AI Patents in 2026: Navigating Section 101 Eligibility for AI and Software Inventions

If you’ve built something genuinely new using artificial intelligence — a novel training algorithm, a machine learning system that detects anomalies in real time, a neural network architecture that solves a problem competitors haven’t cracked — you may be wondering whether a patent can protect it. The short answer is: maybe, but the path is harder in the United States than almost anywhere else in the world, and it just got more complicated.

The fundamental obstacle is 35 U.S.C. § 101, the provision of the Patent Act that defines what kinds of inventions are eligible for patent protection in the first place. Courts have interpreted § 101 to exclude abstract ideas, laws of nature, and natural phenomena from patent protection. In practice, this has meant that enormous categories of software and AI innovation — the kind that drives the modern technology economy — have been swept out of patent eligibility before examiners ever get to ask whether the invention is novel or non-obvious.

This post explains where things stand in mid-2026: what the law says, how courts have applied it to AI and machine learning patents, what the USPTO has been telling examiners to do, and what practical strategies are available if you’re a startup or small business trying to protect an AI-based product.

The Alice/Mayo Framework: What It Is and Why It Matters

The current framework for analyzing patent subject matter eligibility under § 101 traces to two Supreme Court decisions: Mayo Collaborative Services v. Prometheus Laboratories, 566 U.S. 66 (2012), which addressed medical diagnostic methods, and Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014), which addressed software patents. Together they established a two-step test — universally called the “Alice” test or the “Alice/Mayo” framework — that has governed software and AI patent eligibility ever since.

Step One: Is the Claim Directed to an Abstract Idea?

The first question is whether the patent claim, considered as a whole, is “directed to” a patent-ineligible concept: an abstract idea, a law of nature, or a natural phenomenon. For AI and software patents, the relevant category is almost always “abstract idea,” which courts have interpreted to include mathematical concepts, mental processes that a human could perform, and methods of organizing human activity.

This step is deceptively difficult. Courts look at what the claim is fundamentally “about” — its focus and character — rather than taking a hyper-literal reading of its words. A claim that recites steps performed by a computer does not automatically escape the abstract idea label. If those steps could in principle be done mentally, or if the claim is really just describing an algorithm without tying it to a concrete technological improvement, courts will find the claim directed to an abstract idea at step one.

For AI inventions specifically, this has meant that claims framed around “using machine learning to predict X” or “applying a neural network to classify Y” tend to fail step one unless the claim also explains what specific technical improvement results — not just what result is achieved, but how the particular technical architecture or method achieves something that generic computing could not.

Step Two: Is There an “Inventive Concept”?

If a claim fails step one — if it is directed to an abstract idea — the analysis moves to step two. Here the court asks whether the claim elements, individually or as an ordered combination, contain an “inventive concept” that amounts to “significantly more” than the abstract idea itself.

This is where many AI patent claims die. The Supreme Court in Alice made clear that simply appending “do it on a computer” to an otherwise abstract method does not supply an inventive concept. Routine and conventional steps — even if they involve sophisticated hardware — do not transform an ineligible abstract idea into a patentable invention. For AI claims, generic references to “machine learning,” “a neural network,” or “a processor” have consistently been held insufficient to meet the step-two threshold when the underlying method is conventional.

What can supply an inventive concept? The cases suggest that a specific, unconventional technical approach — something that actually improves the functioning of the technology itself, not just uses the technology to perform an abstract task — can meet the bar. But identifying and claiming that kind of improvement is the core challenge of AI patent prosecution.

How Courts Have Applied Alice to AI and Machine Learning Patents

The Federal Circuit — the appellate court with nationwide jurisdiction over patent cases — has been the primary arena for working out how the Alice framework applies to AI. The decisions from 2025 and early 2026 have hardened the law in ways that AI patent applicants must understand.

Recentive Analytics v. Fox Corp. (April 2025): The Landmark Ruling

In Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437 (Fed. Cir. Apr. 18, 2025), the Federal Circuit issued its first major ruling directly addressing machine learning patents under § 101. The patent-owner had sued Fox Corporation asserting four patents relating to machine-learning-generated network maps and broadcast schedules for television events. The Federal Circuit affirmed the district court’s dismissal, finding all four patents ineligible.

The court’s central holding was blunt: applying established machine learning techniques to a new data environment, without reciting a specific improvement to the underlying machine learning process itself, is directed to an abstract idea and does not supply an inventive concept. The opinion drew an explicit line between two categories of claims: (1) claims that improve how machine learning works — improving the mathematical algorithm, improving training efficiency, reducing computational overhead — and (2) claims that simply use machine learning, however impressive the application, to accomplish a task in a new domain.

