What are the AI patent eligibility requirements under current USPTO guidance in 2026?

The Current State of AI Patent Eligibility in 2026

Patent eligibility for artificial intelligence inventions in the United States operates inside the same Section 101 framework that governs every other category of patentable subject matter, but the practical experience of applying that framework to machine-learning systems has shifted noticeably since 2024. The USPTO has issued a series of memoranda through 2025 and into 2026 that attempt to clarify how examiners should evaluate claims reciting neural networks, training methods, and AI-implemented processes. The two-step Alice/Mayo analysis remains the operative test, and the agency has not abandoned the abstract-idea screening that has defined eligibility doctrine since the Supreme Court's 2014 Alice Corp. decision. What has changed is the agency's posture toward the second prong of that test, particularly the way examiners weigh "additional features" that go beyond purely mathematical modeling.

Also worth reading: How do I build an AI patent enablement compliance checklist that actually satisfies patent office requirements? · What are the AI patent disclosure requirements in 2026 and how do they impact patent filings? · What are the exact requirements for filing a CFR 1.137 revival petition to recover an abandoned patent application?

The October 2024 USPTO memorandum on AI patent eligibility was followed by additional guidance in 2025 and early 2026, including a revision to how Rule 132 secondary considerations evidence ("SMED" declarations) is treated when applicants seek to demonstrate that a claim recites an inventive concept. These memoranda did not rewrite the statute, but they did give examiners more permissive instructions about when AI claims directed to training models, optimizing loss functions, or applying trained networks to specific technical fields can satisfy Step Two of the Alice test.

How the Two-Step Eligibility Test Applies to AI Claims

Under the Alice/Mayo framework, examiners evaluate every patent claim, including AI claims, in two steps. Step One asks whether the claim is "directed to" one of the three judicial exceptions: laws of nature, natural phenomena, or abstract ideas. For AI inventions, the relevant exception is almost always the abstract-idea category, which courts have repeatedly described as encompassing mental processes, certain methods of organizing human activity, and mathematical concepts. A claim that recites nothing more than "train a neural network on data and output a prediction" will usually be characterized as directed to an abstract idea at Step One, because the steps can largely be performed in the mind or on paper.

Step Two is where most AI claims survive or fail. This step searches for an "inventive concept" sufficient to transform the claim into a patent-eligible application of the abstract idea. Conventional steps performed after the fact, generic computer implementation, and well-understood, routine, and conventional activity will not supply that inventive concept. The USPTO's 2025-2026 guidance tells examiners that certain AI-specific features can qualify as inventive concepts when properly claimed, including architectural limitations tied to a specific model structure, training-time constraints that alter the optimization, and technical applications that produce a non-conventional technological improvement in a stated field of use.

## Why AI Claims Get Rejected and How to Avoid the Common Traps

The single most common reason AI patent applications fail is that the claims are drafted at a level of abstraction that does not survive Step One. When an applicant claims "a method comprising receiving data, processing the data with a machine-learning model, and outputting a result," the examiner is essentially being asked to patent the concept of using AI to analyze data, regardless of what data, what model, or what result. The Federal Circuit has invalidated similar claims in decisions preceding Enfish and after it, and the USPTO has flagged this pattern repeatedly in training materials for its examiners.

A second frequent trap involves treating AI as if it were a "black box" and claiming only inputs and outputs without describing any of the structural or functional specifics of the model itself. The 2025-2026 guidance rewards claims that include details about network architecture, training methodology, data preprocessing, and post-processing steps that integrate the model into a specific technological pipeline. A claim that recites a particular layer configuration, a particular training algorithm with specified hyperparameters, or a particular integration with a sensor or industrial control system will tend to fare better at Step Two than a generic recitation.

Applicants also make the mistake of relying exclusively on arguments about the underlying mathematics being "novel" or "complex." The Supreme Court has been explicit that novelty under Section 102 and Section 103 cannot rescue an otherwise ineligible claim under Section 101. The mathematics of a new training method, no matter how original, is still mathematics, and mathematical concepts remain one of the three patent-ineligible categories.

Practical Steps for Drafting and Prosecuting an AI Patent Application

The first practical step is to draft claims at multiple levels of abstraction. A portfolio should generally include at least one independent claim that recites the broad method of using AI for a purpose, at least one independent claim that recites the specific architecture or training method, and at least one independent claim that recites the AI component as part of a larger technical system, such as an autonomous vehicle controller, a medical imaging workstation, or a manufacturing process controller. This stratified approach gives the applicant fallback positions if the broadest claim is rejected under Section 101.

