Less than a year after holding that generic machine-learning patents are abstract in Recentive Analytics, Inc. v. Fox Corp., the Federal Circuit may be refining where to draw the line on patent eligibility. Monday’s oral argument in Rensselaer Polytechnic Institute v. Amazon.com, Inc. suggests the court is closely scrutinizing whether the AI patent claims at issue there capture genuine technical improvements—or instead merely describe known techniques applied to new fields.
The Recentive Backdrop
The Federal Circuit’s April 2025 decision in Recentive established that patents merely applying generic machine-learning techniques to new data environments are patent ineligible under § 101. The key holding was 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.” That framework dominated oral arguments in RPI.
The Technology
RPI’s U.S. Patent No. 7,177,798 covers a method for processing a natural language input using an interconnected metadata database to process user queries. The specification refers to case-based learning, which RPI described as “a problem-solving methodology conceived in the 1970s that uses past experiences or cases to solve new problems.” Amazon’s Alexa voice assistant is the accused product.
The District Court
The lower court had granted Amazon’s summary judgment motion, finding the ’798 patent invalid under § 101. Applying the two-step Alice framework, the lower court found, at step one, that “the ’798 patent claims recite an approach to collecting, analyzing, and displaying information and [were] therefore directed to an abstract idea,” even if limited to Natural Language Processing (NLP). It then found, at step two, that the claim elements failed to transform the underlying abstract idea into a patent-eligible application. RPI appealed.
The Parties’ Arguments
RPI argued that the district court improperly disregarded its own claim constructions when finding the claims abstract. For example, RPI contended that the court’s construction of the term “case information” suggested that the claimed natural language processing method uses “case-based reasoning.” But then in its invalidity analysis, the court, according to RPI, determined that case-based reasoning is not reflected in the claims. This is important because to get past Alice Step One, RPI argues that its “patent claims are directed to using metadata-specific case-based reasoning to solve a critical technological problem within NLP.”
Further, unlike the generic machine learning in Recentive, RPI contended, the ’798 patent claims a specific technological improvement: an interconnected metadata database architecture implementing case-based reasoning in a way never before used in NLP systems. RPI also emphasized that its claims require constrained searches based solely on natural language input across four enumerated metadata categories—structural requirements that go beyond applying known AI to a new field.
Amazon countered that RPI’s purported improvements exist only in the specification, not the claims. Citing Recentive, Erie, and SAP America, Amazon also argued that patent eligibility is not conferred by applying machine-learning techniques generically to new fields, having new or interconnected data in a database, or identifying new combinations of information in the database. Since RPI admitted that case-based reasoning has been known since the 1970s, Amazon characterized the patent as merely applying a “conventional learning technique” to NLP—precisely what Recentive prohibits.
What the Court Focused On
Within the opening minute, the court asked: “How is this different from Recentive?” The court then pressed whether applying case-based reasoning to NLP is meaningfully distinguishable from the generic machine learning found abstract in that case. The court pressed that it can’t be non-abstract to use AI for natural language processing. The judges also scrutinized whether any technological improvement is captured in the claims rather than merely disclosed in the specification, asking pointedly: “What is in the claims other than just functional language?” The panel further explored whether Recentive’s outcome turned on the patentee’s concession of no technological improvement—a concession RPI has not made here. Amazon responded that courts routinely find claims abstract regardless of such concessions.
Practical Takeaways
- Claims should capture the improvement. The court’s skepticism reinforces that disclosing a technical advance in the specification may not be enough—the improvement should be reflected in the claim language itself. Functional recitations of results (“provide, perform, provide, identify and determine”) without the technical “how” may remain vulnerable.
- Recentive may apply beyond machine learning. Although Recentive addressed machine-learning specifically, the court’s questions suggest its logic extends to other AI techniques. Applying any known AI method to a new field, without claiming a specific technical implementation, may be insufficient.
- Claim construction may not save abstract claims. RPI’s reliance on favorable Markman constructions did not appear to move the panel. Narrow constructions of individual terms may not be enough to transform otherwise functional claims into patent-eligible subject matter if the claims still read on abstract concepts at a high level.
We will continue to monitor this case and report on the Federal Circuit’s decision when issued.
