# What are the best practices for AI patent prosecution in 2026?

aitrademarkreview.com · August 28, 2026

> What "AI Patent Prosecution" Actually Means in 2026 The phrase "AI patent prosecution" now covers two distinct workstreams that practitioners routinely...

## What "AI Patent Prosecution" Actually Means in 2026

The phrase "AI patent prosecution" now covers two distinct workstreams that practitioners routinely conflate. The first is the prosecution of patents that claim AI inventions — machine-learning models, neural network architectures, training pipelines, and AI-driven systems. The second is the use of AI tools by patent attorneys and examiners during the prosecution process itself, including AI-assisted prior-art search, claim drafting, office-action response generation, and disclosure tracking. Both workstreams now carry concrete risk profiles that did not exist five years ago, and the USPTO's updated 2024–2025 guidance on Subject Matter Eligibility Declarations under Rule 132 has made the dividing line between them more visible.

**Also worth reading:** [What are the best practices for authenticating AI evidence in intellectual property and trademark disputes?](https://aitrademarkreview.com/knowledge/what_are_the_best_practices_for_authenticating_ai_evidence_in_intellectual_property_and_trademark_disputes.php) · [What are the trademark clearance workflow best practices for modern businesses?](https://aitrademarkreview.com/knowledge/what_are_the_trademark_clearance_workflow_best_practices_for_modern_businesses.php) · [What are the definitive AI trademark enforcement best practices for protecting brand identity in generative models by 2026?](https://aitrademarkreview.com/knowledge/what_are_the_definitive_ai_trademark_enforcement_best_practices_for_protecting_brand_identity_in_generative_models_by_2026.php)

For applicants, the practical question is no longer whether to file AI-related patent applications, but how to draft and prosecute them so they survive Section 101, Section 102, Section 103, and Section 112 challenges while also managing the new duty of disclosure issues that arise when generative-AI tools are used during preparation. The Bloomberg Law analysis of patent lawyers as "shields" against AI misuse and the National Law Review's warnings about generative-AI disclosure risk both point to the same operational reality: the tools you use to draft the application can become the evidence used to challenge it.

A third layer sits beneath the other two: institutional AI adoption at the USPTO itself. The Office has been rolling out internal AI tools for classification, search, and examiner support, and IPWatchdog's reporting on the USPTO's AI agenda indicates practitioners should expect examiners to be working with AI-assisted prior-art results by the time a 2026 application reaches action. That changes how an applicant should structure the specification, the claims, and the information disclosure statement.

## The Current Legal Framework for AI-Claiming Patents

Section 101 remains the single largest source of rejection for AI-related applications, and the 2025 Holland & Knight review of Section 101 stories confirms that the USPTO's application of the Alice/Mayo two-step framework continues to invalidate abstract idea rejections on appeal. The Office's updated best-practices memorandum on Rule 132 Subject Matter Eligibility Declarations — issued through Crowell & Moring's coverage — gives examiners a more structured template for evaluating whether a claimed invention recites an "inventive concept" sufficient to transform an abstract idea into patent-eligible subject matter.

Three doctrinal points matter for prosecution strategy. First, the Federal Circuit's 2024–2025 treatment of AI claims has not produced a clean circuit split, but district-court invalidation rates for software-and-AI patents remained above 70 percent in NPE litigation according to the Reuters analysis of AI in patent litigation. Second, the USPTO's 2024 guidance on inventorship for AI-assisted inventions (following Thaler v. Vidal and the Supreme Court's denial of cert) confirms that a natural person must be named as inventor, and that an AI system cannot be listed even if it contributed substantively to the conception of the claimed subject matter. Third, the Thaler line of cases has been treated as persuasive by examiners when rejecting claims directed to outputs of autonomous AI systems without a specifically identified human contributor.

For Section 112, written-description and enablement rejections on AI claims are increasingly common when the specification fails to disclose training data characteristics, model architecture details, or reproducible performance metrics. The Mayer Brown commentary on the Thaler cert denial underscores that practitioners should treat the human-inventor requirement as a documentation discipline, not a formality.

## Duty of Disclosure When Generative AI Touches the File

The National Law Review's warning on generative-AI disclosure risk identifies the central problem: a 37 C.F.R. § 1.56 duty to disclose material prior art applies regardless of how the prior art was discovered. If an attorney or agent uses a generative-AI tool that surfaces a reference not previously known, and that reference is material to patentability, the duty of candor still attaches. Failing to disclose because "the AI found it, not the attorney" is not a recognized safe harbor.

The operational best practice is to treat AI tool outputs as if they were the output of a junior associate: reviewable, attributable, and disclosable when material. That means documenting the prompts used, the date the tool was queried, the version of the tool, and the specific references the tool surfaced. It also means evaluating each surfaced reference against the claims as actually filed — not the claims as originally conceived — because claim scope can shift during prosecution in ways that change materiality.

