USPTO Raises Bar for § 101 Rejections in AI Patents
Legal Alerts
8.14.25
Key Takeaways
- Rejection Threshold – If it’s a close call, examiners are reminded not to issue a §101 rejection unless it’s more likely than not (more than 50%) that the claim is ineligible.
- Mental Process – Examiners are reminded to only count steps that a person could realistically do in their head (or with pen and paper).
- Recites vs. Involves – Examiners are reminded that a claim recites an exception if it clearly spells it out (like naming specific algorithms).
- Claim as a Whole – Examiners are reminded to evaluate how the additional elements work together with any recited exception to form a practical application.
- Improvement vs. “Apply It” – Examiners are reminded to check whether the claim actually improves a computer or technology (a specific solution or mechanism) or just uses a computer to do an abstract idea.
The United States Patent and Trademark Office (USPTO) has issued a new memorandum to examiners working in Technology Centers 2100, 2600, and 3600, reminding them of subject-matter-eligibility practices under 35 U.S.C. § 101, with particular attention to software- and AI-related inventions. Building on its prior AI-specific guidance and recent Federal Circuit case law, the memo notably cautions against stretching the ‘mental process’ category of § 101 to limitations that cannot practically be performed in the human mind and on close calls and reminds examiners to make a § 101 rejection only when it is more likely than not that a claim is ineligible. Although the memo does not announce new USPTO practice, these clarifications could lead to fewer § 101 rejections in borderline AI cases, a welcome outcome for companies seeking to expand in the AI space.
Patent eligibility under 35 U.S.C. § 101 requires that an invention fall within one of four statutory categories: process, machine, manufacture, or composition of matter. Even if this threshold is met, the invention must not be directed to a judicial exception (e.g., a law of nature, natural phenomenon, or abstract idea), unless it includes an “inventive concept” that adds significantly more to transform it into a patent-eligible application. This analysis is guided by the Supreme Court’s two-step Alice/Mayo framework: first, determining whether the claims are directed to a patent-ineligible concept; and second, evaluating whether the claims contain an “inventive concept” sufficient to transform the abstract idea into a patent-eligible application, adding “significantly more” than the judicial exception itself.[1]
Last year, the USPTO released guidance on subject matter eligibility under 35 U.S.C. § 101 for AI inventions in the 2024 AI-SME Update.[2] The guidance included examples and reinforced a fundamental principle: just as applying a computer to an abstract idea doesn’t make it patentable, merely incorporating AI, large language models (LLMs), or neural networks into an abstract idea cannot transform it into a patentable application. The guidance tackled the issue of evaluating whether a claim recites an abstract idea in Step 2A, Prong One, and considers improvements in Step 2A, Prong Two. It distinguished between claims that recite an abstract idea as an essential element and those that involve an abstract idea without it being part of the claim language. A claim reciting an abstract idea directly includes the abstract idea as an essential element of the claim itself. On the other hand, a claim involving an abstract idea implies that the claim is based on or utilizes an abstract idea but does not explicitly recite the abstract idea as part of the claim language. This distinction is crucial because claims that merely involve an abstract idea can still be eligible for patent protection if they integrate it into a practical application.
More recently, in Recentive Analytics, Inc. v. Fox Corp.,[3] the Federal Circuit addressed questions about the scope of patent eligibility under 35 U.S.C. § 101, evaluating whether patents that apply machine learning to optimize a technology meet the threshold for patent-eligible subject matter.[4] The Federal Circuit held that the patents were directed to abstract ideas and lacked any inventive concept that would transform them into patent-eligible subject matter. The decision signaled that AI-related patents must go beyond simply using machine learning to solve a problem in a new field. Courts will look for concrete, technical innovations, such as improvements to the ML algorithms, training processes, or the way the technology operates, that advance the underlying technology itself. Merely applying standard AI techniques to different data or automating existing tasks will not be enough.
