In the United States, eligibility under 35 U.S.C. § 101 is governed by the Alice/Mayo¹¹ two-step test, which asks whether a claim recites a judicial exception and, if so, whether it integrates that exception into a practical application or includes an inventive concept amounting to significantly more. Recent USPTO memoranda and decisions signal an intention to reduce difficulties for getting AI innovations through examination.

The August 2025 guidance: key changes for AI applicants

In August 2025, the USPTO issued a memorandum on assessing eligibility for AI and machine learning inventions. The memo distinguishes between claims that "set forth or describe" a judicial exception and claims where AI is merely involved without the claim describing the exception - the latter of which are eligible without further analysis. Examiners are reminded not to expand the mental process exception to limitations that cannot practically be performed in the human mind. Claims reflecting an improvement to computer or technology functioning may qualify as a practical application, and the improvement need not be explicitly stated in the specification if apparent to one of ordinary skill. The memo cautions against oversimplifying the "apply it" analysis; examiners should not dismiss claims merely because they involve a computer implementing an algorithm.

Examiners should reject only if it is "more likely than not" (>50%) that the claim is ineligible. Examiners should not reject simply because they are uncertain. The burden rests on the USPTO to establish ineligibility by a preponderance of evidence. For AI inventions presenting close questions, the benefit of the doubt now goes to the applicant.

Ex parte Desjardins (September 2025)

An Appeals Review Panel (ARP) decision led by the USPTO Director in Ex parte Guillaume Desjardins et al. (Sept. 2025)¹² held that claims directed to a method of training a ML model on multiple tasks were eligible under § 101. The invention, filed by Google’s DeepMind Technologies, was directed to the problem of “catastrophic forgetting” in ML models, where models lose prior task knowledge when learning new ones. The ARP decision vacated the PTAB’s § 101 rejection, finding that the claims sufficiently reflected improvements in how the ML model operated. It also found the PTAB’s rejection “troubling”, noting that excluding AI innovations from patenting jeopardizes America’s leadership in AI technology.

AI inventorship (November 2025)

The USPTO's November 2026 memo on "Revised Inventorship Guidance for AI-Assisted Inventions" addresses who qualifies as the inventor when AI contributes to the inventive process and requires a natural person to make a significant contribution to conception while recognizing AI tools may play a substantial role. Notably, the guidance treats AI systems as analogous to any other tool that assists in the inventive process and that inventors “may use the services, ideas, and aid of others” without those sources becoming co-inventors.

Using declarations to overcome § 101 Rejections (December 4, 2025)

On December 4, 2025, the USPTO Director issued a memo encouraging the use of Rule 132 declarations, ‘Subject Matter Eligibility Declarations’, or SMEDs, to overcome § 101 rejections. SMEDs are optional but allow applicants to submit objective factual evidence supporting eligibility. They must focus on facts (e.g., performance metrics, benchmarks, design differences), not legal conclusions. Effective SMEDs address the state of the art at filing, technical improvement evidence, implementation practicality (e.g., that claimed steps cannot be performed mentally or involve specific data structures), and claim nexus (connecting evidence to specific claim language). Examiners must assess SMEDs under the preponderance-of-the-evidence standard and document their impact in the next Office Action. For AI inventors facing § 101 rejections, SMEDs are a practical tool to build an evidentiary record of real technical improvement.

MPEP updates following Ex parte Desjardins (December 5, 2025)

The following day, the USPTO issued updated MPEP guidance in light of Ex parte Desjardins, broadening the improvements analysis to cover "an improvement in technology or a technical field", not only computer functionality. New AI-specific examples of qualifying improvements include training a machine learning model in a way that protects knowledge of previous tasks while learning new ones and improving system performance through adjustments to ML model parameters. This signals that training methodology and model optimization may be recognised bases for eligibility.

Canada vs. the US: where does each regime leave AI inventors?

While both regimes seek to avoid granting patents on abstract ideas, they diverge in how that principle is applied in practice. In Canada, the Schlumberger question asks whether the algorithm improves the computer's functioning; if not, and the only physical element is a generic computer, the claim fails. In the United States, the August 2025 "close calls" standard resolves uncertainty in the applicant's favour. Critically, the USPTO improvements analysis considers improvements to technology generally, not only to the computer's own functioning which is a broader standard than Canada's Schlumberger question.

The United States currently offers a more permissive environment for AI patents. The SCC's pending Pharmascience decision and US legislative reform of § 101 may further reshape the landscape.

Takeaways for AI inventors

Across jurisdictions, patent eligibility of software and AI continues to hinge on whether the invention is characterized as more than an abstract algorithm, but inventors should be aware that different jurisdictions apply this threshold in different ways. Inventors considering patent protection should make note of any technological improvements or physicality provided for in their AI inventions.

‍