Federal Circuit’s Question of First Impression on the Patent Eligibility of Machine Learning
| May 22, 2025
Case Name: : Recentive Analytics, Inc. v. Fox Corp.
Date of Decision: April 18, 2025
Before: Dyk, Prost, Goldberg
Summary:
In this precedential opinion, the Federal Circuit addressed “a question of first impression: whether claims that do no more than apply established methods of machine learning to a new data environment are patent eligible.” The court held that claiming the application of generic machine learning techniques in a new environment or new field of use, without specific improvements to the machine learning techniques or models themselves, is not patent eligible subject matter.
Procedural History:
Recentive sued Fox for patent infringement. The district court granted Fox’s motion to dismiss the case for failure to state a claim due to the patents being ineligible under 35 USC §101. Recentive appealed.
Background:
Recentive asserted two groups of patents. One group of “Machine Learning Training” patents include USP Nos. 11,386,367, 11,537,960. Another group of “Network Map” patents include USP Nos. 10,911,811 and 10,958,957.
Claim 1 of the ‘367 patent below is representative of the Machine Learning Training patents.
1. A computer-implemented method of dynamically generating an event schedule, the method comprising:
receiving one or more event parameters for series of live events, wherein the one or more event parameters comprise at least one of venue availability, venue locations, proposed ticket prices, performer fees, venue fees, scheduled performances by one or more performers, or any combination thereof;
receiving one or more event target features associated with the series of live events, wherein the one or more event target features comprise at least one of event attendance, event profit, event revenue, event expenses, or any combination thereof;
providing the one or more event parameters and the one or more event target features to a machine learning (ML) model, wherein the ML model is at least one of a neural network ML model and a support vector ML model;
iteratively training the ML model to identify relationships between different event parameters and the one or more event target features using historical data corresponding to one or more previous series of live events, wherein such iterative training improves the accuracy of the ML model;
receiving, from a user, one or more user-specific event parameters for a future series of live events to be held in a plurality of geographic regions;
receiving, from the user, one or more user-specific event weights representing one or more prioritized event target features associated with the future series of live events;
providing the one or more user-specific event parameters and the one or more user-specific event weights to the trained ML model;
generating, via the trained ML model, a schedule for the future series of live events that is optimized relative to the one or more prioritized event target features;
detecting a real-time change to the one or more user-specific event parameters;
providing the real-time change to the trained ML model to improve the accuracy of the trained ML model; and
updating, via the trained ML model, the schedule for the future series of live events such that the schedule remains optimized relative to the one or more prioritized event target features in view of the real-time change to the one or more user-specific event parameters.
The Machine Learning Training patents relate to the scheduling of live events and involve “the unique application of machine learning to generate customized algorithms, based on training the machine learning model, that can then be used to automatically create … event schedules that are updated in real-time.”
Claim 1 of the ‘811 patent below is representative of the Network Map patents.
1. A computer-implemented method for dynamically generating a network map, the method comprising:
receiving a schedule for a first plurality of live events scheduled to start at a first time and a second plurality of live events scheduled to start at a second time;
generating, based on the schedule, a network map mapping the first plurality of live events and the second plurality of live events to a plurality of television stations for a plurality of cities,
wherein each station from the plurality of stations corresponds to a respective city from the plurality of cities,
wherein the network map identifies for each station (i) a first live event from the first plurality of live events that will be displayed at the first time and (ii) a second live event from the second plurality of live events that will be displayed at the second time, and
wherein generating the network map comprises using a machine learning technique to optimize an overall television rating across the first plurality of live events and the second plurality of live events;
automatically updating the network map on demand and in real time based on a change to at least one of (i) the schedule and (ii) underlying criteria,
wherein updating the network map comprises updating the mapping of the first plurality of live events and the second plurality of live events to the plurality of television stations; and
using the network map to determine for each station (i) the first live event from the first plurality of live events that will be displayed at the first time and (ii) the second live event from the second plurality of live events that will be displayed at the second time.
The Network Map patents relate to the creation of network maps for broadcasters, which, prior to computers, was done by humans to determine the types of content to be displayed on particular channels at various times. The Network Map patents use training data and user-provided target features, such as overall ratings for the NFL for particular stations, markets, or audience, to generate optimized network maps.
However, Recentive acknowledged that they do not claim the machine learning technique itself and that any machine learning technique may be employed. Rather, the claims use iterative training on its machine learning model with “different event parameters and … event target features” provided by users to “identify relationships within the data.” But, Recentive also acknowledged that “[t]he process of training the machine learning model[] … is required for any machine learning model” “and then the algorithm is actually updated and improved over time based on the input” or training.
Under Alice Step 1, the district court found that the patents were “directed to the abstract ideas of producing network maps and event schedules… using known generic mathematical techniques.” Under Alice Step 2, the claims are not directed to any inventive concept “because the machine learning limitations were no more than ‘broad, functionally described, well-known techniques.’”