Claims in the first category can potentially survive § 101. Claims in the second category fail. The Federal Circuit’s formulation — that the “generic use of AI without other parameters, such as ‘improving the mathematical algorithm or making machine learning better,’ is abstract” — has become the operative standard.

The practical implication for AI patent applicants is stark. The novelty of your use case, the sophistication of your business application, or the value your system creates in the marketplace are not what matter for § 101. What matters is whether you can point to a specific, claimed technical improvement in the AI system itself.

Rensselaer v. Amazon (February 2026): Recentive Confirmed and Extended

In February 2026, the Federal Circuit affirmed another § 101 invalidity ruling in Rensselaer Polytechnic Institute v. Amazon.com, Inc., No. 24-1725 (Fed. Cir. Feb. 24, 2026). The patent at issue claimed a method for processing natural language input using case-based reasoning — essentially a form of machine learning that uses past experience to resolve ambiguities in human language. Rensselaer argued that its “novel application” of case-based reasoning to natural language processing distinguished the claims from the abstract idea.

The Federal Circuit disagreed, quoting Recentive directly for the proposition that applying AI to a novel environment does not make a claim non-abstract. Judge Dyk, writing for the court, noted that the claims recited a “general purpose computer system” and were “not limited to a particular computer system platform, processor, operating system, or network.” The court also rejected Rensselaer’s argument that a “metadata database” constituted an unconventional technological improvement, finding that the patent’s own specification described a metadata database as “well-known in the art.” The court’s closing formulation captures the current state of the law: “A conventional application of case-based reasoning, even to a novel environment, is abstract.”

Together, Recentive and Rensselaer define the floor. They tell us what will not work. Claiming machine learning in a new application domain, without more, is the wrong strategy.

What Survives: The Positive Signal from Ex Parte Desjardins

Not all recent developments have been bad for AI patent applicants. In September 2025, the USPTO’s Appeals Review Panel — acting under Director John Squires — issued a decision in Ex Parte Desjardins, Appeal No. 2024-000567, that has become a significant positive precedent. The decision was designated precedential on November 4, 2025, meaning it now binds all USPTO examiners and PTAB panels.

The patent application in Desjardins claimed a machine-learning method for sequentially training models while preserving previously learned knowledge — a technique designed to prevent what researchers call “catastrophic forgetting,” where a model trained on new data loses its performance on prior tasks. The PTAB had characterized the invention as abstract mathematical computation and issued a § 101 rejection. Director Squires vacated that rejection on rehearing, finding that the claimed steps — specifically, selective parameter adjustment to optimize performance on a second learning task while protecting performance on a first — reflected a technical improvement in how the model itself functions.

Desjardins establishes an important pathway: claims directed to improving the performance, efficiency, or architecture of machine learning models can be patent-eligible, even when framed in software terms, as long as the improvement is to the AI technology itself and not merely to a downstream application.

USPTO Guidance: What Examiners Are Looking For

The USPTO has issued several rounds of guidance in the last two years that collectively define the examiner’s-eye view of AI patent eligibility.

The July 2024 Guidance Update

In July 2024, the USPTO published a formal guidance update in the Federal Register — 2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence, 89 Fed. Reg. 52,134 (July 17, 2024) — accompanied by three new subject matter eligibility examples directly addressing AI inventions.

The three examples are instructive:

Example 47 covers an artificial neural network that detects anomalies in received data. The guidance explains how claims reciting specific neural network architectures and specific anomaly-detection functions can integrate the underlying mathematical concepts into a practical application, establishing patent eligibility.

Example 48 covers AI-based methods for analyzing speech signals and separating desired speech from background noise. The guidance shows how claims tied to a specific technical problem in signal processing — improving signal quality in a defined context — can satisfy the Alice framework.

Example 49 covers an AI model for personalizing medical treatment based on individual patient characteristics. This example demonstrates eligibility where the AI is integrated into a specific application with a concrete, real-world result that is not merely a display of information.

Across all three examples, the recurring theme is the same as the case law: the claims must do more than describe using AI to accomplish a task. They must tie the AI operation to a specific technical improvement or a concrete, definite application.

The August 2025 Examiner Memorandum

In August 2025, the USPTO issued a memorandum from the Deputy Commissioner for Patent Examination Policy to Technology Centers 2100, 2600, and 3600 — the technology centers that handle most software and AI applications. The memo addressed several examiner behaviors that had been generating improper § 101 rejections.

