The explosive growth of artificial intelligence has created unprecedented challenges for patent protection. While AI may be "the most transcendent and transformative technology of our time," as USPTO Director John Squires remarked to the American Intellectual Property Law Association’s Annual Meeting in Washington in October, securing patent rights for AI innovations requires navigating increasingly complex legal terrain. For in-house IP teams and technology leaders, understanding these evolving requirements is essential to building a robust and defensible patent portfolio. The fundamental issues remain familiar, consistent with the timeless French observation "plus ça change, plus c'est la même chose," but their application to AI demands heightened sophistication. My view: submitting focused claims and a rich, detailed specification is the prescription for success.

Subject Matter Eligibility: A Minefield for AI Patents

Viable AI-related patent applications must meet today’s legal requirements for eligible subject matter, which has evolved considerably since the Supreme Court’s 2014 decision in Alice v. CLS Bank. The USPTO's Subject Matter Eligibility Guidance expresses the USPTO’s view of the applicable legal tests. Its 2024 examples 47-49 offer somewhat narrow signposts for what can establish eligibility for AI-related inventions. These examples highlight potentially successful approaches: reciting a model structure implemented in an ASIC, methods combining training with post-predictive actions (such as dropping network packets), speech waveform signal extraction using deep neural networks, and disease treatment methods that use trained ML models to identify high-risk patients coupled with specific treatment protocols like administering eye drops.

However, relying on USPTO guidance alone could prove costly in litigation. The September 2025 decision of the Court of Appeals for the Federal Circuit in Rideshare Displays v. Lyft explicitly declined to adopt the USPTO's guidance in a revealing footnote, stating unequivocally that the guidance "is not binding on this Court." This decision highlights the ongoing tension between the USPTO's efforts to establish more predictable and patent-friendly eligibility standards and the Federal Circuit's adherence to its own precedent in interpreting Supreme Court doctrine.

The April 2025 Recentive v. Fox Corp. decision further constrains, at least for now, the boundaries around viable AI patent claims. (Recentive petitioned the Supreme Court to overturn the decision in late October 2025, and a responsive brief is due in late November; Baker Botts represents Recentive in this effort.) The Recentive court reviewed two "machine learning training patents" and two "network map" patents, making clear that "wrapping" a known process in AI terminology, even in length claims, won’t establish eligibility. The claimed methods must genuinely be new and different, not merely conventional processes dressed up in machine learning language. Recentive’s Supreme Court petition counters that the Federal Circuit failed to consider that the Recentive claims don’t monopolize the fundamental ideas underlying the claims, and that the court categorically excluded patents for AI applied to new training domains.

Encouragingly, the Patent Trial and Appeal Board has begun shifting its approach. Starting in October 2025, the PTAB dramatically reduced its issuance of new grounds for rejection under Section 101 while increasing reversals of examiner 101 rejections at notably high rates. This trend suggests a more patent-friendly environment at the administrative level.

Nevertheless, patent prosecutors must draft with an eye toward potential litigation, going well beyond USPTO guidance to establish a solid foundation for eligibility findings in future disputes. The Patent Eligibility Restoration Act (PERA) of 2025 looms in the background—a six-year legislative effort with hearings held in mid-October 2025—but practitioners cannot wait for legislative solutions. Consider leveraging AI tools like Harvey.AI, subject to your institution's governance policies, to strengthen Section 101 arguments by analyzing applications against MPEP Section 2106 and Federal Circuit precedent.

Enablement: The EPO Sets an Increasingly High Bar

Enablement requirements for AI patents have become substantially more demanding, particularly following the European Patent Office's guidance, which is influencing global practice. Successful multilateral patent applications must be drafted with the EPO's guidance in mind. Even if a US application isn’t later filed in the EPO, its standards may presage future USPTO requirements.

The February 2024 EPO examination guidelines establish stringent disclosure expectations. Section F-III, 3 states that a patent disclosure is insufficient when "the mathematical methods and the training datasets are disclosed in insufficient detail to reproduce the technical effect over the whole range claimed." Such inadequate disclosures may be characterized as merely "an invitation to a research programme" rather than an enabling disclosure. Section G-II, 3.3.1 clarifies that while technical effects from machine learning algorithms may be established through explanations, mathematical proofs, or experimental data, the characteristics of training datasets required to reproduce these effects must be disclosed unless the skilled person can determine them using common general knowledge. Importantly, there's generally no requirement to disclose the specific training dataset itself.

The October 2024 decision in T 1669/21 under Article 83 EPC provides concrete guidance on EPO expectations for "sufficiency" or reproducibility. The Board of Appeals established that detailed disclosure of ML models is crucial—simply mentioning the use of machine learning is woefully insufficient. Applications must provide comprehensive descriptions of model architecture, including topology, mathematical modeling of nodes, and learning procedures. Name-dropping popular frameworks without explaining specific implementation fails this standard.

The decision further requires a clear definition of input and output variables, rejecting broad terminology in favor of specific, well-defined variables with concrete examples. Guidance on parameter selection within broadly defined categories must be provided, with working examples serving as a strong starting point. Perhaps most significantly, the Board criticized the lack of information about training data, requiring applicants to address the source and characteristics of training data across the full relevant parameter space.

