Patents

Recentive Analytics v. Fox: Generic Machine Learning Meets the Abstract-Idea Bar

The Federal Circuit's first precedential machine-learning eligibility ruling holds that applying off-the-shelf models to a new data environment claims an abstract idea under § 101.

March 30, 2026
Server room corridor with rows of illuminated data servers
Recentive's patents promised optimized broadcast schedules, but the Federal Circuit found only generic machine learning applied to a new field. Shutterstock
Educational content, not legal advice. This article explains general legal concepts. It does not create an attorney–client relationship. For your specific situation, consult a licensed attorney.

In Recentive Analytics, Inc. v. Fox Corp., No. 23-2437, 134 F.4th 1205 (Fed. Cir. Apr. 18, 2025), the United States Court of Appeals for the Federal Circuit affirmed a Rule 12(b)(6) dismissal from the District of Delaware and, in doing so, issued what it expressly characterized as its first precedential decision squarely confronting the patent eligibility of machine-learning inventions under 35 U.S.C. § 101. Writing for a unanimous panel that also included Judge Prost and Chief District Judge Mitchell S. Goldberg of the Eastern District of Pennsylvania, sitting by designation, Judge Dyk held that all four asserted patents claimed nothing more than the application of conventional, generically described machine learning to a new field of use — the scheduling of live events and the construction of broadcast “network maps.” The decision is short, but its reasoning supplies a durable analytical template for the wave of AI patent disputes now reaching the appellate court.

At a glance

Recentive owns four patents — U.S. Patent Nos. 11,386,367 and 11,537,960 (the “machine-learning training” patents) and 10,911,811 and 10,958,957 (the “network-map” patents). The first family describes dynamically generating optimized event schedules using machine-learning models trained on historical data; the second describes automatically generating network maps that determine which programs a broadcaster’s channels display, by geographic market and time slot. Recentive conceded a critical point: its claims do not invent or improve any machine-learning algorithm. They instead apply existing, generic models — the patents themselves disclaim reliance on any particular technique — to a problem domain where, the patentee said, machine learning had not previously been deployed. Fox moved to dismiss on eligibility grounds. The district court granted the motion, and the Federal Circuit affirmed: at Alice step one the claims are directed to the abstract idea of producing schedules and maps with known mathematical methods, and at step two they supply no inventive concept beyond the abstract idea itself.

The two-step inquiry and the “new environment” problem

The court applied the now-familiar framework of Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014), as refined for laws of nature and mental processes in Mayo Collaborative Services v. Prometheus Laboratories, 566 U.S. 66 (2012). Step one asks whether a claim is “directed to” a patent-ineligible concept — here, an abstract idea. Step two, the search for an “inventive concept,” asks whether the claim elements, individually and as an ordered combination, transform the abstract idea into a patent-eligible application by adding “significantly more.”

At step one, the panel located the claims firmly within the abstract-idea category. The asserted methods automate tasks that humans had long performed by hand — assembling event schedules and allocating programming across markets — and the patents’ contribution was to perform those tasks faster and at greater scale by feeding data to a machine-learning model. That, the court reiterated, is not enough: doing a conventional task more efficiently with a computer, or with a generic model, does not convert an abstract idea into patentable subject matter. The opinion leaned on a principle the Federal Circuit has applied across two decades of software cases — that “an abstract idea does not become nonabstract by limiting the invention to a particular field of use or technological environment.” Recentive’s central pitch was precisely that kind of limitation: it argued that nobody had previously applied machine learning to event scheduling or network-map generation. The court treated novelty of the application domain as irrelevant to the eligibility question. The claims recited generic models, generic training, and generic inputs and outputs, distinguished only by the data environment to which they were pointed.

Why iterative training is not an “improvement”

Recentive’s most substantial argument was that its claims do more than apply a static model: they recite machine-learning models that are iteratively trained and dynamically updated as new data and parameters arrive, producing “custom” algorithms tailored to the scheduling problem. If the claims improved machine learning itself, they might clear step one under the line of cases — such as McRO, Inc. v. Bandai Namco Games America and Enfish, LLC v. Microsoft — that protect specific technological improvements to how computers operate.

