AI Patent Eligibility Claims: Can Machine-Learning Inventions Survive Section 101 in 2026?

Direct answer: AI-related claims can qualify, but AI is not a category of patentable subject matter

Under 35 U.S.C. § 101, a claim directed to a machine-learning process is not ineligible simply because it uses artificial intelligence. Section 101 covers new and useful processes, machines, manufactures, and compositions of matter, and its judicial exceptions (laws of nature, natural phenomena, and abstract ideas) are applied through the two-step test announced in Mayo v. Prometheus and refined in Alice Corp. v. CLS Bank. Because a neural network runs on a computer, examiners and courts ask whether the claim recites an abstract idea (most often a mathematical algorithm) and, if so, whether the claim adds an inventive concept, such as a specific technical improvement. A claim aimed at a concrete improvement in medical-imaging reconstruction, for example, has a plausible route through eligibility; a claim aimed at using AI to predict customer behavior, standing alone, often does not.

Also worth reading: What Are the USPTO Subject Matter Eligibility Guidelines for AI Inventions in 2026? · How Does an AI Patent Eligibility Review Work in 2026? · How Should Deepfake Technology Be Drafted for U.S. Patent Eligibility?

So the direct answer in 2026 is a conditional yes. The USPTO's AI-related eligibility guidance, issued July 17, 2024 at 89 Fed. Reg. 58128, did not create a safe harbor, but it did confirm two recognized routes to eligibility: practical applications of abstract ideas, and improvements in the functioning of a computer or another technology. Federal Circuit decisions such as Enfish, Bascom, Thales Visionix, and Recentive Analytics show that claims directed at specific implementations can pass step two. The same line of cases shows that results-oriented phrasing, with no details about how the improvement is achieved, usually fails. The label AI carries no weight in the analysis; the claimed technical improvement carries the weight. 2025 commentary from Bloomberg Law and practitioner analyses describes a field in which AI claims are moving from generous allowance toward closer scrutiny, partly because many applications are drafted at the level of a result rather than a mechanism. An applicant that files without understanding that distinction should expect § 101 rejections, costly amendments, or later vulnerability in post-grant proceedings.

How the two-step eligibility test applies to machine-learning claims

The Alice test has two steps. At step one, the reader asks whether the claim recites a judicial exception. In AI-related claims, abstract ideas most often take the form of mathematical concepts and formulas, methods of organizing human activity (predicting, classifying, optimizing), and mental processes. The USPTO's 2024 guidance gave examples of claims that tend to recite an abstract idea without more: applying a mathematical equation to a data set with no practical application, using AI to predict or analyze a business outcome, and invoking a machine-learning model as a black box that simply produces an answer. These are the patterns examiners and PTAB panelists look for first, and they are the patterns most likely to appear in independent claims drafted by applicants who want maximum breadth.

At step two, the claim must supply an inventive concept. The 2024 guidance grouped eligible routes into improvements to the functioning of a computer or another technology, applications of an abstract idea to a particular machine or industry where the processing produces a technical result, and other meaningful limitations on the abstract idea. A related concept the guidance highlighted, sometimes called specific measurable evidence of difficulty (SMED), helps show that the underlying problem was technically difficult and that the claimed solution produced a measurable technical improvement. Federal Circuit cases illustrate the dividing line. Enfish treated a self-referential database structure as an improvement in computer functionality; McRO and Thales Visionix credited specific rules and tracking techniques that improved computer-generated animation and inertial tracking; Recentive Analytics credited a specific event-calendar recalculation that improved computer functionality. By contrast, IBM v. Zillow held that displaying the results of an analysis was abstract, and SAP America v. InvestPic held that mathematical analysis of financial data was abstract.

For claim drafting, the practical lesson is that every independent claim should identify the specific mechanism by which the technical improvement is achieved, the data that flows through that mechanism, and the technical result produced. A claim that says the system employs AI to adjust parameters, without stating which parameters, how they are adjusted, and what technical condition is improved, gives the reader nothing beyond the algorithm. The same specification language that makes an abstract idea eligible can also support written description under § 112(a), so careful eligibility drafting and careful disclosure drafting are often the same work done once.

Why AI claims fail: recurring failure patterns

The most common failure pattern is reciting the output of a model rather than the mechanism that produces it. Claims that state a result (identifying an object in an image, ranking results, generating a recommendation) and end there are treated as reciting the mathematical algorithm or the mental process, with a computer added in field position. A second pattern is reciting an abstract goal with a functional instruction (optimize, select, adapt) that does not specify the steps, the constraints, or the technical effect. A third is the data-acquisition claim, which courts have long treated as abstract when it merely collects, analyzes, and displays information, as in Electric Power Group v. Alstom. A fourth is assuming that adding a generic computer or cloud implementation cures an abstract idea, which Alice itself shows it does not.

