Recent Microsoft PTAB ruling signals the key role of patent specifications in AI patent eligibility

June 8, 2026

Alert

In a continued development of patent eligibility policy following Ex parte Desjardins, the Patent Trial and Appeal Board (PTAB) reversed a subject-matter eligibility rejection under 35 U.S.C. § 101 for a Microsoft invention directed to AI-assisted code-editing tools, such as GitHub Copilot. Here, in Ex parte Ziegler, Appeal No. 2025-003643, decided May 29, 2026, the Board reversed the eligibility rejection in a 2-1 split decision, highlighting distinct perspectives on what constitutes an eligible claim to inventions in AI. In this decision, the specification made the difference.

Rejection based on typical "performable in the human mind” examiner arguments

Microsoft’s claims were directed to a system for improving code predictions in a programming editor in application No. 17/740,164 filed May 9, 2022. Rather than limiting input to code before the cursor as conventional code language models might, the claimed system assembles "additional context" from text after the cursor, open editor tabs, related code files, and metadata. Further, it specifies constructing a prompt by identifying a tree structure from that context before feeding it to the code language model. The Examiner rejected the claimed system under §101 as covering steps performable in the human mind, and the Board reversed with Administrative Patent Judge Jeffrey S. Smith writing for the majority and Administrative Patent Judge Sharon Fenick dissenting.

Majority found specification support for claimed improvement to code editor technology

Judge Smith found that the specification sufficiently described limitations of prior art code editors and explained how the claimed invention overcomes those limitations by leveraging additional technical context to improve predictions displayed to the programmer. The majority opinion concluded that the specification "provides sufficient details such that one of ordinary skill in the art would recognize the claimed invention as providing an improvement" to code editor technology, and then verified that the claim itself "reflects the disclosed improvement" through specific technical steps including tree structure identification, sibling relationships, and object rearrangement, rather than merely invoking a generic machine learning model in a code editing environment.

Dissent found claims merely applied generic machine learning to new input data

Judge Fenick applied Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), characterizing the claims as merely applying "machine learning to code editing using assertedly new input." Judge Fenick also cited non-precedential Rensselaer Polytechnic Inst. v. Amazon.com, Inc. (Fed. Cir. 2026) for the proposition that "[g]eneric use of AI without other parameters, such as improving the mathematical algorithm or making machine learning better, is abstract." The dissent read Microsoft's arguments that using "additional context" as input constitutes an improvement to code language modeling as "an admission that no improvement to the machine learning model is claimed," and merely "the use of a different source of input data to a generic code language model." Accordingly, Judge Fenick would have sustained the rejection, finding that the asserted improvement does not integrate the abstract idea into a practical application.

The doctrinal divide: Improving model operation vs. generic application to new data

The majority opinion establishes that improving the technology the machine learning model serves, in this case, code editors, is sufficient to integrate the abstract idea into a practical application, while the dissent takes the position that unless the claims improve the machine learning model itself, they fail under Recentive. This split reflects a broader framework emerging across the PTAB following the precedential Desjardins decision in Appeal 2024-000567 decided September 26, 2025, where claims reflecting "an improvement to how the machine learning model itself operates" may be found eligible, while claims that "do no more than claim the application of generic machine learning to new data environments" are not.

The present decision in Ziegler places Microsoft's claimed system for generative code editing at the exact boundary between these two positions. When the asserted improvement to how a claimed machine learning model operates concerns the particular application of machine learning inputs in a relatively new data environment, the majority finds eligibility provided the specification sufficiently supports the improvement, while the dissent remains critical of claims that do not improve the mathematical algorithm or make the machine learning model itself better.

Case highlights for capturing patent-eligible inventions under 35 U.S.C. § 101

1: Prior art limitations in the specification

The majority opinion faulted the Examiner for failing to address why "the specification's disclosure of using additional context to overcome limitations of prior art code editors does not describe an improvement in the technological field of code editors." The specification identifies prior art limitations to precursor text, explains how those limitations reduced prediction usefulness, and describes the claimed structure as the solution. This may have provided critical foundation for the majority finding a technological improvement.

2: How the model operates on context

The dissent argued that the claims merely change "what is used as input to the code language model." The majority opinion determined that the claimed system overcame this characterization because the claim further specifies how the system structures context, including tree identification, sibling analysis, and object rearrangement, rather than reciting only what data enters the model.

3: Technological improvement vs. business outcome

The majority opinion framed the improvement as one to "code editing technology" rather than a generic invocation of hardware applied to an abstract process. This distinction reflects the Desjardins framework, where the claimed improvement must be technological in nature, and a claimed improvement describable only by reference to a business benefit rather than a technical advancement may fall on the Recentive side of the divide.

Conclusion

The Recentive and Desjardins divide will define patent examination at the U.S. Patent and Trademark Office and prosecution strategy for the foreseeable future. In view of Ziegler and other proceedings since Desjardins, there are clear strategic advantages in (1) drafting specifications that articulate the technical shortcomings of existing systems and describe the claimed invention as the solution to those shortcomings, (2) reciting in the claims how the system structures and processes context rather than merely reciting new input data to a generic model, and (3) framing the claimed improvement as a technological advancement in the relevant field rather than a business outcome.

This article examines how AI tools — both public and enterprise — may be silently circumventing your document retention policies, the risks of using public AI tools, including inadvertent public disclosure and unintentional IP infringement and best practices for updating policies, training inventors, and meeting your duty of candor to the USPTO.

Colorado’s new law more clearly defines parameters for AI developers and deployers and significantly scales back the obligations the 2024 CAIA was slated to impose this summer. The new law will help ease the tension between the legislature and AI companies, while also likely rendering the pending DOJ and xAI lawsuit moot.

As AI use accelerates, governance is lagging, leaving companies exposed to evolving regulations and the application of legacy laws to new technologies. Organizations need a practical AI governance framework that aligns use cases with risk, implements clear controls, and evolves alongside regulatory change.