Most significantly, the memo instructed examiners that a claim limitation should be treated as a “mental process” — one of the abstract idea categories — only if it can practically be performed in the human mind or with pen and paper. Examiners were explicitly cautioned not to stretch the mental process category to cover limitations that require machine-based operations, such as those tied to AI model execution or hardware-implemented functions.

The memo also set a threshold for issuing § 101 rejections: examiners should only do so if there is a greater than 50% probability that the claim is ineligible. Mere uncertainty is insufficient grounds for a rejection.

The December 2025 SMED Guidance and MPEP Update

In December 2025, Director Squires issued guidance addressing Subject Matter Eligibility Declarations (SMEDs) — sworn declarations from inventors or experts attesting to the technical nature of a claimed improvement. The guidance explains how examiners must give weight to timely-submitted SMEDs and provide a substantive response explaining why the declaration does or does not change the eligibility analysis.

The MPEP (Manual of Patent Examining Procedure) was also updated in December 2025 to incorporate Desjardins directly. Examiners are now instructed to search for a practical application that constitutes “an improvement to other technology or technical field” and are cautioned not to evaluate claims “at such a high level of generality that potentially meaningful technical limitations are dismissed without adequate explanation.”

The sum of this guidance is more favorable to AI patent applicants than the case law alone might suggest. The USPTO — under the current administration — has signaled that § 101 should not function as a blanket exclusion for AI technology. The Desjardins precedent and the August 2025 memo in particular have created more room for AI patent applicants to overcome examiner rejections at the prosecution stage, even as the Federal Circuit’s standards remain demanding in litigation.

Drafting Strategies: How to Frame AI Patent Claims

The research above points toward concrete drafting strategies for AI patent applications. The goal in every case is the same: position the invention as a technical improvement, not a new use of an existing tool.

Lead with the Technical Problem, Not the Business Result

Claims should identify a specific technical problem — one that exists in the prior art of machine learning or computing — and explain how the claimed invention solves that problem at a technical level. “Our algorithm produces better recommendations” is a business result. “Our method reduces catastrophic forgetting in sequential learning by selectively freezing weight parameters in network layers not activated by the new training task” is a technical result. The latter framing anchors the claim in the technology itself.

The Federal Circuit has been responsive to claims that specify a technical improvement to computer functioning, data processing efficiency, model accuracy, or computational architecture. Claims that merely describe a desirable output — a better schedule, a more accurate prediction, a more relevant recommendation — tend to fail.

Describe the Architecture, Not Just the Function

Generic references to “a neural network,” “a machine learning model,” or “a processor” are insufficient. Claims should recite specific architectural features: the type of network (convolutional, recurrent, transformer-based), the specific training regime, the specific way in which data is processed or parameters are updated. The more the claim reads like a real engineering specification — not a marketing description — the better.

Specific hardware integration can also help. Claims that tie a method to a particular type of hardware, such as an application-specific integrated circuit (ASIC) or a specialized processing unit, are less likely to be categorized as mental processes and more likely to demonstrate a technical character.

Use the Specification to Build the § 101 Record

What appears in the patent specification matters enormously, even though § 101 is officially a legal question about the claims. Courts and examiners look to the specification to understand the technical context of the claimed invention. If the specification describes the prior art problem in technical terms, explains why existing machine learning approaches fail to solve it, and articulates the technical mechanism by which the claimed invention solves it, that record supports the claim’s case for eligibility.

Conversely, a specification that focuses primarily on business advantages, customer benefits, or application-domain novelty — without a corresponding technical narrative — builds a bad § 101 record even if the claims are drafted carefully.

Consider SMEDs as a Prosecution Tool

As noted above, the USPTO’s December 2025 guidance has made Subject Matter Eligibility Declarations a more formal and effective prosecution tool. In a case where examiners are pushing back on § 101, a well-crafted declaration from a technical expert or the inventor — attesting to the specific technical improvement the claimed invention provides over the prior art — can shift the burden and require a more substantive examiner response.

Patents vs. Trade Secrets: An AI Innovation Strategy Decision

For many AI companies, § 101 uncertainty raises a threshold question: should you pursue patent protection at all, or does trade secret law offer a more reliable path?

The honest answer is that it depends on the nature of your specific innovation, and for many AI companies the right answer is a combination of both.

When Trade Secret Protection Makes Sense for AI

Trade secrets are well-suited for protecting AI assets that are difficult to detect or reverse-engineer from the outside: proprietary training datasets, specific model weight configurations, internal optimization techniques, and infrastructure choices that competitors cannot observe by looking at your product. Unlike patents, trade secrets require no disclosure and have no term limit — your protection lasts as long as you maintain reasonable confidentiality.