A different problem often appears alongside § 101: sufficiency of disclosure. An applicant may draft a claim that is eligible in principle but whose specification never explains the steps said to produce the technical improvement. In AI cases that gap is common, because inventors describe a trained model or a result rather than the training data, the model architecture, the inference steps, and the deployment environment. Such a claim can face § 112(a) rejection even after the applicant narrows the abstract-idea language. Ownership raises yet another issue. The USPTO rule effective February 13, 2025 requires that a natural person have contributed to each claim, and the agency will not issue a patent whose inventors are solely an AI author. That inventorship rule is separate from eligibility, but it frequently surfaces in the same prosecution file and the same review.

It is also worth keeping the threshold in view. § 101 eligibility is a floor, not a ceiling. A claim that survives step two can still be anticipated under § 102 or obvious under § 103, and AI claims face a crowded prior-art field of research papers, open-source releases, and product manuals. Many AI applications fail on obviousness because a general method plus one paper supplies the missing motivation. A review that examines only eligibility is incomplete.

What the USPTO changed in 2024 and 2025, and what carried into 2026

The USPTO's July 17, 2024 AI guidance was the first agency-level synthesis of how the Alice framework applies to AI inventions. It withdrew earlier, narrower memoranda, listed categories of abstract ideas relevant to AI, and confirmed the two routes to eligibility discussed above. In 2025, the agency acted in two further areas. On February 13, 2025, the USPTO published its final rule on AI-assisted inventions, applying the patent system's human-inventorship requirement to AI-generated subject matter. Separately, 2025 analyses from firms such as Dykema and write-ups circulated via JD Supra addressed how the USPTO planned to clarify eligibility for AI-related inventions and how that clarification might change prosecution practice. The direction was predictable: generic functional claims face more scrutiny, and specific technical-improvement claims receive better treatment.

One practical tool deserves attention. Under 37 CFR 1.132, an applicant may file an affidavit to traverse a rejection, and the USPTO guidance on AI eligibility describes how evidence of unexpected results and of specific measurable evidence of difficulty can be supplied during prosecution to support a step-two showing. Commentary from Reed Smith has described this use of Rule 132 evidence in AI practice. The tool has limits. A declaration can supply evidence of an effect that the specification already discloses; it cannot add new matter to the application. And evidence created for prosecution may receive less weight in litigation than evidence that was contemporaneous with the filing date.

By September 2026, the framework is stable in outline but unsettled at the edges. The guidance remains guidance; the Federal Circuit's decisions control in court, and the PTAB applies the same law in inter parts reviews. Practitioner coverage, including Bloomberg Law reporting on a shift tied to a single machine-learning case, suggests the ground is moving case by case rather than through a single new rule. The sensible approach for a 2026 filing is to rely on the durable features of the case law, use the USPTO guidance as a drafting template, and expect that future guidance revisions will refine, not replace, the two-step test.

Evidence that the risk is real: studies, cases, and PTAB exposure

Reports published in 2025, including coverage in IPWatchdog on the eve of a USPTO eligibility hearing, described a study finding that AI-related patents face higher rates of § 101 invalidations than other software patents. The study's exact figures depend on how the cohort was defined and how invalidation was counted, so the numbers should be read as directional rather than as a universal invalidation rate for every AI patent. What the finding supports is a simple proposition: the stock of AI patents granted in earlier, more permissive years contains a higher share of claims that later reviewers, whether PTAB panels or district courts, view as abstract. That is a retrospective fact about a filing cohort, not a prediction about well-drafted 2026 applications.

The mechanics of post-grant review make the exposure concrete. An owner facing an inter parts review must live with the statute's one-year final-decision deadline after institution, and software and AI claims are among the most common subjects of § 101 findings in those proceedings. Meanwhile, Federal Circuit and district-court decisions continue to narrow the margins for generic functional claims. Morgan Lewis has described a Federal Circuit dental-related machine-learning decision in which claims were rejected as too generic, and Bloomberg Law has reported that a single machine-learning case shifted the eligibility analysis. Even a favorable PTAB ruling is not precedential, so the same claim can be treated differently in the next review or in court.

The evidence should be read with its limits in mind. A higher invalidation rate does not mean AI inventions are ineligible as a class. The FDA-approved drug, the improved sensor, and the specific inference engine that changes how a computer schedules resources are all still potentially eligible. What the rate figure does mean is that survival in the office is a weak predictor of survival in review, and that a meaningful number of AI patents will be challenged on eligibility grounds. For any patent that matters to a business, eligibility should be stress-tested before an assertion letter is sent, not after a demand letter arrives.

A practical review workflow for AI claims

A useful review begins with a claim chart. For each independent claim, the reviewer writes down the abstract idea, the alleged technical improvement, the specific mechanism that produces the improvement, and the specification paragraphs that support that mechanism. Claims where any of the four cells is empty are the claims at risk. The reviewer then asks whether the improvement can be expressed as a concrete technical result, such as reduced latency, a new accuracy range, a new control loop, or resource savings, rather than a business result such as lower cost or higher engagement. Dependent claims are read in the same way, because they are the usual place where a fallback position is stored after an examiner narrows the independent claim.

