these three pillars become formal MPEP guidance in 2026 remains to be seen. As an articulation of the institutional direction, however, they provide a useful drafting and prosecution checklist — particularly for AI applications where the technical improvement is architectural rather than purely algorithmic.
The Fault Line That Practitioners Must Navigate
Two years of institutional activity at the USPTO have moved the examination environment meaningfully toward AI applicants. But the Federal Circuit — the court that actually decides § 101 disputes in litigation — has not moved at the same pace. Recentive Analytics in April 2025 and GoTV Streaming v. Netflix in February 2026 both sustained § 101 invalidity findings against technology claims that lacked specific grounding in technical improvement. The Federal Circuit has not held that AI inventions receive any different treatment under Alice than other technology.
The practical consequence is a two-level compliance obligation. For USPTO prosecution, the current institutional climate supports more aggressive eligibility arguments, earlier allowance on AI claims with specific technical improvement language, and the strategic use of SMEDs to build a factual record. For litigation and post-grant proceedings, the Federal Circuit’s standard remains what it has always been: claims must be directed to a specific technological improvement, not to the abstract application of known methods to new data or business domains. The drafting discipline described in this series — technical mechanism framing, specific architectural detail, problem-solution specification structure — is calibrated to satisfy both levels simultaneously.
What This Means for Your AI Patent Strategy Right Now
The window that Director Squires has created is real and worth using. AI applications drafted with genuine technical specificity—particularly those claiming novel ML model architectures, training methodologies that solve defined technical problems such as catastrophic forgetting or computational efficiency constraints, or AI systems integrated into concrete technical workflows with measurable performance improvements—are finding a more receptive examination environment in 2026 than at any point since Alice.
Track 1 prioritised examination for AI applications is worth serious consideration: getting claims before an examiner under the current institutional guidance, rather than waiting for a possible future shift, has strategic value. Applications whose specifications describe technical problems and improvements in concrete terms — with measurable benchmarks, specific architectural departures from prior art, and clear non-abstract claim limitations—are precisely the applications that the SMED framework and the Prong Two analysis currently favor.
The AI patent landscape is not settled. The Federal Circuit will continue to refine Recentive Analytics, formal MPEP guidance from Director Squires may yet arrive, and Congress has been discussing § 101 reform for years. But for practitioners working with AI patent portfolios today, the direction is clearer than it has been in a decade: draft specifically, claim technically, and document the problem you are solving.
About the Author: [Your Name] is a registered patent attorney practising before the USPTO, with a focus on software, artificial intelligence, and technology patent prosecution and portfolio strategy.
Protect Your Innovation
Schedule your consultation with our expert IP attorneys to secure your assets and scale your brand globally. Partner with us to build a lasting legacy.
Disclaimer
This article is for general informational purposes only and does not constitute legal advice. Please consult a registered patent attorney for advice specific to your situation.