The Bloomberg Law framing of patent lawyers as "shields" against AI misuse applies here in a literal sense. The attorney is the only party with both the authority and the expertise to determine what becomes of an AI-generated reference, and the duty runs from the attorney to the USPTO, not from the AI tool to the client. Confidentiality is a separate issue: practitioner use of public commercial AI tools can transmit client confidential information to third-party servers, which implicates both state-bar confidentiality rules and the duty of loyalty.

## Drafting Specifications That Survive §112 Challenges

AI patent specifications fail enablement and written-description review for predictable reasons. The specification describes the AI component at a level of abstraction that matches how a research scientist would describe it to a colleague, not how a patent examiner would evaluate it against §112. Practitioners drafting for the 2026 examination environment should include at least the following concrete disclosures, structured to be both technically accurate and examiner-friendly.

The specification should disclose the specific architecture (e.g., transformer with named attention mechanism, convolutional backbone with specified depth, or a graph neural network with named aggregation function). It should disclose the training data at a level that allows a person of ordinary skill in the art to obtain comparable results without undue experimentation: data type, dataset size, source distribution, labeling methodology, and any preprocessing steps. It should disclose at least one quantitative performance metric tied to a specific benchmark or held-out test set, with the metric expressed in the same units used in the relevant technical field.

A comparison table is useful both in the application and in prosecution strategy. The table should compare the claimed invention against the closest prior-art systems on at least three dimensions: performance metric, computational cost, and training-data requirement. That structure preempts the examiner's most common written-description rejection, which is that the specification fails to demonstrate that the inventor had possession of an invention that achieves a result better than what the prior art already achieved.

| Specification Element | Minimal Acceptable Disclosure | Best-Practice Disclosure |
| --- | --- | --- |
| Model architecture | Named model class (e.g., "neural network") | Specific layer types, dimensions, activation functions, connectivity |
| Training data | "A labeled dataset" | Size, source, label schema, preprocessing, augmentation strategy |
| Performance metric | "High accuracy" | Specific metric, dataset, baseline comparison, statistical significance |
| Inventive contribution | "Improved over prior art" | Concrete technical effect with measured delta and root cause |

## Drafting Claims That Survive §101
Section 101 rejections on AI claims have settled into a recognizable pattern. The examiner applies Alice step one and concludes that the claim recites an abstract idea — usually a mathematical concept, a mental process, or a method of organizing information. The applicant responds at step two by arguing that the claim recites an "inventive concept" sufficient to transform the abstract idea. The 2024–2025 Federal Circuit case law has not loosened this framework, but it has clarified the evidentiary burden: bare assertions that "the integration provides a technical improvement" are no longer enough; the applicant must point to a specific, measurable improvement in a technical field.

Claim drafting best practice in 2026 is to anchor each independent claim to a specific technical application, with the AI component described in terms of what it does inside that application rather than in terms of its internal mathematical structure. A claim to "a method for training a neural network" is more vulnerable than a claim to "a method for controlling a chemical reactor, comprising: receiving sensor data; processing the sensor data through a trained neural network to predict a reactor condition; and adjusting a control parameter based on the prediction." The second claim recites an inventive concept tied to a concrete technical field, which is the structure the USPTO's 2024 examiner guidance rewards.

Dependent claims should build outward from the independent claim with technical limitations, not abstract ones. "The method of claim 1, wherein the neural network is a transformer with eight attention heads" is a useful dependent claim; "The method of claim 1, wherein the step is performed on a cloud server" is not, because it adds a non-technical limitation that does not change the abstract-idea analysis.

## Examiner Interviews, RCEs, and the 2026 Procedural Toolkit

The 2025 USPTO data and reporting on the AI agenda both indicate that examiner interviews have become a higher-leverage tool for AI applications than for conventional mechanical or chemical applications. The reason is structural: AI applications raise §101 and §112 issues that are often resolved by agreement about what the specification discloses, and those agreements are easier to reach in a 30-minute interview than in three rounds of office-action correspondence.

The Request for Continued Examination (RCE) is a less attractive tool than it was before 2023. RCE fees have increased, and the USPTO's internal AI tools mean that an examiner reading a new amendment often has better prior-art retrieval than the original examiner had at first action. A continuation application with a narrowed claim set is often a more efficient path than an RCE on the original application, especially when the original application's specification supports the narrowed claims without needing new matter.

Appeals to the Patent Trial and Appeal Board (PTAB) have become a more attractive option for AI cases, particularly after the 2024–2025 PTAB decisions clarifying how Alice step one is applied at the board level. The board has shown a willingness to reverse §101 rejections when the examiner's analysis fails to identify a specific abstract idea, and a successful appeal can produce a patent with broader claim scope than a negotiated allowance.

