Written by Charles Dresser, Steve Wang, and Kurt Hoppmann.
Abstract ideas (e.g., mathematical concepts and tasks that may be performed in a human mind) are not eligible for patent protection, through a longstanding judicially created doctrine under 35 U.S.C. § 101. The question: what is an invention that is merely an abstract idea is nebulous and plagued with seemingly competing legal precedent. Accordingly, this question introduces risk that certain inventions may not be entitled to patent protection.
Artificial intelligence (AI) and machine learning (ML) have become sources of innovation. But, AI and ML inventions exemplify inventions that can be considered abstract and found ineligible for patent protection.
Recently, on August 4, 2025, the United States Patent and Trademark Office (USPTO) released a memorandum to patent examiners further clarifying its position on subject matter eligibility for artificial intelligence and machine learning inventions. This memo comes one year after, the USPTO July 2024 Subject Matter Eligibility Guidance provided initial clarity the USPTO’s position on AI and subject matter eligibility.
This paper summarizes main points from the August 2025 Memo and other USPTO guidance on AI and ML inventions.
1. Only AI Inventions that can be Performed in the Human Mind are Mental Processes
AI inventions, almost by definition, provide outputs that are similar to outputs of the human mind. Accordingly, AI patents can be drafted in such a way that they are abstract and ineligible. However, there are aspects of most AI inventions that cannot be performed in a human mind. For instance, generating a sequence of words from spectral features extracted from a speech signal is considered a task that cannot be performed in a human mind. See USPTO, July 2024 Subject Matter Eligibility Examples, Example 48, claim 2, p. 23 (July 2024) (“[s]ynthesizing speech waveforms from a cluster of numbers is not a process that can be practically performed in the human mind”). Conversely, clustering vectors is an AI function that the USPTO considers performable in the human mind. See id., p. 22 (“‘partitioning . . . into clusters’ encompasses a human arbitrarily selecting groups of vector and mentally assigning them to clusters”).
The August 2025 Memo admonishes examiners to consider whether an AI function may be performed in a human mind before declaring the patent an ineligible mental process, stating:
The mental process grouping is not without limits. Examiners are reminded not to expand this grouping in a manner that encompasses claim limitations that cannot practically be performed in the human mind. The MPEP and the AI-SME Update provide examples of claim limitations that cannot be practically performed in the human mind. Claim limitations that encompass AI in a way that cannot be practically performed in the human mind do not fall within this grouping.
USPTO, Memorandum to Tech. Centers 2100, 2600, and 3600, p. 2 (August 4, 2025) (emphasis added).
2. The Search for a Technological Improvement
According to the July 2024 guidance, an AI invention may be incorporated into a practical application by (1) linking any abstract idea to a “particular field of use,” within the claims (e.g., “field of network intrusion detection”); (2) presenting “ a technical explanation of [an] asserted [technological] improvement . . . in the specification” (e.g., “detect[ing] network intrusions and tak[ing] real-time remedial actions”); and (3) limiting the claims with non-abstract terms “to reflect the disclosed improvement” (e.g., dropping malicious packets in real-time after being identified by a neural network”). July 2024 Subject Matter Eligibility Examples, Example 47, claim 3, p. 12.
Practically, to comply with the 2024 guidance, a patent application should include a specification that frames the AI as an improvement over presently available technology. Additionally, the claims should be limited to a technological field and use non-abstract terms to recite further limitations implementing the improvement described in the specification. Examples of non-abstract terms in a machine-learning context can include digital images, neural networks, receipt\transmission of data, computer hardware, real-time automatic performance, or anything else that cannot be accomplished within a human mind.
Explaining that an invention improves a technology usually requires something to be said (or at least implied) of the current state of the technology being improved (e.g., what features does it lack, what performances are inadequate, and the like.) However, prior to the July 2024 guidance most practitioners were opining very little, in their patent applications, on technological background. Accordingly, many applications drafted before the USPTO’s July 2024 guidance may not explain an AI invention as being an improvement in AI technology.