The second step is a prior-art search, aimed not just at anticipation but at obviousness combinations that pair a generic machine-learning method with one piece of prior art. The third step is a sufficiency check under § 112(a): does the specification enable and describe the mechanism the claim now recites? If the answer is no, the remedy is usually to narrow the claim to what the specification actually teaches, or to file a continuation that can add the missing detail. The fourth step is deciding when to use evidence. A Rule 132 affidavit can demonstrate unexpected results or measurable difficulty, but only for effects already disclosed in the application. The fifth step is amendment strategy: preserve the broadest defensible independent claim, move breadth into dependent claims, and reserve a position for a continuation filed before a final office action if the first-pass rejection is severe.

The sixth step is calendar discipline. A provisional application preserves a priority date for twelve months, and a U.S. filing made within twelve months of a public disclosure, sale, or use may still receive the benefit of the § 102(b) grace period, but that grace period does not apply abroad. Claims that depend on evidence or data developed after the first filing should be evaluated for a continuation or divisional rather than forced into the original application. The seventh step is documentation. A written review that records the abstract idea, the improvement, the supporting passages, and the amendment options is the record that later supports a defense, a negotiation, or a continuation. That is the kind of record a structured third-party claim review is meant to produce.

Comparing claim strategies: result, pipeline, and hardware coupling

Claim strategy is usually the largest single lever an applicant controls, and the three common approaches differ mainly in where they place the technical detail. The table below compares them on the dimensions that drive § 101 outcomes and downstream cost.

Each option trades breadth for survival. A result-oriented claim is cheap to draft and expensive to defend, because its breadth is also its vulnerability at every stage. A mechanism-specific claim is more work up front, but it converts the abstract idea into a concrete improvement that examiners, PTAB panels, and courts have accepted in analogous cases. A hardware- or process-coupled claim offers the strongest eligibility position and the strongest infringement story, yet it requires a mature product and enough implementation detail to describe. Most mature portfolios use a hybrid: one reasonably broad independent claim, a tier of mechanism-specific dependents, and one or more apparatus claims tied to the deployed system. Whichever option is chosen, the same constraint applies. If the specification does not describe the mechanism, narrowing the claim to it creates a § 112(a) problem instead of a § 101 solution.

Common mistakes that sink otherwise good AI applications

The first mistake is treating AI as the improvement. The words neural, deep learning, or generative do not supply an inventive concept, and the USPTO guidance says so expressly. The second mistake is assuming a grant resolves the question. Allowances are often made without a written § 101 opinion, and a claim that issues can still be invalidated in an inter parts review. The third is confusing eligibility with sufficiency. A narrowed claim that survives Alice may still fail under § 112(a) if the specification never described the narrowed mechanism. The fourth is overlooking the human-inventorship rule, which can invalidate naming and complicate licensing even when eligibility is fine.

The fifth mistake is measuring success by the independent claim's breadth after an examiner forces a narrowing, when the narrowing was so severe that the claim no longer covers the product. The sixth is ignoring the prior art that the abstract idea assumes, particularly when a competitor can point to a research paper plus a general method. The seventh is running a § 101-only review before an office action, because the cheapest time to fix a defective claim is before it is filed, and the next cheapest time is while a response is being drafted. None of these mistakes is exotic. They are the ordinary consequences of treating an AI application as a software template rather than as a technical disclosure that must satisfy a two-step legal test and a sufficiency test at the same time.

When to act, and what a review costs

Timing matters more than most applicants expect. A provisional filing holds a priority date for twelve months, the § 102(b) grace period covers only U.S.-derived disclosures within twelve months, and a patent term runs twenty years from the non-provisional filing. Within the USPTO, Track One aims for a final disposition within twelve months and expedited examination within one year, while an inter parts review must reach a final written decision within one year of institution. The practical message is that the first twelve months determine what evidence and what claim language are available, and eligibility defects that are cheap to fix at drafting become expensive once a reviewer, a panel, or a court is involved.

The second message is that review options range from free to tens of thousands of dollars, and the right level depends on the decision at hand.

These are market estimates, not quotes, and USPTO fees change annually; on the 2025 fee schedule, utility filing fees were roughly $1,290 for a large entity, $645 for a small entity, and $320 for a micro entity, so confirm the current schedule for 2026. Any external review should run under a confidentiality agreement, because the same technical detail that supports eligibility is often the detail a competitor would value most. None of this is legal advice, and the outcome of any single claim depends on its text and its specification. The durable takeaway is simple: define what the technology actually improves, write that improvement into the independent claim, support it in the specification, test it against the prior art, and keep the record. If a concrete improvement cannot be stated, a narrower claim or a different filing strategy will usually serve a product business better than a broad one that will not survive review.