The U.S. Court of Appeals for the Federal Circuit’s recent conclusion that applications of machine learning are not necessarily eligible for patenting has drawn significant notice from the AI community. In Recentive Analytics v. Fox Corp., the Federal Circuit affirmed that claims directed to using machine learning to predict TV viewership were patent-ineligible abstract ideas. While the decision grabbed attention, it was unremarkable for practitioners with deep experience in AI, given the way the patents presented their technology. The Federal Circuit’s decision instead reinforces what such experienced practitioners have long known: The challenge in obtaining strong AI patents isn’t the technology itself, but rather in crafting a specification that tells the right story about technological innovation. Best practices for protecting machine learning fundamentally follow best practices for other types of software. When protecting high-value applications of AI, it remains critical to work with counsel who understand both the technology and how to position it strategically.
Avoiding Surface-Level Innovation Claims
An example may help frame the question. Nearly a decade ago, I worked with a client seeking to help patients with a chronic condition predict whether they would experience their symptoms each day (details being omitted for obvious reasons). In this context, there was a decades-old technique for evaluating certain physiological measurements using charts that provided useful but imperfect forecasts. The physician had implemented this known evaluation technique using machine learning—the first time AI had been applied to this particular clinical question. This was years before the current wave of AI, but AI in general was not new, particularly in the medical context. The critical question we wrestled during patent strategy discussions was this: If the physiological measurements were known, the analysis was known, and the output was known, if the AI was characterized as routine engineering, then what exactly was “new” and “nonobvious” that should form the foundation of the patent protection?
Especially at that time period, other counsel might have stopped at “first application of AI to this problem.” Recentive demonstrates, though, that that approach fundamentally misunderstands how to draft strong patent protection for AI.
The Recentive Decision
The Federal Circuit’s decision in Recentive applied the Supreme Court’s two-step Alice framework that has become central to patent eligibility analysis under Section 101. Under Alice, courts first determine whether claims are “directed to” an abstract idea, and if so, whether the claims contain an “inventive concept” sufficient to transform the abstract idea into patent-eligible subject matter.
The Recentive patents claimed methods for using machine learning to predict television viewership by analyzing historical data and audience behavior patterns—certainly a valuable commercial application. However, the Federal Circuit found these claims generally directed to applying machine learning to a data set, and went on to criticize the patents for their lack of detail. This is what proved fatal to Recentive’s patents: Not the use of AI itself, but rather how the patents’ specifications presented that use. The court specifically noted that the patent failed to describe how the claimed invention improved machine learning models or achieved technological advancements beyond applying generic machine learning to a new data environment.
A Pattern Across AI and Software Patents
The reaction to Recentive mirrors the response that followed the Federal Circuit’s 2019 decision in American Axle, where many were surprised that technology for creating axle vibration dampeners could be found patent-ineligible. But in both cases, the patent specifications told stories that inadvertently encouraged readers—including the judges in those cases—to conclude that the only innovation lay in patent-ineligible mathematical relationships rather than technological advancement. The American Axle patent focused on new mathematics to be used in the existing context of dampening, rather than telling the story from the perspective of improved dampeners.
The 2022 district court decision in Health Discovery Corp. followed similar reasoning, finding that an early AI patent focused too heavily on the mathematics underlying feature selection rather than its practical applications and technological benefits. Critically, the court noted that while the specification described improvements in data quality relative to conventional mathematical methods, it failed to articulate concrete technological improvements that would support patent eligibility. The specification emphasized mathematical and statistical advantages without explaining how the claimed invention improved computer functionality or solved computer-specific technical problems.
The Spring 2021 Federal Circuit decisions in the Stanford cases reinforced this analytical framework, finding bioinformatics analysis patent-ineligible as mathematics applied to a specific context. The Federal Circuit specifically criticized the specifications for focusing on the mathematical relationships underlying haplotype analysis rather than technological innovation. The court noted that the patents described computerized statistical methods but failed to explain how the claimed inventions improved computer systems or solved technical problems beyond implementing known mathematical concepts on generic computer hardware.
The CardioNet Success Story
These outcomes can be contrasted with the Federal Circuit’s 2020 decision in CardioNet. The CardioNet patent involved a cardiac monitoring device that detected beat-to-beat timing of cardiac activity and determined the relevance of timing variability to specific heart conditions. From a purely technical perspective, this technology might seem vulnerable to the same eligibility challenges—signal processing fundamentally involves mathematical operations on data. But the CardioNet specification told a compelling story of technological innovation that distinguished it from mere mathematical analysis.
Rather than focusing primarily on algorithms, the CardioNet specification explained how the claimed invention achieved meaningful technological improvements with concrete, measurable benefits. These included more accurate detection of specific medical conditions, reduction of false positives and negatives, and enhanced suitability for monitoring ambulatory patients. The Federal Circuit emphasized that the written description successfully demonstrated that the claims were directed to a specific technological improvement by documenting tangible advantages in solving real-world technical problems.
Crafting the Right Narrative
Section 101 analysis can sometimes appear inconsistent and unpredictable. Since the Supreme Court’s 2010 Bilski decision, though, it has increasingly seemed that outcomes depend significantly on how patents cast their innovations—and this represents an area where skilled prosecution and deep familiarity with the subject matter can make a decisive difference.
The most effective specifications focus on identifying and explaining the technological difficulties that existed before the invention and describing how the claimed technology overcomes specific technical hurdles. They demonstrate why the invention represents genuine technological advancement rather than routine application of existing technologies. Perhaps most importantly, they explain the technology clearly to readers who may lack deep technical expertise—recognizing that patent examiners and judges need sufficient context to understand why the claimed invention merits patent protection.
Strategic Implications for AI Innovation
The difference between CardioNet‘s success and Recentive‘s failure illustrates that patent eligibility for AI technologies doesn’t depend on the sophistication of the underlying algorithms, but rather on how effectively the patent connects those algorithms to concrete technological advancement. Both patents involved processing data and generating useful results, but CardioNet emphasized practical technological problem-solving and improvements over prior technologies, while Recentive remained focused on business analytics without adequately demonstrating technological contribution.
This suggests that successful AI patent prosecution requires counsel to engage in substantive technical discussions with inventors, examining not just what the technology accomplishes, but why it represents technological advancement worthy of patent protection. Such conversations prove most productive when counsel genuinely comprehends the technology being discussed and understands how patents will support the client’s business objectives.
The Recentive decision ultimately reinforces a fundamental principle: Simply applying AI to existing problems doesn’t automatically create patentable subject matter. Success requires articulating the myriad ways in which your specific application represents genuine technological advancement, and not merely application of existing technology to a business case or mere mathematical advancement. For AI innovators, this means working with counsel who can understand your technology deeply enough to identify and present the technological contributions that distinguish meaningful innovation from routine engineering—and who can craft specifications that persuasively communicate that distinction to patent examiners and courts.