Best Practices for AI Patent Disclosures

Given these heightened requirements, best practices for AI patent disclosures should include several key elements. For an AI architecture description, explain the model's structure, including its layers and connections. This requirement can often be satisfied by identifying a specific open-source package (such as TensorFlow, PyTorch, or specific model architectures) along with the configuration parameters used.

For training disclosures, describe the training data, process, hyperparameters, and optimization techniques. While there's no requirement to disclose entire training datasets, applications should provide an example record structure and identify 20-50 key features or columns that characterize the training data. Optional performance metrics, which provide quantitative measures of the AI solution's effectiveness, can strengthen the disclosure, although they're not strictly required.

For applications involving large public language models, consider including prompt structures, example prompts, context data, and retrieval-augmented generation (RAG) query approaches. These disclosures strike a balance between the need for enablement and legitimate trade secret concerns—a tension that looms large in AI patent strategy.

After Recentive, assume examiners will be skeptical and seek to characterize innovations as non-inventive applications of AI to specific problem domains. That said, there's no requirement to disclose model validation or performance data, nor is there a requirement to prove data integrity. Ultimately, no special rule exists specifically for AI disclosures. The fundamental principle remains: avoid purely functional disclosures and provide sufficient information about the hardware and software environment for a skilled person to implement the invention, taking into account the high level of skill in this art.

What Innovations Qualify as Patentable?

Not all AI improvements qualify as inventive subject matter, making the distinction between patentable innovations and trade secrets crucial for portfolio strategy. Consider four common disclosure scenarios:

First, claims asserting "it works better because we have access to training data that no one else has" are unlikely to be regarded as inventive. This competitive advantage is better protected as a trade secret than pursued through patents.

Second, discovering and claiming "the top N features in our training data that are most predictive of correct output classification" can be regarded as inventive if the features are properly recited in the claims. US Patent No. 10,860,961 provides an example of this approach.

Third, innovations where "core AI algorithms are entirely new" are likely to be regarded as inventive. Examples include specific forms of deep learning classifiers, transformer architectures with Low-Rank Adaptation heads, and novel parameter changes or neural network configurations that represent genuine architectural innovations.

Fourth, innovations that "reduce the memory and storage needed to contain and run deployed models" are likely to be regarded as inventive, addressing the practical challenges of model deployment and efficiency.

Recent successful patents illustrate these principles in practice. US Patent No. 12,353,827 covers the improvement of LLMs via non-LLM systems. US Patent No. 11,886,800 protects specific innovations in transformer model architecture. US Patent No. 12,379,948 addresses agentic AI, covering the integration of specific AI agents into larger processes with defined roles and coordination mechanisms.

The rapid evolution of generative AI creates new patenting opportunities beyond traditional machine learning. Prompt engineering and context formation in RAG-based systems represent a fertile area, particularly in the structure of templates for prompts and context, and their use in solving domain-specific problems. These innovations bridge the gap between general-purpose LLMs and specialized applications.

Graphs of AI agents present another opportunity, including the definition of agent roles, the prompts or queries they use to access databases or LLMs, and the criteria for transitioning from one agent to another based on per-agent analysis of LLM responses. These multi-agent systems represent genuine architectural innovations that solve coordination and orchestration challenges.

Ethics and Compliance: Navigating USPTO Requirements

The USPTO issued guidance (though not formal rulemaking) on the use of AI tools in drafting submissions in April 2024 (89 Fed. Reg. 25609). This guidance, grounded in duties of candor and good faith, signature requirements and corresponding certifications, confidentiality obligations, foreign filing licenses and export regulations, USPTO electronic systems policies, and duties owed to clients, establishes clear expectations for practitioners.

When making submissions to or corresponding with the USPTO, a practitioner's signature certifies that all statements are true and that reasonable inquiry was performed. The practitioner must have reviewed and verified the paper and its contents—delegation to an AI tool without human verification violates these requirements.

If the use of an AI tool is material to patentability, disclosure to the USPTO is required. This disclosure enables examiners to determine whether human inventors contributed to conception or whether inventor designations should be modified. It also guards against presenting unpatentable claims and helps limit excessive Information Disclosure Statement submissions.

Perhaps obviously, AI systems cannot obtain uspto.gov accounts and cannot be sponsored as support staff individuals. (One wonders if the USPTO saw someone attempt this.) Practitioners must guard against AI systems retaining client confidential information or data subject to export or national security controls as training data. Non-compliance can lead to disciplinary action on grounds of fraud or intentional misconduct.

Notably, the USPTO has not taken a position on whether practitioners have a duty to disclose, as prior art, any aspect of training data that a generative AI tool used to generate part of a patent application. This remains an open question that careful practitioners should monitor.

Strategic Recommendations for In-House Teams

As AI patent law continues evolving, successful protection strategies require several key commitments. First, invest in sophisticated drafting that goes beyond USPTO guidance to address Federal Circuit precedent. Second, maintain awareness of international standards, particularly EPO requirements that may influence future US practice, and consider requiring US applications to conform to those standards of description. Third, carefully balance trade secret protection with patent disclosure requirements, recognizing that not all AI innovations are best protected through patents. Fourth, implement robust AI tool governance policies that address confidentiality, verification, and disclosure obligations. Finally, conduct thorough inventor interviews to ensure proper inventorship determination and complete disclosure of AI tool involvement in conception and reduction to practice.

The intersection of AI and patent law presents both challenges and opportunities. Organizations that master these complexities will build stronger, more defensible patent portfolios in this transformative technological era.