An AI patent eligibility checklist is a structured way to test whether a proposed artificial-intelligence invention will survive examination under 35 U.S.C. § 101 before money is spent on drafting, filing, or foreign filing. It is not a substitute for attorney judgment, and it is not a promise of validity. Eligibility under § 101 is only the first gate: an application must also satisfy the novelty and nonobviousness requirements of §§ 102 and 103, the written-description and enablement standards of § 112, and all filing formalities. A practical checklist works backward from the broadest claim and asks whether each element recites a patent-eligible concept, whether the combination is markedly more inventive than the prior art, and whether the specification and evidence support a concrete technological improvement rather than a result that the law deems unpatentable on its face. Used carefully in 2026, such a checklist helps applicants focus prosecution resources, choose between patent, trade-secret, and publication strategies, and anticipate how a reviewer will read the claim.
What the core eligibility test actually asks
Also worth reading: What Does the 2026 USPTO AI Patent Eligibility Guidance Actually Change for Applicants? · How Do Patent Examiners Evaluate Subject Matter Eligibility for Machine Learning Inventions Under Current 2026 Guidelines? · What are the definitive best practices for drafting AI patent claims in 2026 to survive eligibility challenges?
The governing American framework has been stable since the Supreme Court's 1981 decision in Diamond v. Diehr, 450 U.S. 175, which held that a process using a mathematical formula can be eligible when integrated into an industrial process as a whole. The framework was restated in Bilski v. Kappos, 561 U.S. 593 (2010), Myriad Genetics, Inc. v. Clarke, 569 U.S. 576 (2013), and Mayo Collaborative Services v. Prometheus Laboratories, 566 U.S. 207 (2012). The two-step test the USPTO applies in practice asks, first, whether the claim falls within one of four judicially created exceptions — abstract ideas, natural phenomena, laws of nature, and some categories of natural products such as DNA — and, second, if it does, whether the claim as a whole is markedly more inventive than anything found in the prior art. The USPTO's 2019 Revised Patent Subject Matter Eligibility Guidance, published at 84 Fed. Reg. 50 on January 7, 2019, remains the primary examination guidance, and it instructs examiners to look at the claim as a whole rather than dissecting it into ineligible components. A checklist that omits this whole-claim step is incomplete, because the same element that looks abstract in isolation can be part of an eligible combination.
How the two-step framework changed under Alice and Berkheimer
The 2014 decision in Alice Corporation v. CLS Bank, 573 U.S. 208, added the operative language for nearly every AI eligibility analysis today: an ineligible concept becomes eligible only if the claim adds an "exception" that applies the concept in a nonconventional way, and the elements collectively amount to "significantly more" than the underlying exception. In 2026 checklists, that second prong is written as a markedly inventive leap over the prior art, evaluated against what a skilled artisan could already do. The 2018 decision in Berkheimer v. HP Inc., 881 U.S. 19, tightened the boundary by making eligibility a question of law for courts while leaving factual questions of ordinary skill and conventionality to a factfinder; in practice, claims that merely recite a generic computer with the abstract concept stapled on it tend to fail both Alice prongs. Checklists that treat the generic-computer step as a box to tick rather than a genuine technical question will misclassify many applications, because examiners and courts ask whether the claimed improvement is in the computer technology itself or merely in the math or data being processed. Berkeley-era guidance from the USPTO, including the 2019 memorandum on artificial intelligence, reinforces that focus by telling examiners to look for technical improvements rather than field-of-use limitations.
Applying the test to machine-learning inventions
Machine-learning inventions force the checklist to be specific rather than generic. A neural-network claim that merely trains a model to predict an outcome, without specifying the architecture, the data features, the loss function, or the improved technical result, usually reads as reciting a mathematical relationship plus a generic processor. Stronger claims identify a particular training technique, a memory or latency improvement, a specific hardware configuration, or an unusual data pipeline that cannot be performed by conventional means. For example, a claim directed to how a federated-learning system reduces communication overhead by selectively updating model shards on edge devices can be framed as an improvement in how the computer operates, while a claim directed to "classifying images using a neural network" usually is not. Patent offices outside the United States apply related but not identical tests: the European Patent Office's technical-effect approach, embodied in the EPO's Guidelines G-II 3.3.1 and related problem–solution methodology, often routes around the American abstract-idea exceptions entirely, and China's Supreme People's Court has been drawing its own lines between AI-generated works protected by human intellectual contribution and training-data disputes reported in connection with eligibility and authorship questions. A 2026 checklist should therefore flag which jurisdiction's standard governs, because a claim that survives in Washington may still face a different analysis in Munich or Beijing.
What the 2024-2026 AI rules changed, and what they did not
The clearest recent change concerns inventorship, not eligibility. On July 17, 2024, the USPTO issued its Guidance Update on AI-Assisted Inventions, updated MPEP sections accordingly, and issued a notice about the inventorship determination for an application naming a DABUS-style inventor. In practice, a human who conceived the claimed invention generally must be named, while a person who only provided AI assistance usually does not qualify; the USPTO's threshold is a natural-person conception test, and the guidance is part of the MPEP, 9.09. Eligibility rules did not change materially in 2024 or 2025, and recent commentary about Congress "fixing" § 101, as discussed in 2025 industry analysis, has not produced enacted statutory change as of this writing. Instead, the pressure on AI applications has come from the growing number of § 101 rejections and post-grant invalidations involving machine-learning claims, with one 2025 study reported as finding a markedly higher rate of § 101 invalidations for AI patents, particularly in the context of pending Supreme Court eligibility proceedings. The lesson for a checklist is to treat AI eligibility as a live prosecution risk requiring claim-level analysis, not as a settled matter that the USPTO's AI guidance has resolved. A 2026 checklist should track both the 2019 eligibility guidance and any new memoranda or Federal Register notices, because the framework is stable but its application is contested.
