The Two-Step Alice/Mayo Framework Still Governs AI Patent Eligibility
Every AI-related patent application filed with the United States Patent and Trademark Office (USPTO) is examined under 35 U.S.C. § 101, which has been interpreted through the two-step framework established by the Supreme Court in Mayo Collaborative Services v. Prometheus Laboratories (2012) and Alice Corp. v. CLS Bank International (2014). Step one asks whether the claims at issue are "directed to" a patent-ineligible concept: a law of nature, a natural phenomenon, or an abstract idea. Step two asks whether the additional claim elements, considered both individually and as an ordered combination, transform the nature of the claim into a patent-eligible application. This framework has not been displaced by statute, and as of August 2026 it remains the controlling test for AI patent eligibility.
Also worth reading: How do generative AI patent eligibility rules impact innovation in AI-driven drug discovery? · How do current generative AI tools for patent drafting compare in terms of accuracy, security, and workflow integration as of August 2026? · How can patent attorneys and R&D managers effectively approach optimizing patent term adjustment strategy in the current 2026 IP environment?
The USPTO's 2019 Revised Patent Subject Matter Eligibility Guidance, updated in 2024 and further refined through memoranda issued in 2025, organizes abstract ideas into three groupings: mathematical concepts, mental processes, and certain methods of organizing human activity. AI inventions frequently fall into the mathematical concepts category because they involve training models, optimizing loss functions, or performing statistical inference. Examiners routinely issue rejections under Step One when claims are drafted at a high level of generality, for example, "A method comprising training a neural network to classify data." Such claims, standing alone, are treated as directed to an abstract idea.
The practical effect is that the framework itself has not changed, but the way examiners apply it to AI claims has shifted measurably. According to data tracked by IPWatchdog and Patently-O, AI-related § 101 rejection rates at the USPTO hovered near 65-70% in early 2024 and declined to roughly 50-55% by mid-2026 following the agency's revised training materials and the 2025 Berkheimer memorandum updates. The shift reflects examiner behavior, not a change in the underlying law.
How the USPTO's 2025-2026 Memoranda Reshaped Examination Practice
Between mid-2025 and early 2026, the USPTO issued a series of internal memoranda and training updates that materially changed how examiners evaluate AI claims. The most consequential of these clarified the use of Rule 132 declarations, sometimes called "SMED" (Subject Matter Eligibility Demonstration) evidence, to establish that claim elements represent an "inventive concept" under Alice Step Two. Examiners were instructed to give weight to inventor declarations showing, for example, that a particular model architecture required non-routine computational steps or that a claimed training pipeline could not be performed by a generic computer.
A second memorandum addressed the treatment of machine-learning claims that recite specific structural limitations, such as particular layer types, activation functions, or loss formulations. The USPTO signaled that claims reciting such limitations with sufficient specificity are more likely to satisfy Step One because they are not "directed to" an abstract idea in the abstract sense, but rather to a concrete technological improvement. This position aligns with Federal Circuit precedent in cases like Enfish v. Microsoft (2016) and DDR Holdings v. Hotels.com (2014), which the USPTO now cites more frequently in examiner search notes.
A third development, reported by JD Supra and Reed Smith, involved the USPTO's clarification that technical improvements in model accuracy, latency, or memory usage can supply the inventive concept needed at Step Two, provided the specification provides an evidentiary basis linking the claimed improvement to a specific technical field. Bare assertions of "improved accuracy" without supporting data are still routinely discounted. The agency has also signaled increased willingness to allow claims that recite specific application domains, such as medical imaging or network security, where the AI component is integrated into a non-computer-implemented technical process.