The panel rejected the characterization on its own terms. Iterative training on updated data, it explained, is inherent to how machine learning works; describing that ordinary behavior at a high level of generality does not disclose an improvement to the technology. The claims, the court found, do not recite a specific method for improving the mathematical model or “making machine learning better.” They offer, at most, “a mere concept” — use a trained model to generate a schedule — “without disclosing how to implement that concept.” That functional, result-oriented framing is what doomed the patents. A claim that recites the desired outcome (an optimized schedule generated by a model that gets better with more data) while leaving the actual algorithmic mechanism to the generic state of the art claims the goal, not the inventive means of reaching it.

Step two followed almost automatically. Because the machine-learning components were, in the court’s words, “broad, functionally described, well-known techniques,” they could not themselves supply the inventive concept missing at step one. The conventional generic-computer-implementation elements — servers, databases, a graphical interface for entering parameters — added nothing significant. Reciting the abstract idea while invoking generic technology to carry it out is the paradigm the Supreme Court foreclosed in Alice.

Open questions

The opinion is careful to disclaim the broadest reading. The court expressly stated that it was not holding machine-learning inventions categorically ineligible, and it acknowledged that claims directed to genuine improvements in how models are built, trained, or deployed may well survive § 101. That leaves the boundary unmapped: how specific must a claimed training methodology, architecture, or feature-engineering technique be before it crosses from “applying generic machine learning” into “improving” it? The opinion gives the negative — high-level recitations of ordinary training are not enough — without supplying a clean positive test.

A second open question is procedural and now pending at the highest level. Recentive filed a petition for certiorari (No. 25-505) in October 2025, pressing the recurring complaint that the Alice/Mayo framework is unworkable and that the Federal Circuit’s “directed to” inquiry collapses eligibility into questions of novelty and enablement better addressed under §§ 102, 103, and 112. Whether the Supreme Court takes up § 101 again — it has repeatedly declined invitations since Alice — will shape how much weight Recentive ultimately carries. A third question concerns the Rule 12(b)(6) posture: the court resolved eligibility on the pleadings without claim construction or factual development under Berkheimer v. HP, reinforcing that, where a patentee concedes the genericness of its computational components, eligibility remains a threshold legal question.

Implications

  • Drafting must target the model, not the field. Claims that recite applying machine learning to a previously untouched domain are exposed; claims should recite a concrete, non-generic improvement to architecture, training, feature selection, or inference, described with enough specificity to show how the improvement is achieved.
  • “Iterative training” and “dynamic updating” are not magic words. Reciting the ordinary mechanics of machine learning at a functional level invites a step-one finding that the claim is directed to an abstract idea.
  • Concessions are dangerous. Recentive’s acknowledgment that it used generic, off-the-shelf models effectively decided the case; patentees should preserve, where truthful, the specificity of their technical contribution.
  • Expect early dismissals. The decision validates resolving AI eligibility on the pleadings when the computational components are admittedly conventional, raising the value of a well-pleaded § 101 motion for accused infringers.
  • Trade secrecy gains appeal. For genuinely novel training pipelines and model improvements that are hard to claim with the required specificity, trade-secret protection becomes a more attractive complement or alternative to patenting.

Frequently asked questions

Did the Federal Circuit hold that machine-learning patents are categorically ineligible? No. The panel expressly declined to adopt a categorical rule and acknowledged that claims directed to specific improvements in machine-learning technology itself may be eligible. The holding is confined to claims that apply generic, functionally described models to a new data environment without improving the underlying technology.

Why didn’t the novelty of applying machine learning to scheduling save the claims? Because eligibility under § 101 does not turn on novelty of the application domain. The court reaffirmed that confining an abstract idea to a particular field of use or technological environment does not make it non-abstract; novelty and non-obviousness are addressed under §§ 102 and 103, not § 101.

What should patent drafters take from the decision? Specificity about the technological improvement is paramount. Claims should describe a concrete mechanism — how the model is structured, trained, or deployed in an unconventional way — rather than reciting the desired result of using a generically trained model, and the specification should disclose how that improvement is implemented.

Authorities and sources