Surviving the new AI prosecution gauntlet

Navigating the AI patent landscape requires moving beyond § 101 eligibility to master the complex interplay with § 112(a) enablement, says Bradford Fritz of BSKB.

In 2023, the main hurdle when prosecuting patents for artificial intelligence (AI) and blockchain technologies was often surviving a 35 U.S.C. § 101 challenge. At the time, overcoming a § 101 rejection required a targeted strategy of moving away from generic hardware recitations and instead focusing on how the alleged abstract idea integrated into a “practical application.”

As I detailed in a previous WIPR article three years ago, amending a claim to recite a real-world result that improves upon existing technology is an effective way to overcome a § 101 rejection.

Fast forward to 2026, and the patent landscape has fundamentally shifted. Passing the subject matter eligibility threshold is now only part of the battle. Securing robust, enforceable AI patents demands mastering the delicate interplay between the established Alice framework and escalating § 112(a) enablement hurdles. 

Navigating this narrowing path successfully requires a robust disclosure that heads off potential pitfalls before they become fatal rejections or litigation challenges.

The diverging § 101 landscape: Examination vs litigation

For prosecuting AI inventions today, a noticeable divergence has emerged between the United States Patent and Trademark Office (USPTO) examination trends and strict litigation outcomes in the federal courts.

Fortunately, the USPTO is starting to push back on examiners relying on aggressive, boilerplate § 101 rejections for AI. Under the pro-innovation leadership of director John Squires, the agency has signalled a clear policy shift. In the 2025 precedential appeals review panel decision Ex parte Desjardins, the USPTO warned against overbroad subject matter eligibility rejections that equate any machine learning with an unpatentable algorithm. 

Instead, the panel instructed examiners to utilise §§ 102, 103, and 112 as the traditional, appropriate tools to limit patent scope. For applicants, Desjardins provides an important shield during examination, proving that AI claims improving the operation of the model itself are patent-eligible.

However, the courts present a stricter reality. In its landmark April 2025 decision, Recentive Analytics for the Federal Circuit held that simply applying generic machine learning techniques to a new data environment (in that case, event scheduling and network maps) fails § 101. To survive litigation, an invention cannot merely perform a known human task faster using existing AI. Rather, the claims must disclose an improvement to the underlying machine learning model itself.

This disconnect between the patent office and the courts is frustrating for practitioners. The Alice/Mayo framework frequently overlaps with traditional novelty (§ 102) and obviousness (§ 103) considerations, such as whether a technique is “conventional”.

While legislative efforts like the pending Patent Eligibility Restoration Act (PERA) aim to eliminate the judicially created “abstract idea” exception entirely, practitioners must currently draft applications capable of satisfying both the USPTO’s shifting standards and the Federal Circuit’s exacting scrutiny.

Leveraging example 39 as a useful tool for AI eligibility

While tying claims to a practical application or a real-world improvement is helpful, practitioners often overlook a relevant tool for navigating § 101 in AI cases, which is USPTO subject matter eligibility example 39.

Example 39 illustrates a method for training a neural network for facial detection. It outlines a multistage training process involving the collection of digital facial images, applying transformations, and utilising distinct training sets to detect human faces despite variations in scale and rotation.

Notably, the ultimate result of this claimed process is the generation of a trained AI model. The USPTO’s guidance confirms that such a claim is statutory because it avoids reciting a judicial exception altogether. The office recognised that a person cannot realistically generate a trained AI model in their mind, meaning the computational steps transcend any abstract “mental process”.

This framework is highly adaptable to complex electromechanical and AI arts. Instead of claiming the abstract idea of organising data, the claim can focus on the concrete, nonhuman computational operations. For example, when prosecuting a patent for a predictive AI model, the claims could recite the specific preprocessing of the input datasets and the complex data transformations required to train a specific attention layer within the network architecture. Focusing the claims on the specific technical architecture and training steps can prove the method is too complex for a mere “mental process”, making it an effective strategy to overcome a § 101 rejection.

Surviving § 101 can provoke a rising § 112(a) threat

The very techniques patent prosecutors use to survive the § 101 gauntlet are now acting as a lightning rod for § 112(a) enablement scrutiny.

By aggressively reciting functional results, data transformations, and complex models to satisfy § 101, practitioners expose their applications to the strict enablement standards established by the Supreme Court in Amgen v Sanofi and the Federal Circuit in In re Starrett. As the USPTO pivots away from § 101 per Desjardins, examiners are increasingly turning to § 112(a) to reject AI claims that lack sufficient technical depth.

Amgen effectively ended the era of broad functional claiming, prohibiting applicants from claiming a broad functional result without providing a commensurate disclosure of how to achieve that result across the full scope of the claim. If an application requires engineers to engage in a “research assignment” to practice the full scope of the invention, it likely is invalid.

Starrett compounded this vulnerability for software and AI. The Federal Circuit invalidated claims where the specification merely assigned a name to a function but failed to describe how the machine learning elements actually operated. Claiming “generation of a trained video upscaling model” might satisfy example 39 and pass § 101. However, if the specification merely provides a “black box” label and a single working example without disclosing the underlying neural network architecture, the training data structures, or the specific logic for handling edge cases, then the claim could face a fatal § 112(a) enablement challenge.

Furthermore, relying on the knowledge of a person having ordinary skill in the art (POSITA) to fill in the gaps is a dangerous gamble. Under the Wands factors, the predictability of the art is crucial. Because AI is a highly unpredictable field, courts and examiners are likely to conclude that an engineer would have to engage in unreasonable experimentation to build the system if the specification lacks concrete, detailed guidance.

Practical drafting guidelines for AI applications

Practitioners cannot rely on clever claim amendments alone to survive both § 101 and § 112(a). It ultimately comes down to pressing the inventors to crack open that specification. A robust, exhaustive specification is the most reliable defence against § 112(a) attacks.

To successfully navigate this dual threat at the USPTO and in the courts, patent prosecutors should consider these practical drafting guidelines:

1) Detail the architecture and data: Never assume the AI model is a standard component. The specification should explicitly describe the training data (how it is collected, structured, and labelled) and the specific neural network architecture being utilised. If the invention adjusts any part of a transformer configuration or modifies network layers to solve a specific problem, those modifications should be documented in exhaustive detail.

2) Expose LLM prompting techniques: Many modern systems rely on large language models (LLMs) or generative agents. If an invention uses an LLM, the specific prompting techniques are often central to how the invention actually works. Include example prompts or prompt templates in the specification to prove exactly how the system guides the LLM to produce the claimed result.

3) Map intermediate steps and transformations: Do not skip over the messy middle steps. Explain the intermediate input/output transformations to prove exactly how an algorithm handles edge cases, ambiguities, and variable data. Explain the logic as a step-by-step recipe.

4) Provide diverse working examples: Do not rely on a single use case. To satisfy Amgen and support broad claims, the specification should provide multiple, diverse working examples across the entire claimed scope.

Takeaways
•  AI patents require robust disclosure
•  Example 39 can help overcome rejection
•  Inventors need to disclose model details

Bradford Fritz is a partner at Birch, Stewart, Kolasch & Birch and a former USPTO patent examiner. He can be contacted at bfritz@bskb.com