Last week, USPTO released its decision to make a rare appeal review panel (ARP) Ex Parte DesjardinsAppeal 2024-000567, September 26, 2025 (“ARP Decision”), reversed the discovery of ineligible subjects in patent applications directed at training machine learning models. Written in his first week of office by new USPTO Director John Squires, the decision appears to demonstrate the intent of the new USPTO leadership to limit the judges' reliance on subject eligibility, particularly when read in conjunction with the August 4, 2025 memo from Charles Kim's subcommittee.
The patent application in question (US Patent Application No. 16/319,040) was directed to training machine learning models to perform several different tasks. I have read the typical claims:
1. Computer-implemented methods for training machine learning models;
Machine learning models have at least several parameters;
Trained in the first machine learning task using the first machine learning data to determine the initial values of multiple parameters in the machine learning model,
Here's how:
Determine each scale for each of the multiple parameters
The importance of parameters for the first machine learning task, including:
Computing based on the first value of multiple parameters
Determined by training a machine learning model in the first machine learning
Approximation of the posterior distribution across possible values of tasks
Multiple parameters,
Use an approximation to assign the values for each parameter. The values are each measure of the importance of the parameters for the first machine learning task, and the first value of the post-training parameters after training in the first machine learning task is the correct value of the first training data given the first training data used to train the first machine learning model.
Get second training data to train a machine learning model in 1 second;
Various machine learning tasks. and
Train the second machine learning model with the second machine learning model, train the machine learning model with the second training data, adjust the first values of multiple parameters to optimize the performance of the machine learning model for the second machine learning task while protecting the performance of the machine learning model in the first machine learning task.
By adjusting the first value of multiple parameters, it involves adjusting the first value of multiple parameters to optimize objective functions that rely in part on penalty terms based on a critical measure of the importance of multiple parameters to the first machine learning task.
The patent examiner refused to argue that it was clear under 35 USC §103, but did not reject the claim that it was directed towards a patent-essential subject under 35 USC §101. At the appeals trial, the Patent Trials and Appeals Board (PTAB) confirmed the rejection under §103 and introduced a new basis for rejection under §101. Coach Squire exercised his authority to convene the ARP to reconsider the PTAB decision.
The ARP left the §103 rejection unharmed, but it invalidated the §101 rejection introduced by the PTAB. Apply the two-stage patent qualification enquiries listed in Alice Corporationv. CLSBank International573 US 208 (2014), and in the Patent Inspection Procedure (MPEP) manuals of §2106, ARP agreed with the PTAB that each claim recited a judicial exception, in particular “at least one abstract idea.” (ARP decided at 6-7).
Moving on to the next prong of the investigation, ARP found that the PTAB had made the error. This request has been to “integrate judicial exceptions into practical applications” as provided by MPEP §2106.04(ii)(a)(2). ARP was dependent Enfish, LLCv. Microsoft Corp.822 F.3d 1327, 1339 (Fed. Cir. 2016), “The Federal Circuit cited ARP's decision in 8, “whether claims are directed towards improvements to abstract ideas and improvements to abstract ideas,” and “determined that we should turn on whether claims are directed towards improvements to computer functions.” Enfish822 F.3d at1336. Claims Language Reciting Claims Language “We have protected the performance of the machine learning model for the first machine learning task by adjusting the first values of multiple parameters to optimize the performance of the machine learning model for the second machine learning task.”[] It is an improvement on how the machine learning model itself works, not, for example, a identified mathematical calculation. “ARP decided at 9am.
In particular, the ARP decision also included a language supervisory examiner and a PTAB panel on how to assess eligibility for a patent under §101.
Under the Charitable View, the overloaded reasoning in the original panel below is probably understandable given the nature of the confusion of existing §101 jurisprudence, as this case highlights what is on in crisis. Determiningly excludes AI innovation from US patent protections puts American leadership in this critical emerging technology at stake. However, under panel reasoning, many AI innovations are potentially innovated when not properly explained or potentially unpatentable. December 24th. Examiners and panels should not evaluate claims at such a high level of generality.
…
At the same time, the claim in question was rejected under §103. This case shows that § 102, 103, and 112 are traditional appropriate tools to limit patent protection to the appropriate scope. These statutory provisions should be the focus of the review.
ARP decision was made 9-10.
With the release of this decision, the new USPTO director appears to be informing of policy changes. In particular, improvements in the operation of machine learning models (and potentially other complex computer algorithms) should be considered patent-friendly improvements to the functionality of computer systems. § 102, 103, and 112 (the more general statements are that novelty, non-obviousness, and clarity of satisfaction and clarity disclosure and claim clarity could be carried over to other technologies.
With this ARP decision, the USPTO may find more claims under §101, but applicants should note that USPTO's decision is not binding on the court. The Federal Circuit recently repeated this point. RideShare Displays, Inc. v. Lyft, Inc.No. 2023-2033 (Fed. Cir. September 29, 2025) (Non-precedent).
Takeaway to evaluate inventions for patents:
-
The claim to recite improvements to the manipulation of machine learning models is more likely to see USPTO as eligible for patents.
-
Patent examiners can begin to focus more on novelty and non-obvious questions, focusing on subject eligibility. Applicants will find refusal less than 35 USC §101 easier to overcome. This change could extend beyond machine learning or computer implementation inventions.
-
The court is not bound by USPTO interpretation of the statute. While patents in some technology areas may be easier, enforcement may still remain a challenge.