## Common Mistakes That Cost Applicants Their Patents

Five recurring mistakes account for a disproportionate share of AI patent prosecution failures. The first is treating AI as a black box in the specification. The Federal Circuit's 2024 enablement decisions in the life sciences, while not directly binding on AI cases, have been cited by examiners as support for requiring more disclosure of internal model mechanics. The second is naming an AI system as a co-inventor, which the Thaler line of cases confirms will result in an inventorship rejection. The third is failing to disclose material prior art surfaced by AI tools, which creates both inequitable-conduct risk and a separate ethical exposure for the prosecuting attorney. The fourth is drafting claims at the wrong level of abstraction — either too abstract (vulnerable to §101) or too narrow (vulnerable to design-around). The fifth is relying on continuation chains to fix drafting errors that should have been corrected in the original application; continuation practice has tightened, and the USPTO's 2024–2025 internal guidance on double-patenting has increased the difficulty of maintaining parallel application families.

The most expensive mistake is the failure to document the human contribution to AI-assisted inventions. Thaler establishes that conception must be performed by a natural person, and the documentation of that contribution is the only evidence that can rebut a later inventorship challenge. That documentation should exist in the inventor's own words, dated before the patent application's filing date, and should describe what the human inventor conceived that the AI tool merely implemented or optimized.

## When to Act and What to Budget

AI patent prosecution is a 12- to 36-month process, and the 2026 fee structure makes early decisions more expensive than they used to be. The basic filing fee for a large-entity utility application is now $1,820, with the examination fee, search fee, and excess-claim fees pushing the realistic first-stage cost above $4,000 for a typical AI application with multiple independent claims and more than 20 total claims. An RCE adds another $2,000 to $2,800, and a PTAB appeal filing fee exceeds $1,000 before the briefs are prepared.

The right time to involve a patent attorney is before the first line of code is written or the first model is trained, not after the invention is complete. Conception evidence is harder to reconstruct than to preserve, and the specification is easier to draft when the underlying engineering documentation still exists. Practitioners who wait until the invention is "done" often end up drafting specifications that describe what the system does rather than how it works, which is the structural pattern most likely to produce a §112 rejection.

For applicants who cannot afford full-service patent prosecution, the right alternative is a focused prior-art search and freedom-to-operate analysis, which can be completed in two to four weeks at a cost of $3,000 to $10,000 depending on the technology area. That investment often prevents the much larger cost of filing a patent that is later invalidated. For applicants with a working AI system and limited budget, provisional patent applications remain useful as priority-establishing vehicles, but they do not buy examination, and the 12-month deadline for filing a non-provisional continuation is non-extendable.

## The Bottom Line

AI patent prosecution in 2026 is a documentation discipline as much as it is a legal discipline. The cases that succeed are the cases where the specification discloses enough technical detail to satisfy §112, the claims recite a concrete technical application sufficient to survive §101, the duty of disclosure has been honored with respect to every AI-sourced reference, and the human contribution to conception is documented before the application is filed. The cases that fail are the cases where the applicant treated the patent as a marketing document rather than a technical disclosure, or where the prosecuting attorney used AI tools without maintaining the same documentation discipline that the rules require of human researchers. Neither technology is going to replace the other, and the practitioners who do the best work in 2026 are the ones who treat both with appropriate skepticism.

## Quick answers

### What is the most common reason AI patent applications are rejected?

Section 101 rejection under the Alice/Mayo two-step framework remains the most common rejection, with examiners frequently concluding that AI claims recite an abstract idea without an inventive concept tied to a concrete technical application. District-court invalidation rates for software-and-AI patents exceeded 70 percent in NPE litigation according to recent Reuters analysis.

### Can an AI system be named as an inventor on a US patent?

No. Following Thaler v. Vidal and the Supreme Court's denial of cert, the USPTO requires a natural person to be named as inventor, and the Federal Circuit has affirmed that an AI system cannot be listed as inventor even if it contributed substantively to conception. Practitioners should document the human contribution in dated records before the filing date.

### Do I have to disclose prior art that an AI tool surfaced during patent drafting?

Yes. Under 37 C.F.R. § 1.56, the duty of disclosure applies to material prior art regardless of how it was discovered, including references surfaced by generative-AI tools. Practitioners should log prompts, dates, tool versions, and surfaced references, then evaluate each for materiality against the claims as filed.

### What specification disclosures help AI patents survive Section 112 challenges?

Disclose specific model architecture (layer types, dimensions, activation functions), training data characteristics (size, source, label schema, preprocessing), and at least one quantitative performance metric tied to a benchmark with a baseline comparison. Specifications that describe AI components as black boxes are increasingly vulnerable to enablement and written-description rejections.

### Should I file a continuation or an RCE after an AI patent rejection?

Continuations are often more efficient than RCEs for AI cases in 2026 because RCE fees have increased and the USPTO's internal AI tools give examiners better prior-art retrieval than the original examiner had. A continuation with narrowed claims supported by the original specification can preserve priority without paying the RCE surcharge.

Canonical: https://aitrademarkreview.com/knowledge/what_are_the_best_practices_for_ai_patent_prosecution_in_2026.php
Markdown: https://aitrademarkreview.com/knowledge/what_are_the_best_practices_for_ai_patent_prosecution_in_2026.php/index.md