For companies facing § 101 uncertainty, trade secret protection has an additional practical advantage: it provides legal coverage immediately, without the 2–4 year USPTO prosecution process, and without the risk that a § 101 rejection leaves you with a published application but no issued patent.

The limits of trade secret protection are real, however. Trade secrets offer no protection against independent discovery: if a competitor develops the same technique on their own, you have no claim against them. Trade secrets are also vulnerable to employee mobility, data breaches, and reverse engineering of deployed products. The Defend Trade Secrets Act (DTSA), 18 U.S.C. § 1836, provides federal civil remedies for misappropriation, but enforcing those rights requires showing that the secret was taken, not independently developed.

When Patents Make More Sense

If your AI innovation is visible in your deployed product — if competitors can observe what your system does and work backward to understand how it works — trade secret protection may already be eroding the moment you launch. In that scenario, a patent that specifically claims the technical method is a stronger long-term defense, because it protects against independent development as well as copying.

Patents are also valuable in licensing contexts, in investor due diligence, and in establishing ownership rights in an acqui-hire or acquisition. A portfolio of AI patents, even in a challenging § 101 environment, signals technical depth and creates negotiating leverage.

The practical strategy for most AI startups is to map their innovation portfolio carefully: use trade secrets for model weights, training data, and internal optimization techniques that are not exposed in the product; pursue patents for the technical methods that are embodied in the deployed product or that describe a specific architectural improvement that could be reverse-engineered.

The International Landscape: Europe and China

US companies competing globally should understand that the United States is an outlier in the difficulty of its AI patent eligibility rules. Both the European Patent Office (EPO) and China’s National Intellectual Property Administration (CNIPA) apply standards that, while not without their own complications, are generally more receptive to AI patents than the US § 101 regime.

Europe: Technical Character Is the Key

The European Patent Convention (EPC) excludes “mathematical methods,” “programs for computers,” and “methods for doing business” from patentability when claimed “as such.” But the EPO has developed a well-defined framework for when software and AI inventions can be patentable: the claimed invention must make a “technical contribution” — it must solve a technical problem by technical means.

In practice, this means that AI inventions that are tied to a concrete technical process — industrial control systems, medical devices, signal processing, image analysis, scientific instrumentation — are generally patentable in Europe as long as the technical contribution is adequately described. The EPO has published detailed examination guidelines for AI and machine learning, recognizing that a neural network trained for a specific technical application can constitute a patentable technical contribution even when the network architecture itself is known.

The EPO standard is not identical to the US standard, but it is applied more predictably, and the rate of § 101-equivalent rejections for AI applications is substantially lower.

China: Quantifiable Technical Effects

China’s CNIPA requires that AI patent claims demonstrate a clear technical effect: the AI algorithm must be described in relation to a specific technical problem, and the technical effect must be quantifiable or at least concretely described. Claims that show how an AI algorithm correlates with the internal structure of a computer — specifying input-output transformations, architectural interactions, or measurable performance improvements — tend to fare well.

China has become one of the world’s largest generators of AI patent applications. EPO data from 2025 showed China entering the top three sources of European patent applications for the first time, reflecting both domestic growth in AI innovation and an increasing sophistication in global patent strategy among Chinese technology companies.

For US companies, the international picture is a strategic opportunity: if your AI invention faces § 101 headwinds in the US, European and Chinese protection may still be achievable, and a global portfolio can provide meaningful competitive protection even if the US patent proves difficult to obtain.

The Legislative Landscape: Congressional Reform Efforts

Congress has been aware of the § 101 problem for years, and reform legislation has been introduced in multiple sessions. As of mid-2026, however, no § 101 reform bill has been enacted.

The most significant pending legislation is the Patent Eligibility Restoration Act of 2025 (PERA), S. 1546, reintroduced in the 119th Congress by Senators Thom Tillis (R-NC) and Chris Coons (D-DE) and Representatives Kevin Kiley (R-CA) and Scott Peters (D-CA). PERA would amend § 101 to eliminate all judicially created exceptions to patent eligibility. Under PERA’s proposed framework, any process that cannot practically be performed without the use of a machine (including a computer) would be patent-eligible. PERA specifies narrow statutory exclusions — pure mathematical formulas not part of an invention, processes that a human could perform entirely in the mind, unmodified human genes, and unmodified natural materials — but would otherwise sweep away the Alice/Mayo framework entirely.