Evidence, training data, and the SMED question
One recurring practical dispute is whether laboratories may use experimental evidence during prosecution to rebut an eligibility rejection, and the USPTO's 2024 clarification on Rule 132 "SMED" evidence is directly relevant. Under the MPEP, § 608.02(p), Rule 132 evidence showing that a claimed arrangement unexpectedly produced a surprising result or that a cited reference taught away an element can rebut certain prima facie rejections of obviousness under § 103, and examiners are reminded that such evidence is not limited to experimental-use data. That said, Rule 132 evidence does not transform an abstract idea into an eligible concept; it can show nonconventionality and unexpected results, but the claim language itself must still recite an eligible arrangement. For AI inventions, the checklist should therefore ask whether experiments were documented, whether the surprise arose from a technical feature of the model or data pipeline, and whether the specification explains the result so that the evidence is tied to what a skilled artisan would understand. Applicants who collect laboratory data after the critical date without a data-deposition plan often discover later that the evidence cannot be substantiated, so the checklist should include a documentation step at the experimental stage, not at the response-to-office-action stage. Related issues, such as whether model weights or generated training data constitute patentable subject matter under § 101, remain contested and should be flagged for counsel rather than assumed.
Prior-art searching and comparing review options
Eligibility analysis and prior-art searching are separate exercises, but a good checklist pairs them, because the Alice second prong is measured against the prior art. The USPTO has extended its AI-driven prior-art search pilot into 2026 and is waiving petition fees for participating applicants, which lowers the cost of that option but does not make it a substitute for a professional search. A human search remains valuable for non-patent literature — product manuals, open-source repositories, and conference papers — that automated tools may underweight. The table below contrasts the main review options a 2026 applicant is likely to consider.
| Feature | Attorney-led full review | AI search pilot | Internal screening |
|---|---|---|---|
| Cost | Roughly $400–$800/hour at large firms; $150–$400/hour at smaller firms | USPTO search fees apply; petition fee waived during the pilot as reported in 2026 | Staff time only |
| What it covers | § 101, § 102, § 103, § 112, inventorship, and filing strategy | Automated prior-art retrieval with human review | Early go/no-go opinion |
| Time | Weeks to a few months | Search results typically delivered on the USPTO schedule | Days |
| Main risk | Higher cost, slower pace | Misses non-patent literature | False confidence from an unreviewed opinion |
| Best for | Commercial portfolios and contested claims | Screening a wide claim set cheaply | Budget triage |
Common mistakes that undermine the checklist
The first common mistake is testing eligibility before testing novelty, which inverts the risk order. If a claim is anticipated by a single reference, drafting around § 101 is wasted effort. The second mistake is writing claims around a product marketing description rather than around the technical contribution; checklist owners frequently confuse "improved accuracy" with an inventive feature, when accuracy alone is usually a result the abstract-idea framework treats skeptically. The third mistake is assuming that adding a generic processor, a database, or a field-of-use limitation cures an otherwise abstract claim, which the Alice decision directly contradicts. The fourth mistake is ignoring § 112, because a specification that describes a model at the level of a black box can be eligible and still fail written description. The fifth mistake is treating an internal AI search as exhaustive; the reported USPTO pilot extension and the wider availability of search tools in 2026 mean the search is cheaper, not more complete. The sixth mistake is deferring inventorship questions until after the application is drafted, even though the USPTO's July 17, 2024 guidance makes clear that inventorship errors can require correction. A disciplined checklist addresses all six mistakes in order, because each one compounds the cost of the next.
Timing, cost, and when to act
Timing matters because resources spent on an application that fails § 101 may need to be redirected to trade-secret protection, defensive publication, or a narrower claim. Most commercial applicants review a disclosure within two to four weeks of receiving it, complete a prior-art and eligibility screen within one to three months, and file before any public disclosure, public demo, or conference talk; the America Invents Act created the one-year grace period for the inventor's own disclosures, but foreign rights are often lost immediately. Costs for a full AI eligibility review typically fall in the range of a few thousand dollars for a simple screening and $5,000 to $20,000 for a complex portfolio review at typical US rate ranges, with additional cost for claim amendments, search opinions, and foreign filings. A practical threshold is to act now if the invention is close to a product launch, if a competitor has filed, if the invention relies on training data that may be challenged, or if the business plan depends on an issued patent rather than a pending application. The USPTO pilot's petition-fee waiver, as reported in 2026, is a reason to move quickly while it lasts, but not a reason to skip attorney review. The strongest 2026 posture is to preserve options early, document experiments as they happen, and revisit the claim set when eligibility guidance or case law shifts.
A working sequence for the next 90 days
Within the first 30 days, assemble a disclosure that identifies the technical problem, the specific model or pipeline improvement, and any measured results, and run an internal screen against the common mistakes above. Between days 31 and 60, commission either a USPTO pilot search or a professional prior-art search, supplement it with non-patent literature, and have counsel test the broadest claim against the two-step framework, comparing it with the closest references found. Between days 61 and 90, file if the position is acceptable, or use the results to narrow claims, switch to trade-secret treatment for the model weights, or publish defensively if the eligibility risk is too high. During all three phases, the evidence file should track experimental dates, inventor contributions, and Rule 132 material, and the inventorship record should follow the USPTO's 2024 guidance. This sequence fits most organizations, and each stage is optional if a lawyer recommends skipping it for cost reasons. The checklist is finished when the team can state in one paragraph why the claim is markedly inventive over the closest prior art and how the specification supports that position.