AI IP Year in Review – Navigating the Crossroads of AI Patent Eligibility: Translating USPTO Optimism into Long-Lasting Protection

Artificial intelligence (AI) has become an engine of growth in the modern economy. Innovations in AI-based technologies are reshaping how industries operate by integrating speed and efficiency into fields such as healthcare, financial services, and content creation. As these technologies increasingly define market compe­tition, protecting intellectual property (IP) is becoming an ever more urgent priority for AI companies. Yet, securing robust patent protection for AI inventions has proven to be a persistent area of difficulty given the nature of the technology and the industry.

Trade secret protection, which has traditionally served as the fallback method for guarding proprietary algo­rithms, source code, and training data, requires the implementation of ongoing confidentiality precautions. However, these measures have become increasingly difficult to maintain for AI companies given the collab­orative work patterns and high employee turnover that are typical of the tech industry. Compounding the problem, advanced AI systems themselves can some­times derive or “reverse engineer” proprietary meth­ods by analyzing model outputs, raising concerns that core system behavior may be partially inferable despite conventional barriers.

Patents often serve as the typical alternative to trade secrets for protecting scientific inventions, but they also face practical problems in tech. AI and machine learn­ing (ML) innovations frequently encounter rejections by examiners at the United States Patent and Trademark Office (USPTO) and federal judges under 35 U.S.C. § 101. This is primarily due to AI and ML’s close relationship with algorithms and computational processes, which courts often treat as unpatentable abstract ideas under § 101.

However, recent actions by the USPTO’s new lead­ership suggest a potential broadening of what qual­ifies as patent-eligible subject matter at the agency. The USPTO has indicated a more permissive stance toward AI-related inventions under the management of Director John Squires. This optimism regarding the patentability prospects for AI applicants is expressed most clearly in the Director’s recent sua sponte inter­vention in Ex Parte Desjardins, where he touted the eligibility of a patent application that claimed a method of training an ML model.

Still, optimism alone cannot substitute for enforceabil­ity. The value of a patent depends not only on whether it can be obtained but also on whether it will with­stand validity challenges years down the line. Because federal courts have the final word on patent validity, AI stakeholders must approach drafting and prose­cution with caution, even if the USPTO appears more welcoming on eligibility. This article examines (1) the USPTO’s shifting posture on AI patentability, (2) the continuing obstacles that remain for AI patents, and (3) strategies for drafting AI applications that can endure both USPTO and Federal Circuit scrutiny.

New Directions at the USPTO

Upon taking office, USPTO Director John Squires made clear that patent eligibility reform would be a priority. His early actions signal a desire to broaden protection for emerging technologies at the agency.

During his first week in the position, Director Squires ceremonially signed two patents in technology areas traditionally thought to be unpatentable (diagnostics and cryptocurrency) in an effort to underscore a policy interest in expanding eligibility.[1] He also provided testi­mony before the Senate in support of the Patent Eligibil­ity Restoration Act (PERA), where he framed the expan­sion of patent eligibility as a matter of national security, global leadership, and economic competitiveness.[2]

The Director’s most notable effort to elucidate his view on AI patentability was his decision in Ex Parte Desjardins.[3] Exercising the Appeals Review Panel in only its second use to date, the Director authored a panel opinion over­turning a decision by the Patent Trial and Appeal Board (PTAB) that had found an ML-related patent ineligible.

The application was directed to an ML model designed to address “catastrophic forgetting,” a degradation phenom­enon in continual learning systems where models lose performance on previously learned tasks when trained on new data. After finding the claims obvious under 35 U.S.C. § 103, the PTAB issued a new ground of rejection under § 101, concluding the claims were directed to abstract mathematical processes lacking a “practical application.”[4]

Director Squires disagreed, finding that adjusting model parameters to “optimize performance… while protect­ing performance on the first task” reflected a technical improvement to the operation of the ML model itself.[5] Because he found the claims integrated the abstract idea into a practical application and provided a concrete tech­nological solution, he reversed the § 101 rejection.