The second step is to ensure that the specification provides explicit support for the technical improvement that the invention is supposed to produce. Courts and examiners have repeatedly emphasized that an applicant cannot rely on after-the-fact statements about technical effects that do not appear in the original disclosure. The specification should describe baseline systems, the technical problem with those systems, and how the claimed AI invention improves upon them in measurable terms, such as reduced inference time, lower memory consumption, improved detection accuracy, or reduced power usage.

The third step is to prepare a Rule 132 declaration early in prosecution if there is credible evidence of technical effect. The USPTO's clarified treatment of secondary considerations evidence in 2025 and 2026 means that inventor declarations, expert declarations, and comparative test data can carry real weight at Step Two when they are tied to specific claim limitations. Declarations that simply restate the specification tend to be discounted, but declarations that include side-by-side comparisons with prior-art systems and quantify an improvement are treated more seriously.

Comparison of Eligibility Strategies for Different AI Inventions

Different categories of AI inventions face different eligibility headwinds, and the strategy for each must be tailored accordingly. A pure software invention, such as a new method of training a generative model, will usually need to lean on specific architectural and training-step limitations to survive Step Two. An AI invention integrated into a physical device, such as a smart sensor or a robotic manipulator, can rely on the device-side limitations as a source of inventive concept. An AI invention used in a regulated or technical field, such as medical diagnostics or semiconductor manufacturing, can often tie its claims to the technical application and the field-specific data processing that occurs there.

## Common Mistakes That Undermine Eligibility Arguments

Applicants often lose eligibility arguments by overrelying on the Berkheimer v. HP and Enfish v. Microsoft line of cases as if those decisions created a blanket safe harbor for software and AI inventions. In fact, both decisions turned on specific claim language and specific disclosures, and neither insulates a poorly drafted AI claim from a Section 101 rejection. The USPTO's 2025-2026 guidance repeatedly reminds examiners that Berkheimer is a factual inquiry about whether the additional claim elements represent an inventive concept, not a presumption that any software recitation qualifies.

Another common mistake is the failure to address the examiner's specific rationale for rejection. When an examiner rejects an AI claim as directed to the abstract idea of "using a computer to perform mental steps," the applicant's response should explain why the specific combination of recited features is not merely a generic computer implementation of a mental process. Generic arguments about AI being a "transformative technology" tend to be ineffective because they do not address the claim language at issue.

A third mistake is ignoring Section 101 until the examiner raises it. By the time the examiner issues a first Office Action with a Section 101 rejection, the applicant has often already spent considerable time on Sections 102, 103, and 112. A pre-emptive eligibility analysis during drafting, including a claim chart comparing the draft claims to the closest analogous Federal Circuit decisions, can save prosecution time and reduce the likelihood of an adverse final rejection.

Timing, Costs, and When to Act

The USPTO's current fee schedule places the basic filing fee for a small entity at roughly $730 and a micro entity at roughly $365, with the examination fee bringing the entry cost to approximately $1,600 to $2,200 depending on entity status. AI inventions are not assigned to any special fee category, but their examination tends to take longer than non-AI applications because of the eligibility complexity. The current average first-action pendency for applications in technology centers that handle AI inventions is approximately 18 to 24 months as of mid-2026, compared with the agency-wide average of roughly 16 to 20 months.

Applicants should treat AI patent strategy as a portfolio decision rather than a single-application decision. A typical AI patent budget should account for at least two continuation applications, one or more response extensions, and the cost of preparing technical declarations to support eligibility. For inventors with limited budgets, a provisional application followed by a carefully drafted non-provisional within twelve months can be a reasonable starting point, provided the provisional contains enough technical detail to support the eventual non-provisional claims.

Where the Doctrine Is Heading

The USPTO's 2025-2026 guidance has shifted the practical experience of AI patent prosecution in a more permissive direction, but it has not solved the underlying tension between abstract-idea doctrine and the realities of modern software and AI innovation. Congressional proposals to reform Section 101, including the Patent Eligibility Restoration Act and its successors, have continued to surface in both chambers of Congress through 2025 and 2026, with Senate Judiciary hearings in 2025 featuring testimony from industry groups including the CCIA. None of these proposals had been enacted into law as of September 2026, but the continued legislative interest signals that the eligibility framework may eventually be rewritten by statute rather than agency guidance.

Until then, applicants should expect eligibility analysis to remain the single largest source of uncertainty in AI patent prosecution. Drafting claims that survive Alice/Mayo is achievable but requires more care than drafting claims for many other categories of technology, and it requires the applicant to think about eligibility as an integral part of claim drafting rather than as a hurdle to be cleared later.