The August 2025 Memo helps clarify that explicit explanation of a resulting technological improvement does not need to be in the specification for an abstract idea to be incorporated into a practical application. Specifically, the August 2025 Memo states: “[t]he specification does not need to explicitly set forth the improvement, but it must describe the invention such that the improvement would be apparent to one of ordinary skill in the art.” August 2025 Memo, p. 4.
Additionally, the August 2025 Memo explains that incorporation of an abstract idea into a practical application may be evidenced by claims that are limited to a particular solution. These are contrasted with claims that seek to cover every possible AI solution to achieve a desired outcome and are usually not incorporated into a practical application. Id. (“[a]n important consideration in determining whether a claim improves technology or a technical field is the extent to which the claim covers a particular solution to a problem or a particular way to achieve a desired outcome,” emphasis added).
3. Training a Neural Network is Non-Abstract (again)
According to USPTO guidance from January 2019, training a neural network was considered subject matter eligible, because it was non-abstract. Specifically, the guidance asserted training a neural network was neither (1) a mathematical concept “because the mathematical concepts are not recited [as mathematical relationships, formulas, or calculations] in the claims [when training is recited at a high level]”; nor (2) a mental process because “the steps are not practically performed in the human mind.” USPTO, Subject Matter Eligibility Examples: Abstract Ideas, Example 39, p. 9, (January 2019). A little over five years later, the USPTO guidance states that training a neural network, using a selected algorithm, such as “a backpropagation algorithm and a gradient descent algorithm,” is ineligible as a mathematical calculation. July 2024 SME Examples, Example 47, claims 2 and 3, p. 6.
While the 2019 and 2024 guidance on training a neural network is not technically contradictory, the change in guidance has had large practical consequences for those who prepare and prosecute AI patent applications at the USPTO. Before the July 2024 guidance, training a neural network was generally considered non-abstract and eligible at the USPTO, and, after the July 2024 guidance, training a neural network is much more likely to be considered abstract and ineligible at the USPTO.
Apparently in response to this effective change in practice at the USPTO, the August 2025 memorandum clarifies the distinction between training in Examples 39 and 47. August 2025 Memo, p. 3. With reference to non-abstract language in Example 39, the August 2025 Memo states:
Even though “training the neural network” involves a broad array of techniques and/or activities that may involve or rely upon mathematical concepts, the limitation does not set forth or describe any mathematical relationships, calculations, formulas, or equations using words or mathematical symbols.
Id. Explaining why the training language in claim 2 of Example 47 is abstract, the August 2025 Memo states:
Contrast this with the limitation “training, by the computer, the ANN based on the input data and a selected training algorithm to generate a trained ANN, wherein the selected training algorithm includes a backpropagation algorithm and a gradient descent algorithm” of claim 2 of example 47. This limitation requires specific mathematical calculations by referring to the mathematical calculations by name, i.e., a backpropagation algorithm and a gradient descent algorithm, and therefore recites a judicial exception, namely an abstract idea.
Id. Practically, going forward, we expect patent language that claims training a neural network will be held non-abstract and subject matter eligible, as was routinely the case before the July 2024 guidance.
4. A “Two-Consideration” Test to Avoid Oversimplification in “Apply It” Scenarios
“Examiners are cautioned not to oversimplify claim limitations and expand the application of the “apply it” consideration. Moreover, examiners are reminded that the “apply it” consideration often overlaps with the improvements consideration.” Id., p.4. Just like the analysis of the “mental process grouping,” which is “not without limits,” the Deputy Commissioner has reiterated two considerations when evaluating an “apply it” scenario, namely, (1) “[w]hether the claim recites only the idea of a solution or outcome, i.e., the claim fails to recite details of how a solution to a problem is accomplished, or the claim covers a particular solution to a problem or a particular way to achieve a desired outcome[ and, (2) w]hether the claim invokes computers or other machinery merely as a tool to perform an existing process, or whether the claim purports to improve computer capabilities or to improve an existing technology (emphasis added).” Id.