In doing so, the Director cautioned against “categorically excluding AI innovations from patent protection,” warn­ing that such an approach “jeopardizes America’s leader­ship in this critical emerging technology.”[6] These remarks signal a USPTO policy preference for treating § 101 as a coarse filter rather than a strict gatekeeping doctrine.

For AI innovators, this suggests a warmer reception at the examination stage. Yet even a favorable agency posture does not eliminate the broader legal headwinds facing AI patents.

Remaining Obstacles for AI Patents

AI patent applicants will continue to face signifi­cant hurdles despite a more receptive USPTO. Even while § 101 rejections may decrease in some art units, applicants will continue to face challenges under §§ 102, 103, and 112. Indeed, in Desjardins, although Director Squires reversed the § 101 rejection, he left the § 103 rejection in place. If the agency adopts the Direc­tor’s view that the remaining statutory bars to patent­ability are “the proper statutory tools for limiting the scope of patents,”[7] then the overall number of §§ 102, 103, and 112 rejections may actually increase, becoming the means by which examiners may invalidate claims that would traditionally have been viewed as patent ineligible under § 101.

More important, regardless of the Director’s views on eligibility, practitioners might not experience much change in practice if the Federal Circuit does not share the agency’s perspective. Following the Supreme Court’s decision in Loper Bright Enterprises v. Raimondo, courts are no longer obligated to defer to agency interpretations of law. As a result, the Federal Circuit will continue applying its own framework for eligibility, irrespective of USPTO policy preferences.

In April 2025, the Federal Circuit decided Recentive Analytics, Inc. v. Fox Corp., the court’s first § 101 decision focused squarely on machine-learning claims.[8] The court affirmed the ineligibility of two patents directed to ML-based TV scheduling. After characterizing the claims as directed to mathematical algorithms, the court held that requiring the model to be “iteratively trained or dynamically adjusted” did not convert the abstract idea into a patent-eligible technical improve­ment.[9] In the court’s view, this behavior was merely “incident to the very nature of machine learning.”[10] The court further clarified that computer-implemented speed improvements or efficiency gains must stem from enhancements to computational techniques or system operation, not simply from automating a process that could be performed mentally.

Although Recentive Analytics may temper enthusiasm generated by Desjardins, the decision also clarified that not all ML models are ineligible. Judge Dyk emphasized that the ruling addressed only claims that “apply estab­lished methods of machine learning to a new data envi­ronment,” signaling that claims describing concrete improvements to model architecture, system perfor­mance, or computational behavior remain viable under the Federal Circuit’s application of § 101.[11] This distinction creates a meaningful, but navigable, tension. While the USPTO may be willing to recognize the practical applications of AI models as sufficient for eligibility, the Federal Circuit requires that those appli­cations reflect concrete technical improvements. Practi­tioners must therefore draft with both audiences in mind.

Drafting Strategies for AI Patent Applications

The juxtaposition between Desjardins and Recentive Analytics provides guidance on how applicants can strengthen their AI claims. The following drafting prin­ciples can help bridge the gap between a more permis­sive USPTO and a more demanding Federal Circuit.

1. Start with a Technical Narrative

A strong AI patent begins with a clear technical story. From the outset, applicants should articulate the specific technical problem addressed by the inven­tion, whether tied to data structures, training behav­ior, inference latency, model stability, resource utili­zation, or deployment constraints. In finding the ML model from Desjardins to be patent eligible, Director Squires relied heavily on the application’s descrip­tion of reduced storage usage and decreased system complexity, treating these as indicators of a genuine technological improvement.

The specification should contextualize the improve­ment and avoid describing the invention solely at a high level. Clear articulation of how the model modifies gradient updates, alters feature-embedding behav­ior, reduces computational cost, or improves memory access patterns can strengthen § 101 positions and reinforce inventive concepts under §§ 102 and 103.

