Artificial Intelligence-related patent applications have faced heightened scrutiny under U.S. patent eligibility law (35 U.S.C. § 101). The precedential PTAB decision in Ex Parte Desjardins, followed by USPTO memoranda, signals a policy clarification that Artificial Intelligence inventions may be patent-eligible, including inventions that show technical improvements to machine learning systems themselves. The change is presented as reducing uncertainty for companies investing in Artificial Intelligence research and development while offering a clearer roadmap for drafting and prosecution.
The Desjardins decision was made precedential by Acting USPTO Director John Squires, requiring Examiners and PTAB judges to apply the policy. A statement in the decision says, “Categorically excluding AI innovations from patent protection in the United States jeopardizes America’s leadership in this critical emerging technology” (Ex parte Desjardins, Appeal No. 2024-000567 p. 9, (PTAB Decision September 26, 2025)). Two memos issued on December 4 and 5, 2025, identify patent-eligible improvements such as “reduced storage,” “reduced system complexity,” “streamlining, and preservation of performance attributes associated with earlier tasks during subsequent computational tasks” and “improvements in computational performance, learning storage, data sets, and structures.”
The December 5 memo sets out a drafting-first approach described as Plan A. Its core principle is that a patent application should disclose a technical improvement in the specification and ensure the claims reflect that improvement. The guidance states that “the specification identified the improvement to machine learning technology … and that the claims reflected the improvement identified in the specification” (MPEP § 2106.04(d)(1), second paragraph). That means the application should include technical detail, such as architecture, parameter updates, data structures, or algorithmic steps, rather than conclusory claims that an invention is simply better. If the improvement is a specific neural network architecture that reduces memory usage, the claims must recite that architecture rather than only its intended outcome. The memo also added two examples to the MPEP covering improved training that preserves knowledge from prior tasks while learning new ones, and improvements to computer component or system performance based on adjustments to machine learning model parameters associated with tasks or workstreams.
The December 4 memo provides a fallback path called Plan B through subject matter eligibility declarations, or SMEDs. A SMED is a sworn expert declaration explaining why a practitioner of ordinary skill would understand the invention as a technological improvement. The memo states that “a SMED may demonstrate how one of ordinary skill in the art would interpret a specification that describes a technological improvement to show that the claimed invention is patent eligible subject matter” (Dec. 4 Memo, p. 3). A SMED cannot add new technical disclosure, but it may clarify how an existing specification would be understood by a POSA. Examples include evidence showing that claim limitations cannot practically be performed in the human mind, objective comparative testing showing superior performance of a claimed neural network architecture, and expert testimony supported by trade magazine articles establishing the state of the art at filing. Together, Plan A and Plan B create a more structured framework for pursuing patent protection for Artificial Intelligence inventions.