The USPTO’s 2025 memo reinforces key eligibility concepts already introduced in the 2024 AI guidance and clarified through recent case law, most notably Recentive Analytics, Inc. v. Fox Corp. While the memo does not introduce new legal standards, it sharpens examiner expectations, especially in the context of AI and software-related inventions, and serves as a practical check on the overextension of § 101 rejections.
One major area of emphasis is the “mental process” category. Echoing the 2024 AI-SME Update, the memo reminds examiners that a limitation should be treated as a mental process only if it can be practically performed in the human mind or with pen and paper, such as evaluations or subjective judgments. Critically, examiners are cautioned not to stretch this category to cover limitations that require machine-based operations, like those tied to AI models or hardware-executed functions. This clarification directly supports AI applicants, whose claims often include steps far exceeding what the human mind can perform in practice.
The memo also reinforces the distinction between claims that recite a judicial exception and those that merely involve one, a distinction that the 2024 guidance made central to Step 2A, Prong One. For instance, a claim referencing “training a neural network” may involve mathematical ideas but doesn’t explicitly recite them. By contrast, a claim that names specific algorithms, such as backpropagation or gradient descent, recites an abstract idea and requires further analysis under Step 2A, Prong Two. This distinction proved crucial in Recentive, where the Federal Circuit held that using machine learning to optimize business decisions was still abstract and ineligible without more. The message to examiners: eligibility hinges not just on the subject matter but on how concretely it is presented in the claim.
At Step 2A, Prong Two, the memo continues to emphasize evaluating the claim as a whole, not dissecting it element by element. This guidance aligns with both the 2024 examples and judicial precedent. Examiners are instructed to focus on whether the claim integrates the abstract idea into a practical application, such as a technical improvement to a computer system or specific computing architecture. Here, the memo distinguishes between claims that offer technological improvements, like measurable gains in efficiency, security, or processing throughput, and those that simply apply an abstract concept using routine computing functions. This is consistent with the Recentive decision, where the court rejected claims that lacked any specific improvement to machine learning models or infrastructure.
Finally, the memo addresses examiner discretion on close calls. It reiterates that a § 101 rejection should only be made when it is more likely than not that the claim is ineligible, in line with the MPEP’s preponderance-of-the-evidence standard. This is a critical shift in tone as it discourages unnecessary rejections in borderline cases and encourages thoughtful application of the law. At the same time, compact prosecution remains essential, and examiners are reminded to evaluate and include all other statutory grounds for rejection (e.g., novelty, non-obviousness, definiteness) in the first Office action, regardless of whether a § 101 rejection is made.
For businesses seeking patents based on AI technology, the USPTO’s reminder memo doesn’t change §101, but it should rein in overbroad “mental process” theories and discourage close-call rejections, reducing friction for borderline filings. To benefit, draft claims that read like technical improvements, not “apply-it” automation, anchor AI features to concrete mechanisms, data structures, or hardware constraints, and tie them to measurable gains (latency, throughput, security). Use the recites vs. involves distinction strategically and don’t name specific algorithms unless you also show how your architecture integrates them into a practical application. And because compact prosecution remains the rule, be ready on §§102/103/112 with clear support and fallback positions. In short, the memo opens the door a bit wider, but only for AI inventions that advance the technology itself.
If you have any questions, please contact Diego Freire or your Dykema relationship attorney.
[1] Alice Corp. Pty. v. CLS Bankint ‘l, 573 U.S. 208 (2014); Mayo Collab. Servs. v. Prometheus Lab’ys, Inc., 566 U.S. 66 (2012).
[2] ttps://www.dykema.com/news-insights/navigating-ai-patent-eligibility-insights-from-the-usptos-2024-subject-matter-eligibility-guidance-update.html
[3] Recentive Analytics, Inc. v. Fox Corp., 1:22cv1545
[4] https://www.dykema.com/news-insights/ai-and-patent-eligibility-strategies-in-the-wake-of-recentive-analytics-v-fox-corp.html
Related People