2. Embed Technical Detail and Avoid Result-Oriented Claiming

The claims in Recentive failed largely because they implemented generic machine learning processes in a new environment without altering the underlying tech­nology. Applicants should instead make AI central to the claimed invention by describing:

• Model architectures (e.g., encoder-decoder path, attention mechanisms, gating functions);
• Training pipelines and parameter update techniques;
• Feature-engineering strategies;
• Deployment-specific constraints (e.g., on-device inference, memory bounds, latency targets); and
• Real-world operational challenges (e.g., model drift, catastrophic forgetting, noisy labels).

Grounding claims in the “how,” not merely the “what,” helps differentiate AI inventions from abstract algorith­mic processing.

3. Address § 112 Head-on While Balancing Trade Secret Sensitivities

AI inventions commonly face written-description and enablement challenges, particularly after Amgen, Inc. v. Sanofi.[12] Applicants must provide enough techni­cal detail to show possession of the invention and to enable others to make and use it, but they must also be mindful that disclosures become public and may reveal aspects of the system that a company would prefer to maintain as a trade secret.

A balanced approach requires articulating the inven­tion’s core technical mechanisms without expos­ing proprietary implementation details that provide competitive advantage. Applicants should consider describing, at an appropriate level of abstraction:

• Inputs, intermediate representations, and outputs;
• Training datasets and preprocessing pipelines (e.g., categories or characteristics, not necessarily raw data);
• Iterative training behaviors and update logic;
• Inference workflows and data flow sequencing; and
• System-level interactions and deployment constraints.

Providing this detail not only satisfies § 112 but also helps demonstrate technical improvements under § 101, while thoughtful scoping of the disclosure helps preserve sensitive information that may be better protected through trade secrets.

4. Layer and Diversify Claims

Applicants should draft claims across multiple statu­tory classes, including method, system, non-transitory computer-readable medium, and component-level claims. Dependent claims can capture:

• Specific model components;
• Data preparation steps;
• Training environment features;
• Memory management techniques;
• Optimization strategies; and
• Deployment configurations.

Layered claiming increases the likelihood that at least one claim survives eligibility and prior art challenges.

***

The USPTO’s recent actions under Director Squires point toward a more flexible interpretation of § 101 for AI-related inventions. But the Federal Circuit’s decision in Recentive underscores that judicial scrutiny remains a critical barrier. AI inventors should be encouraged by the USPTO’s signals, but they must draft with the understanding that a patent examined today may be litigated years later before a court that applies § 101 more narrowly.

Long-term protection requires careful alignment of specification detail, technical storytelling, and layered claiming. Applicants who articulate concrete techno­logical improvements, rather than functional or result-driven descriptions, will be best positioned to secure durable protection for their most important AI assets.

[1] Dani Kass, Squires Jumps Right Into Patent Eligibility Reform, Law360 (Oct. 1, 2025)

[2] The Patent Eligibility Restoration Act–Restoring Clarity, Certainty, and Predictability to the U.S. Patent System: Hearing on S. 1546 Before the S. Subcomm. on Intell. Prop. of the S. Comm. on the Judiciary, 119th Cong. (2025) (statement of John A. Squires, Director of the United States Patent and Trademark Office).

[3] Ex Parte Desjardins, 16/319,040, 2025 WL 3095778 (P.T.A.B. Sept. 26, 2025).

[4] Id. at 4.

[5] Id. at 9.

[6] Id.

[7] The Patent Eligibility Restoration Act–Restoring Clarity, Certainty, and Predictability to the U.S. Patent System: Hearing on S. 1546 Before the S. Subcomm. on Intell. Prop. of the S. Comm. on the Judiciary, 119th Cong. (2025) (statement of John A. Squires, Director of the United States Patent and Trademark Office).

[8] Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025).

[9] Id. at 1212.

[10] Id.

[11] Id. at 1211.

[12] Amgen Inc. v. Sanofi, 598 U.S. 594 (2023).

This article appeared in the 2025 AI Intellectual Property: Analysis & Trends Year in Review report.