AI Patent Risks - Recentive vs Fox Corp (Fed. Cir. 2025)
Hey everyone — Sean Holder here. I want to quickly break down a recent Federal Circuit decision — Recentive Analytics v. Fox — and why it could have big implications for anyone working in AI or machine learning patents.
The court acknowledged that “this case presents a question of first impression” and specifically held 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 section 101.”
I’m going to be honest, I hate this case. This case is troubling for what I call “AI lite” patents, or in other words patents where the original idea of the AI invention is not particularly technical in nature. This could create issues for essentially any AI patent that just processes data without some tie into to a specific machine. I’ve already had an Examiner cite this case in a financial related AI patent. I am really skeptical of this decision, though.
Is it bad law?
Does it just have really bad facts?
Let’s get into the details.
In Recentive, the court invalidated several patents covering systems that used machine learning to generate optimized event schedules and network maps for live television broadcasts. The claims were divided into two categories: one set focused on dynamically generating schedules for live events — like concerts or sports games — based on input parameters such as venue availability, ticket prices, and expected revenue. The other set focused on producing real-time, market-specific broadcast network maps — essentially determining what content should air on which station in which city, at what time — using machine learning to optimize for factors like overall ratings or audience engagement.
The patents claimed that the systems used machine learning models, like neural networks or support vector machines, trained on historical data to identify patterns between input features and desired outcomes. These models were then retrained and updated in real time as new data became available, allowing the system to adjust schedules or broadcast maps on the fly. The inventors argued this enabled more efficient and responsive planning than traditional human-driven methods.
Here is a quick look at a representative claim:
[discussion of claim]
The court concluded the patents were directed to abstract ideas — namely, producing network maps and event schedules — and that the use of machine learning added nothing inventive.
The courts main arguments are that 1) the machine learning technology described in the patents is conventional, 2) the fact that the model is iteratively updated with real time data is not an improvement but just a conventional machine learning technique.
There are a few difficult facts here to consider as well. The court used Recentive’s own admissions against them. For example, Recentive admitted that much of the claims relate to processes that were previously done manually by humans. They admitted that that machine learning in general necessarily includes an interactive training step. They also admitted that their patent did not make machine learning better. In essence, the court takes these admissions to mean that all this patent is, is just an attempt to claim a conventional process done by humans but using a generic machine learning model to do so. The claims were pretty functional, and basically the court wanted to see the claims more clearly show some kind of technical improvement.
I want to end with some questions here. What if there weren’t so many of these admissions here? In retrospect, it seems like it was probably a bad move to frame the invention as doing something that people previously did manually. It would have been better to have shown what the technical solution was previously, and then show how this technical solution is an improvement to that prior technical solution.
Also, what on earth is a generic machine learning model? That sounds non-sensical to me. There is no such thing as a generic machine learning model. The obvious comparison here is to prior law in which a generic computer was added to a claim.
Unlike a generic computer processor, which simply executes instructions, a machine learning model is shaped by the data it’s trained on. That training process creates a unique internal structure—weights, decision boundaries, and learned relationships—that didn’t exist before. The trained model becomes a data-derived artifact, not a plug-and-play tool that can execute any arbitrary program like a processor can. Recentive’s system didn’t just use machine learning—it trained models to optimize schedules and maps based on user-defined priorities and real-time changes. That’s not trivial. It requires domain-specific engineering, feature selection, and dynamic updating logic. In my opinion, just reciting what data is used to trained the machine learning model and what the inputs / outputs of the model are should be enough to get you out of the realm of a generic machine learning model. For example, you could take the same generic Resnet model but if it is trained at all, it will not be a copy of any other Resnet based model, will it? Its like taking a generic process, adding some special logic to it so that it can only function in a certain way and then saying that yes this a generic processor that can be used for any purpose.
I’m also taking beef with the idea that iterative training is conventional machine learning. The court seems to completely equivocate between two entirely different parts of training a model. There is in fact a difference between the initial training of a model and then later updating a model used in production with real time data.
Certainly initial training seems standard. You use historical data to build a model and deploy it. But retraining a model in production — that’s a whole different beast.
Retraining in production is not universal and therefore not really generic:
So, anyways, I think it was wrong for the court to completely blur the distinction between those two types of iteration. There are actually a couple of patent eligible examples the USPTO has put forth that include the use of retraining a model. Example 39 and 47 both include retraining, and I have heard examiners literally point to example 39, the fact that includes retraining, as a guidepost for getting a difficult machine learning related claim allowed.
To wrap it up, the biggest risk here is that Recentive becomes a template for rejecting AI patents that clearly deserve protection—not because they lack real innovation, but because the legal standard is being misapplied. If courts and examiners start treating any applied use of machine learning as “generic,” companies working on practical AI—especially in fields like finance, logistics, and media—may face an uphill battle just to get meaningful and broad claims allowed. That doesn’t just raise costs. It forces inventors to water down their claims, abandon broad protection, or give up entirely.
Anyways, that’s just my two cents.
Thanks for watching. If you’re building AI systems and want to talk about how to protect them — the right way — let’s connect. I am leader in the field of AI related patents with over a decade of experience drafting and prosecuting AI patents in a large variety of technologies for Ivy League schools, top 5 tech companies, fintech companies, multinational scientific companies, and more.