To forestall anticipated subject matter eligibility rejections in AI-powered or AI-assisted inventions that address a specific industrial problem, patent practitioners should guide examiners through the “two-consideration.” For instance, explanation of the specific problem and how a solution to the problem is accomplished in the Specification and recitation of an extra step for the execution of the solution in the Claims ties the AI-related steps recited in the claims to a practical effect on a technical problem that is supported by the Specification.
5. Preponderance of the Evidence is the Standard Proof in Subject Matter Eligibility Determinations
The Memorandum further reminds Examiners that the standard of proof for assessing a claim’s patent eligibility is a preponderance of the evidence. Id., p. 5. Particularly, the memo relates:
A rejection of a claim should not be made simply because an examiner is uncertain as to the claim’s eligibility. In order to make a rejection of a claim under any of the statutory bases (i.e., 35 U.S.C. 101, 102, 103, 112), unpatentability must be established by a preponderance of the evidence.
Id.
The preponderance of the evidence standard is a well-established evidentiary standard for civil cases. See John Leubsdorf, The Surprising History of the Preponderance Standard of Civil Proof, 67 Fla. L. Rev. 1569, 1583 (2016). A common expression of the preponderance of the evidence standard is: “when a party has the burden of proving any claim. . . by a preponderance of the evidence, it means you must be persuaded by the evidence that the claim . . . is more probably true than not true.” Manual of Model Jury Instructions for the District Courts of the Ninth Circuit, 1.6 Burden of Proof—Preponderance of the Evidence (2022).
The memo admonishes Examiners to not make rejections, when they have substantial doubts as to the eligibility of a claim:
[e]xaminers are reminded that if it is a “close call” as to whether a claim is eligible, they should only make a rejection when it is more likely than not (i.e., more than 50%) that the claim is ineligible under 35 U.S.C. 101. A rejection of a claim should not be made simply because an examiner is uncertain as to the claim’s eligibility. In order to make a rejection of a claim under any of the statutory bases (i.e., 35 U.S.C. 101, 102, 103, 112), unpatentability must be established by a preponderance of the evidence.
August 2025 Memo, p. 5 (emphasis added).
The reiteration of the “burden of proof” is an indication urging examiners to strictly adhere to the MPEP and avoid deviation or inconsistency in their decisions with respect to “close calls” in AI inventions due to the high-volume influx of AI-related patent applications submitted to the USPTO and the significant and fast-paced evolution of AI/ML inventions. Pairolero, N.A., Giczy, A.V., Torres, G., Islam Erana, T., Finlayson, M.A., & Toole, A.A. The artificial intelligence patent dataset (AIPD) 2023 update, J. Tech. Transfer (2025).
While, formally, this does not represent a change to the standards deciding claim eligibility (the standard has been established as a preponderance of the evidence.) Manual of Patent Examining Procedure, § 706(I). This clarification from the Deputy Commissioner for Patents may result in more favorable outcomes for inventors seeking to patent AI or AI-related inventions, because examiners are being instructed not to declare claims ineligible just because they think that they might be ineligible. Instead, considering the evidence, it must be more likely than not that the claims are ineligible.
6. Conclusion
AI and ML inventions present unique challenges in the patent system, but also great opportunities for innovators who understand how to navigate the process. In the past year, the USPTO has greatly explained how patent examiners may analyze AI and ML patents for subject matter eligibility, reducing uncertainty in the patent process for AI and ML innovators.
As patent practitioners, we are cautiously optimistic as the Patent Office tries to procedurally ensure consistent outcomes under 35 U.S.C. 101 analyses. As such, we also suggest that the relevant language of this memo, when it applies, should be cited in upcoming office action responses and appeal briefs.
This publication is distributed with the understanding that the author, publisher, and distributor of this publication and/or any linked publication are not rendering legal, accounting, or other professional advice or opinions on specific facts or matters and, accordingly, assume no liability whatsoever in connection with its use. Pursuant to applicable rules of professional conduct, portions of this publication may constitute Attorney Advertising. The choice of a lawyer is an important decision and should not be based solely upon advertisements.