Software and AI Patents in 2026: Why the USPTO and the Federal Circuit No Longer Agree

Two Audiences, Two Standards

If you are patenting software or machine learning, you now have to persuade two audiences that apply different versions of the same test. The USPTO has spent the past 15 months making eligibility easier to establish during examination. Over the same period, the Federal Circuit has kept invalidating software claims that describe a result without claiming how it is achieved.

The practical consequence is simple: an allowance is not validity insurance. USPTO guidance has no force of law (89 FR 58128), and in a nonprecedential 2026 opinion the Federal Circuit repeated that courts do not have to defer to it (Innovaport v. Target). A claim has to clear both bars.

The Test Has Not Changed

Courts still apply the two-step Alice/Mayo framework: is the claim directed to an exception such as an abstract idea, and if so, does it add an inventive concept? Adding "apply it with a computer" to an abstract idea fails (Alice v. CLS Bank).

The USPTO runs the same inquiry as Step 2A Prong One (does the claim recite an exception), Step 2A Prong Two (is it integrated into a practical application) and Step 2B (does it add significantly more) (MPEP 2106). Whether AI helped create the invention is not part of that test (89 FR 58128).

Recentive Set the Floor for Applied Machine Learning

In April 2025 the Federal Circuit addressed applied machine learning directly. The holding is narrow but sharp:

"we hold only that patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101."

The court also said that iterative training and real-time updates are incident to the very nature of machine learning, and that limiting the idea to a new field does not help (Recentive Analytics v. Fox). The Supreme Court denied certiorari on December 8, 2025 (Docket 25-505).

The claim most exposed by Recentive is the common commercial one: a known model, trained on domain data, applied to a new problem.

Desjardins Helped, but Less Than the Headlines Suggest

Ex parte Desjardins is the USPTO's main counterweight. An Appeals Review Panel led by Director Squires considered a DeepMind continual-learning application and credited a limitation that adjusted model parameters while protecting performance on an earlier task. The panel called that an improvement to how the machine learning model itself operates (Ex parte Desjardins).

Three details often get lost:

The panel vacated a rejection the Board had raised on its own. The claims still stand rejected under § 103, so this was not an allowance.
The specification alone was not enough. The panel required that the claim itself reflect the disclosed improvement.
It binds the USPTO, not the courts. It was designated precedential on November 4, 2025 (USPTO Precedential Decisions), and none of the Federal Circuit opinions we reviewed cites it.

A December 2025 notice folded the decision into examination practice (USPTO Advance Notice on MPEP). The online MPEP 2106 still does not mention it, so read the notice alongside the manual.

What the USPTO Added

The rest of the shift came through memos, and the USPTO describes them as procedure rather than new tests:

August 2025: examiners were told not to stretch the mental process grouping, and to reject only when ineligibility is more likely than not (USPTO Memo, August 4, 2025).
December 2025: the Subject Matter Eligibility Declaration (SMED) arrived. It is optional, it needs a factual nexus to the claim, and it cannot supply what the original disclosure lacked. Legal opinion on eligibility gets no weight (USPTO SMED Memo, December 2025).
September 2026: a dedicated Patent Center filing code and five best practices followed (USPTO SMED Memo, September 2026), along with a flowchart resolving an even split in the applicant's favor (USPTO SMED Flowchart).

We found no published data on how often SMEDs succeed.

Meanwhile, in the Federal Circuit

The three precedential 2026 software decisions we reviewed all went against the patentee:

US Patent No. 7,679,637 LLC v. Google (January): result-focused functional language failed, and dismissal on the pleadings was affirmed (637 LLC v. Google).
GoTV v. Netflix (February): the court reversed a judgment entered after a jury verdict, and a concrete improvement required more than result-focused functional language (GoTV v. Netflix).
Columbia v. Gen Digital (March): features not reflected in the claim language did not count at step one, and the case went back for step two (Columbia v. Gen Digital).

We did not find a precedential Federal Circuit decision holding an AI or ML claim eligible. The last clear precedential software win we found is Contour v. GoPro (2024), where the specification disclosed, and the claims reflected, a specific way of generating high- and low-quality video streams in parallel (Contour v. GoPro). This was not a full census of every decision, but the pattern is clear.

Where the Two Standards Split

Issue	USPTO	Federal Circuit
Must the claim state the improvement?	The claim "does not need to explicitly recite the improvement" (USPTO Memo, August 4, 2025)	Features "not reflected in the relevant claim language" do not count (Columbia v. Gen Digital)
Are speed and efficiency enough?	Better, cheaper and faster can support eligibility (USPTO Memo, April 30, 2026)	Adding computer functionality for speed or efficiency does not confer eligibility (GoTV v. Netflix)

How We Would Draft for Both Audiences

This is our reading of the sources rather than a rule from any of them. Eligibility decisions turn on the facts of each claim, so treat it as a starting point for a conversation with counsel.

Name the mechanism in the specification. State the limitation of the prior approach, the technical change and the measured effect.
Put evidence in the original filing. A SMED cannot fill a disclosure gap, and the USPTO encourages applicants to include favorable evidence in the specification (USPTO SMED Memo, December 2025).
Claim how, not what. Recite the structure, update rule, architecture or data flow that produces the result.
Do not rest on domain data or retraining. Recentive treats both as ordinary machine learning.
Put every argued feature in an independent claim. After Columbia, an improvement that lives only in the specification may not count.
Tie any speed gain to a specific technical change. That wording works for both audiences.
Weigh trade secrecy for applied ML. If the value is in a pipeline competitors cannot inspect, a patent may not be the best fit.

Do Not Wait for Congress

The Patent Eligibility Restoration Act would replace the judicial exceptions with enumerated exclusions (S. 1546). According to Holland & Knight, it received its first full-committee hearing on July 14, 2026. We found no markup or vote as of this writing, so plan around current law.

The Bottom Line

Draft for the stricter reader. A claim that names its technical mechanism, in a specification that explains and measures it, is positioned for both the examiner and a later court challenge.

Obsidyan's drafting tools help you structure a specification and claim set, and our claims guide covers the fundamentals. Neither replaces review by a registered practitioner, especially on a question this fact-dependent.

Sources

Sources checked November 6, 2026. Fees, pilot programs and pending rules change, so confirm current figures with the USPTO. This article is general information, not legal advice, and Obsidyan is not a law firm. Talk to a registered patent attorney or agent about your situation.