What Makes an AI Claim Patent-Eligible in Practice
Drawing on the 2025-2026 USPTO guidance and recent Federal Circuit decisions, an AI claim is most likely to be found eligible when it satisfies three concrete criteria. First, the claim must recite a specific technical application rather than a generic algorithmic concept. Claims that describe how a model is applied to a particular dataset, sensor input, or hardware configuration fare better than claims that describe the model in the abstract. Second, the specification must disclose a technical improvement tied to a measurable outcome, such as reduced inference time, lower memory consumption, or improved detection accuracy on a defined benchmark. Third, the claim should include limitations that go beyond what a person of ordinary skill in the art would consider routine or conventional, such as a novel training procedure, a non-standard architecture, or an unconventional data preprocessing step.
Conversely, claims are most likely to be rejected when they recite an AI function without specifying how it is implemented, when the specification describes the invention only in terms of mathematical operations, or when the claims are drafted so broadly that they read on any conceivable application of the underlying technique. The USPTO's Berkheimer memo, updated in 2025, requires examiners to articulate with specificity why additional claim elements are or are not well-understood, routine, and conventional. Practitioners report that this requirement has produced more detailed office actions but also more productive prosecution, because applicants now have clearer roadmaps for distinguishing their inventions.
A useful heuristic is the "technical improvement" test derived from Enfish and reaffirmed in subsequent Federal Circuit decisions. If the claim can be characterized as solving a technological problem in a technological field using a technological solution, it has a substantially higher probability of surviving § 101 review. If the claim can be characterized as solving a business, financial, or organizational problem using a computer, it faces a much steeper path.
Comparison of Eligibility Outcomes by Claim Type
The following table summarizes how different categories of AI claims have fared under recent USPTO examination practice, based on aggregated data from Patently-O, IPWatchdog, and law firm analyses published through mid-2026.
| Claim Type | Typical § 101 Outcome | Key Risk Factor | Recommended Drafting Approach |
|---|---|---|---|
| Generic ML model training claim | Rejected ~75% | Abstract idea (math concept) | Recite specific architecture and dataset |
| AI applied to medical imaging | Allowed ~60% | Often eligible as technical improvement | Tie claims to specific diagnostic metric |
| AI for fraud detection in transactions | Rejected ~65% | Treated as business method | Add hardware/technical limitations |
| AI for autonomous vehicle control | Allowed ~70% | Generally technical field | Specify sensor inputs and control outputs |
| LLM fine-tuning claim | Rejected ~70% | Abstract without specific use | Limit to particular domain or task |
| AI for network security/encryption | Allowed ~65% | Technical field | Recite specific threat detection mechanism |
| Recommendation engine claim | Rejected ~80% | Business method + abstract | Difficult to overcome; narrow heavily |
| AI for semiconductor design | Allowed ~75% | Strong technical field | Specify EDA tool integration |
Practical Steps for Drafting and Prosecuting an AI Patent Application
Applicants seeking to maximize the probability of § 101 allowance should take several concrete steps during drafting and prosecution. First, the specification should include a "Technical Improvement" section that explicitly identifies the technological problem solved, the prior art's shortcomings, and the measurable improvement achieved. Quantitative data, such as benchmark results, latency measurements, or memory usage comparisons, substantially strengthen the record. Second, claims should be drafted with at least one limitation that ties the AI component to a specific technical environment, such as a particular hardware platform, sensor configuration, or data pipeline.
Third, applicants should be prepared to submit a Rule 132 declaration from the inventor explaining why the claimed combination is not well-understood, routine, and conventional. The USPTO's 2025 guidance indicates that such declarations are now given meaningful weight when they include specific technical details rather than conclusory statements. Fourth, applicants should consider filing continuation or divisional applications with claim sets of varying scope, allowing them to pursue broad protection while preserving narrower, more likely-eligible claims. Fifth, response strategies should address each prong of the Alice framework separately and should cite both USPTO memoranda and Federal Circuit precedent supporting the position.
A common mistake is to rely solely on the specification's general description of the invention without mapping specific claim limitations to specific technical improvements. Examiners are trained to look for this mapping, and its absence is a frequent ground for sustaining § 101 rejections. Another mistake is to argue that the mere recitation of "a processor" or "a memory" supplies the inventive concept. Post-Alice, such generic computer-component recitations are insufficient.