PERA has bipartisan support and has attracted attention in Senate Judiciary Committee hearings, where supporters have emphasized the cost of § 101 uncertainty for AI investment and diagnostics. The bill’s fate remains uncertain, as it continues to grapple with how to carve out purely economic, financial, or social methods without inadvertently sweeping in the business-method patents that Alice was designed to eliminate.

The practical reality is that legislative reform, even if it happens, will take years to implement and litigate into settled doctrine. Companies making IP decisions today cannot count on PERA or any other reform bill as part of their near-term strategy.

Prosecution Tips for Startups and SMBs

For early-stage companies with limited budgets, navigating AI patent prosecution requires prioritizing carefully.

File provisionals early. A provisional patent application establishes your priority date without requiring full claim drafting. This gives you 12 months to assess whether the technology is worth pursuing and to gather technical evidence for the § 101 analysis. Given how quickly AI evolves, locking in your priority date while the invention is still novel is critical.

Invest in the specification, not just the claims. A well-written technical specification that thoroughly describes the technical problem, the prior art’s limitations, and the specific technical mechanism of your invention costs the same to file as a thin one — but it is the foundation on which successful prosecution is built. Examiners and courts read specifications. Don’t skimp.

Claim at multiple levels of specificity. Draft a claim set that includes broad claims, medium-scope claims, and narrow claims. The broad claims may face § 101 challenges; the narrow claims that recite specific architectural features or specific technical mechanisms are your fallback positions. A patent with narrow but valid claims is far more valuable than a patent application that never issues.

Use continuation applications strategically. After an initial patent issues, continuation applications allow you to pursue additional claims based on the same specification. As the case law evolves, continuations let you adapt your claiming strategy to the current legal landscape without filing a new application with a new priority date.

Respond to § 101 rejections with technical evidence. The August 2025 examiner memo and the December 2025 SMED guidance have given practitioners new tools to respond to over-broad § 101 rejections. If you receive a § 101 rejection, consider whether an inventor declaration or expert declaration attesting to the specific technical improvement could change the examiner’s analysis.

When to Fight a Section 101 Rejection — and When to Move On

Not every § 101 rejection is worth fighting. The decision to appeal, continue prosecuting, or abandon a patent application should be driven by a realistic assessment of the claims’ merits and the commercial value of the technology.

Fight when the claims describe a genuine technical improvement. If your invention solves a specific technical problem in machine learning or computing — not just applies AI to a new domain — then you have a viable argument under both Desjardins and the USPTO’s current guidance. The legal landscape is not uniformly hostile, and examiners are under instructions to be more careful about over-broad § 101 rejections.

Amend rather than appeal when possible. Prosecution amendments that add specific technical features — even at the cost of narrowing the claim scope — are often faster and cheaper than appeals. A narrower but issued patent protects your priority date and gives you an asset that can be enforced or licensed.

Consider abandonment when the claims are application-domain claims. If your strongest arguments for eligibility amount to “we’re using machine learning in a new industry,” that argument has been explicitly rejected by Recentive and Rensselaer. Spending thousands of dollars on an appeal that is unlikely to succeed is not good capital allocation for a startup.

Pair patents with trade secret protection regardless. Whatever you decide about a specific patent application, keep trade secret protection in place for the proprietary elements of your AI system. The two forms of protection are complementary, not mutually exclusive.

Conclusion

The landscape for AI patents in 2026 is genuinely challenging, but it is not hopeless. The Federal Circuit has drawn clear lines: claims that improve AI technology itself — its training efficiency, its architectural performance, its ability to avoid degradation — can survive § 101. Claims that simply use AI to accomplish a task in a new domain, however valuable or novel that task, will not.

The USPTO’s guidance over the past two years has created meaningful room for AI patent applicants who understand these distinctions and draft accordingly. Ex Parte Desjardins, now precedential, provides a positive signal that claims focused on technical improvements to machine learning systems can succeed before the USPTO even in the current environment. The August 2025 examiner memo and the SMED guidance have given practitioners new tools to respond to over-broad § 101 rejections.

For businesses and startups building AI products, the strategic imperative is the same as it has always been in patent law: work with qualified patent counsel before you start drafting, invest in claim language and specification quality, and build a protection strategy that combines patents, trade secrets, and — where appropriate — international filings.

The law will continue to evolve. Congress may yet reform § 101, and the Federal Circuit will continue to decide cases that refine the boundaries. What will not change is the basic principle that meaningful patent protection requires claiming specific technical improvements, not abstract results. Companies that internalize that principle and draft to it are the ones whose AI patents will survive.

This post is for general informational purposes only and does not constitute legal advice. Patent eligibility questions are highly fact-specific. Contact a qualified patent attorney to evaluate your specific invention and prosecution strategy.