AI and the ARP: Patent Eligibility under Ex Parte Desjardins

Are inventions in Artificial Intelligence (“AI”) eligible for patent protection?  A new front has emerged in that question.  In response to a Request for Rehearing, an Appeals Review Panel (the “ARP”) has vacated an eligibility rejection issued by the Patent Trial and Appeal Board (the “PTAB”) on a patent application claiming methods of training a machine learning model.  The ARP was apparently led by John Squires, the recently appointed Director of the USPTO, and includes the following passage.

This case demonstrates that §§102, 103 and 112 are the traditional and appropriate tools to limit patent protection to its proper scope.  These statutory provisions should be the focus of examination.

This statement is not explicitly limited to PTAB proceedings, and by implication applies to the examining corps in the USPTO.  Based on the omission of §101 from the recited statute sections, this statement may be construed to be a prohibition on the issuance of rejections based on eligibility.  Even as a routine decision, the ARP decision may be available to patent prosecutors facing eligibility rejections. 

The ARP decision cites Claim 1 for a step that “optimize[s] performance of the machine learning model” as evidence of improvement in the functioning of a computer, a defense to noneligibility.  In drafting software and AI patent applications, patent practitioners should continue to include such performative language in the claims, and support same in the specification, even without exhibit data.

If the USPTO ceases examination of patent applications for eligibility, a tsunami of software and diagnostic applications will likely proceed to grant (which applications would otherwise have gone abandoned, or never have been filed).  Until or unless Congress or the courts act to change the eligibility law to better conform to the Commissioner’s view of “extended eligibility”, this may result in a significant number of granted patents under clouds of ineligibility-based invalidity accumulating in the portfolios of patentees.  Several questions arise under this scenario: 

• If public companies disclose such patents, would prudence dictate disclosure of the accompanying clouds of invalidity?  
• Will accounting standards be modified to recognize the potential impairment of the book value represented by the clouds of invalidity of such patents?

Will the development and spread of Artificial Intelligence compel the courts to adopt a broader view of eligibility by implementing commercial success as a de facto determinant in the invalidity analysis?

Notes:

1. The Supreme Court has long held that 35 USC § 101, which defines the subject matter eligible for patent protection, contains an implicit exception for laws of nature, natural phenomena, and abstract ideas, the last of which is understood to include algorithms.  The Court fashioned what has become known as the two-part “Alice test”, asking: 1) whether the claims at issue are directed to a patent-ineligible concept; and, 2).if so, whether the claim’s elements, considered both individually and as an ordered combination, transform the nature of the claim into a patent-eligible application.  See, Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 573 U.S. 208 (2014).  The second prong of this test, known as the “inventive concept” prong, has been the basis for a plenitude of findings of invalidity for ineligibility of patent applications and granted patents directed to software, financial technology, and AI.  Alice has been cited in 1935 federal cases, 258 of which were before the Federal Circuit.

2. U.S. Ser. No. 16/319,040, “Training Machine Learning Models on Multiple Machine Learning Tasks”, assigned to Google LLC.  First named inventor Guillaume Desjardins is an active researcher in machine learning, and is in the DeepMind group, which was acquired by Google in 2014.  Among DeepMind’s achievements are breakthoughs in predicting protein folding, and beating the world champion of Go.

3. Ex Parte Desjardins, USPTO Appeals Review Panel of the Patent Trial and Appeal Board (Sept. 26, 2025, the “ARP decision”).
4. ARP decision, p. 10.

5. See Manual of Patent Examination and Practice (“MPEP”), §707.06, “Citation of Decisions, Orders Memorandums, and Notices”, [R-11.2013].  This section provides for the citation of a decision of the PTAB (and presumably ARP decisions thereon) which has not been published but which is available to the public.
6. MPEP, §§ 2106.04(d)(1) and 2106.05(a), citing Enfish, LLC v Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016).  The ARP decision makes much of Enfish.
7. FASB ASC 805 requires recognition and fair value measurement of identifiable intangible assets (including patents) acquired in a business combination.  ASC 350 Governs subsequent accounting for intangible assets like patents, including amortization and impairment testing.
8. Did the Supreme Court apply an unannounced fifth fair use factor of commercial success in finding that Google’s copying of 11,500 lines of Oracle’s Java SE code was transformational, where Oracle’s use of its code was as an Application Programming Interface (“API”) for mainframe computers, and Google’s was as the Android operating system (for a hand-held computer)?  See Google LLC v. Oracle Am., Inc., 593 U.S. 1 (2021).  By the time the case was heard at the Supreme Court in 2021, over 6 billion Android handsets were in use, over 63 % global market share of all active smartphones.  Query the consequence of a finding of copyright infringement in the form of royalties on 6 billion consumer products.  By the time the Supreme Court is presented with a case addressing the patent eligibility of AI inventions, what will be the size of the AI market, and what would be the impact of an adverse finding?  Perhaps the Commissioner is pursuing a “facts on the ground” strategy.