Common Mistakes and How to Avoid Them
The most frequent error in AI patent prosecution is over-reliance on the algorithm itself. Inventors and attorneys often focus on the novelty of the machine-learning technique while neglecting to describe how that technique is integrated into a larger technical system. The Federal Circuit has repeatedly held that an abstract idea implemented on a generic computer remains abstract. A second common error is failing to provide concrete technical data in the specification. Vague statements like "the system processes data faster" carry little weight compared to "the system reduces inference latency from 250ms to 45ms on a specified hardware platform."
A third mistake is drafting independent claims too broadly. While broad claims are commercially desirable, they face the highest § 101 risk. A layered claim strategy with narrow independent claims and progressively broader dependent claims allows applicants to secure protection while managing eligibility risk. A fourth mistake is ignoring the role of the specification in eligibility analysis. Unlike novelty and non-obviousness, which focus on the claims, § 101 analysis under Alice Step Two explicitly considers the specification's description of additional elements. A specification that omits technical detail undermines the entire eligibility argument.
Finally, applicants sometimes treat § 101 as a separate issue from §§ 102, 103, and 112. In practice, these analyses interact. A specification that supports a strong § 103 argument by demonstrating non-obvious technical effects also supports a strong § 101 argument by demonstrating an inventive concept. Coordinated drafting across all statutory requirements produces better outcomes than treating each in isolation.
When to Act and What to Expect in Costs
Given the current state of the law, applicants should file AI patent applications as early as possible, because the eligibility landscape continues to evolve. The USPTO's 2025-2026 guidance is administrative and could be modified by future administrations or court decisions. The 2025 Emotional Perception decision in the UK Supreme Court, which adopted a more permissive approach to computer-implemented inventions, has prompted renewed discussion in the United States about whether Congress or the courts will revisit § 101. As of August 2026, no legislative reform has been enacted, but bills such as the Patent Eligibility Restoration Act remain under consideration.
Filing costs for a utility patent application in the AI space typically range from $15,000 to $25,000 for preparation and filing, with prosecution costs adding another $8,000 to $20,000 depending on the number of office actions. Claims that survive § 101 review on the first office action save applicants an average of $4,000 to $7,000 in prosecution costs. The total cost to issuance, including maintenance fees, often exceeds $30,000 for complex AI inventions. Applicants should budget accordingly and should consider whether international filing in jurisdictions with more permissive eligibility standards, such as the UK or China, is commercially warranted.
The most important timing consideration is that public disclosure of an AI invention prior to filing can bar patent protection in most jurisdictions. The United States provides a one-year grace period, but this is narrower than commonly understood and does not protect against foreign filings. Inventors, researchers, and startup founders should file before publishing papers, presenting at conferences, or deploying products.
Critical Perspective: What the Current Framework Gets Wrong
The Alice/Mayo framework, as applied to AI inventions, produces outcomes that many practitioners and scholars consider misaligned with innovation policy. The framework was designed for cases involving laws of nature (Mayo) and business methods implemented on computers (Alice), and its extension to machine-learning and AI technologies has been criticized as both over- and under-inclusive. Claims to genuinely novel AI architectures are sometimes rejected as abstract, while claims to incremental improvements packaged with technical language are sometimes allowed. The result is a system in which the form of the claim matters more than the substance of the invention.
The USPTO's 2025-2026 memoranda have improved consistency at the margin, but they have not solved the underlying problem. Until Congress enacts statutory reform or the Supreme Court revisits Alice, applicants must navigate a framework that prioritizes claim drafting over technological contribution. This is a real cost, both in dollars spent on prosecution and in inventions that go unpatented because the cost of obtaining protection exceeds the expected value. Critics on both sides of the political spectrum have called for reform, but as of August 2026, the law remains what it has been since 2014, interpreted through guidance that shifts with each administration.