Decision:
At Alice step 1, the court examines the “focus of the claimed advance over the prior art.” Here, by Recentive’s own representations, iterative training or dynamic adjustment based on real-time user-specified parameters are “incident to the very nature of machine learning.” Recentive argues that its claims “unearth ‘useful patterns’ that had previously been buried in the data, unrecognizable to humans.” However, Recentive itself acknowledged that its patents do not claim any specific improvements in mathematical algorithms used or improvements to the machine learning itself. Neither does the specification describe any improvements in machine learning. The court did not find any “specific implementation of a solution to a problem in the software arts.” Instead, the court found that “the only thing the claims disclose about the use of machine learning is that machine learning is used in a new environment…event scheduling and the creation of network maps.” And, the court had long recognized that “[a]n abstract idea does not become nonabstract by limiting the invention to a particular field of use or technological environment.”
At Alice step 2, the court did not find any “inventive concept” that would amount to something “significantly more” than the abstract idea itself of generating event schedules and network maps through application of machine learning.
Nevertheless, the court acknowledged that “machine learning is a burgeoning and increasingly important field and may lead to patent-eligible improvements in technology.” But, it concludes that “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.”
Takeaways:
The court’s decision in this “question of first impression” comes as no surprise. Just using machine learning, or for that matter, any other “tool” of artificial intelligence, such as neural networks, AI models, etc., “in new data environments” is not enough to confer patent eligibility. There must be some claimed and disclosed “improvement” in “software arts” or some other technology.
This decision aligns with the recent USPTO AI examples issued July 2024. For instance, USPTO Example 47, claim 2 and Example 48 were ineligible because the mere “use of” a neural network or trained model is just the mere use of a tool/computer, a general link to a technical field, a mere “apply it” feature – all of which are not meaningful limitations for Alice steps 1 and 2 (as reflected in the USPTO Step 2A, Prong 2 and Step 2B analysis in these examples).
Of more interest is the follow-up exploration into the boundaries of patent eligibility for the “improvements” in such AI-related fields. The July 2024 USPTO AI Examples provided some examples for eligible AI-related subject matter where features in a specific technological field are used/impacted/created via the use of the neural network or trained model that leads to some improvement in that specific technological field. This included USPTO Example 47 claim 3’s dropping of potentially malicious packets and blocking of future traffic from the detected source address, which reflected an improvement in the technical field of network intrusion detection by using the information (source address of malicious packets) to enhance security by taking claimed proactive measures to remediate the danger. Other examples of eligible claims reciting AI features leading to improvements in some underlying technology include USPTO Example 48, claims 2 and 3 (speech signal processing). Of course, these are clear cut examples of eligible AI subject matter. We hope to see more Federal Circuit decisions addressing the more grey areas involving AI techniques.
For instance, Recentive tried to argue that that its claims “unearth ‘useful patterns’ that had previously been buried in the data, unrecognizable to humans.” Recentive also focused on the claimed iterative training as conferring patent eligibility. Perhaps an argument could be made that a specific new way for how to train the machine learning model using selected specific data improved the iterative training procedure that would otherwise apply using existing techniques. This, of course, would require a carefully drafted specification spelling out what an existing iterative training might entail, and then, how this new iterative training would improve the accuracy of the model. Again, what does the specification say about the “focus of the claimed advance over the prior art?” If the “focus” is on the new way to perform the iterative training, perhaps the claims might have survived §101.
This is also reminiscent of USPTO Example 39, reciting the creation of first and second training sets and the training of a neural network in first and second stages with those respective training sets. USPTO Example 39 is an example of an eligible claim involving the training of a neural network for facial detection. One key aspect of Example 39 is that it described how prior methods of facial detection using neural networks suffered from an inability to robustly detect human faces in images when the NN was trained with a set of facial and non-facial images. Example 39, instead, focused on expanded training sets of facial images. Under USPTO Step 2A, Prong 1, the “training” in USPTO Example 39 was simply deemed not to fall into any judicial exception category and the claim was deemed patent eligible because no judicial exception was recited.
Contrast, however, with USPTO Example 47, claim 2, where “training” was deemed to fall into the category of math, and the claim was deemed ineligible. One distinction between these two examples is that USPTO Example 39 focused on “how” the “training” led to improvements over the prior training techniques. Such a focus did not exist in Recentive’s case. But, when it is the first time applying AI in “a new data environment,” a next question might be whether there is a way to describe deficiencies in using existing training techniques that would be overcome with newly created sets of training data or a new specific way of changing existing training procedures to improve a desired outcome. This is just one example of a grey area where more Federal Circuit decisions on eligibility would be helpful. Nevertheless, the guidance is clear – spell out the “improvement” in technology in the specification and claim the features that result in that improvement.
