USPTO Director Squires issued a precedential decision, further opening the door for patents on AI inventions. Here’s what this decision means.
Prosecuting AI patent applications is often frustrating. Examiners have been routinely dismissing machine learning claims as “abstract” because they are “algorithms” that “merely” run on “generic computer components.” Fortunately, that practice started to shift on September 26, 2025, when newly confirmed USPTO Director John Squires issued an Appeals Review Panel (ARP) decision in Ex parte Desjardins, Appeal 2024-000567. This Director Squires designated the case as precedential on November 4, 2025. He also incorporated the case into the examiner’s official instructions (MPEP). This is now binding on all examiners and PTAB panels.
What Happened
Inventor Desjardins, working for Google, filed application 16/319,040 for a Google DeepMind project. The application claimed a method for training a machine learning model to learn new tasks sequentially while preserving knowledge from earlier tasks — addressing “catastrophic forgetting.” The specification identified concrete benefits: reduced storage, lower system complexity, and preserved performance across sequential training. The patent application subsequently issued in Europe, China, and elsewhere, but hit a wall in the US.
During US prosecution, the US examiner rejected the claims as obvious under §103, but did not raise any “abstract” (101) rejections. Google appealed the obviousness rejection to the USPTO Patent Trial and Appeal Board (PTAB).
Things went sideways. The PTAB raised a new §101 rejection. They took it upon themselves to treat the claims as an abstract mathematical algorithm. In July 2025, the PTAB denied rehearing, citing the Federal Circuit’s Recentive Analytics v. Fox decision. They argued that applying generic ML (Machine Learning) to new data environments is patent-ineligible.
The Director’s Appeals Review Panel (ARB)
USPTO Director Squires pulled the case for a Director’s Appeals Review Panel (ARP) review and reversed. Recall that the patent eligibility (Alice) process is multi-step. At Alice Step 2A, Prong One, the ARP agreed with PTAB that the claims recite a mathematical concept. But at Prong Two, relying on Enfish v. Microsoft, the ARP disagreed with PTAB. The ARP found that the claims were directed to an improvement in how the model itself operates. This was not merely an application of math. The claim limitation requiring parameter adjustment to optimize a second task while protecting performance on the first was the key.
Director Squires essentially took PTAB out to the woodshed. He called the PTAB’s approach “troubling” and warned:
Categorically excluding AI innovations from patent protection in the United States jeopardizes America’s leadership in this critical emerging technology.
— Ex parte Desjardins, ARP Decision, p. 9
Director Squires further criticized the PTAB for equating all machine learning with unpatentable algorithms. He also cautioned against evaluating claims “at such a high level of generality” that meaningful limitations get dismissed. He further emphasized that §§102, 103, and 112 are the “traditional and appropriate tools” for limiting patent scope. (Sadly for the US version of Desjardins, the original US obviousness §103 rejection still stood. So Squires was promoting machine learning patents generally, but not this one!)
The Aftermath
The USPTO examiner’s instructions (MPEP) were updated to reference this Ex Parte Desjardins case. (It’s not in the main book yet, but the change order has been posted.) Going forward, teaching how a “machine learning model is trained to learn new tasks while protecting knowledge about previous tasks” can be patent-eligible. This will join the previously discussed Enfish and McRo (Planet Blue) decisions, expanding software patent eligibility.
US Courts to USPTO: “Nuts to you, the law is what we say it is!”
Unfortunately, the Desjardins decision binds only the USPTO, not the US courts. The Federal Circuit’s Recentive decision remains good law. It may become easier to obtain AI patents, but enforcing them in litigation could still face §101 challenges.
Recommendations
- Identify a specific technical problem in the specification. Don’t just describe what your model does, describe what it solves and why existing approaches fail.
- Describe the improvement in detail using non-conclusory language: reduced storage, improved efficiency, prevention of specific failure modes.
- Ensure the claims reflect the improvement. The ARP emphasized that specification assertions alone are insufficient; the claim language must embody the disclosed improvement.
- Frame claims as improvements to the model, not applications of it.
- Think about filing in Europe and China; they may be more open-minded.
Like Director Iancu before him, present USPTO Director Squires has made it a bit easier to obtain AI patents. But the reform process is not over yet. Congress still needs to fix §101 legislatively (PERA). In the meantime, consider following the Desjardins playbook (e.g., technical problem, technical solution, claims that reflect the improvement).