From Black-Box AI to Technical Improvement: Where USPTO §101 Is Really Moving

Recent debate around the USPTO and Section 101 has made one point sound simpler than it really is: that an AI patent claim becomes easier to defend as soon as the application says the invention involves model training. The official materials tell a narrower and more useful story. From the 2024 AI subject matter eligibility update, to the August 2025 reminder memo, to the late-2025 MPEP change prompted by Ex parte Desjardins, the clearer direction is not a special AI shortcut. It is that claims framed as a black box that takes in data and produces a result remain vulnerable to abstraction, while claims that reflect a concrete improvement in how the model or system actually operates stand on firmer ground in the Section 101 analysis.

That distinction has real prosecution consequences. An application may contain pages of technical background, but if the claim is ultimately drafted as little more than “receive data, train or infer, output a result,” the applicant is still likely to face the familiar problem of an abstract idea implemented on a computer. What helps more is not a broad statement that the model is more accurate or more efficient, but a claim set that shows where the technical improvement lives and how it changes the internal operation of the system.

What has actually changed in the official materials

The recent shift is easier to understand when separated into layers. The 2024 AI eligibility update used new examples to show that an AI-related claim does not pass Section 101 simply because it mentions neural networks or training. The question remains whether the claim is really directed to an abstract idea, or whether it integrates that idea into a practical application. The August 2025 memo then reinforced several points that matter in software and AI examination: examiners should not overextend the “mental process” grouping, they should distinguish claims that recite a judicial exception from claims that merely involve one, and they must evaluate the claim as a whole when asking whether there is a practical application.

Ex parte Desjardins and the subsequent MPEP update pushed that discussion closer to machine learning itself. The USPTO did not bless AI as a special category. Instead, it accepted a more specific kind of argument: that the claimed invention improved how the learning model operated by preserving performance on earlier tasks while learning new ones, while also reducing storage demands and system complexity. That is an important signal. The doctrinal framework did not change, but the Office is now drawing a clearer line between claiming an AI result and claiming a technological improvement in AI machinery.

Why black-box drafting is becoming a bigger liability

Many AI applications do have real technical content. The problem is that the claims often leave that content behind. If the filing mainly says that some data is received, processed by a model, and turned into a classification, recommendation, score or prediction, the invention becomes easier to characterize at a high level as pattern recognition, data analysis or mathematical processing. Once the claim is framed that way, the applicant is forced back into the familiar Section 101 dispute over whether the computer is doing anything more than applying an abstract idea.

That is the practical danger of black-box drafting. Better accuracy, lower error rates and faster processing may describe the business value of the invention, but they do not necessarily tell the examiner what the technical mechanism is. Without that mechanism, eligibility arguments tend to weaken quickly. In AI cases especially, it is often not enough to present the model as a container that produces a better output. Claims fare better when they identify the training sequence, parameter-update constraints, penalty terms in the objective function, interactions with hardware resources, or the way a model output triggers a further technical operation in real time.

How the specification and the claims need to work together

One of the more useful lessons from Desjardins is that it is not enough for the specification to declare that there is a technical improvement. The claims still need to reflect it. The USPTO is not rewarding slogans. It is looking for correspondence. If the specification reads like a technical breakthrough but the claims reduce the invention to a generic data-processing flow, the application remains exposed even if the underlying engineering work was sophisticated.

For training-related or data-intensive AI inventions, the most persuasive disclosures usually explain where the technical bottleneck arises, which parameters, states or intermediate representations must be preserved or constrained, how training data is organized, filtered or staged, why a particular update rule or loss function addresses the problem, and how that choice changes system behaviour in terms of storage, stability, robustness or real-time performance. Dataset disclosure still matters, but not because it creates a new eligibility shortcut. It matters when it supports a concrete and defensible technical mechanism that can be carried into the claims.

Practical takeaways for drafting and prosecution

For new filings, it is safer to treat Section 101 as a drafting architecture issue from the outset rather than a later-response problem. At least one independent claim should be built around the technical mechanism itself, not only the end result. Where the invention also has a system-level interaction, applicants should consider parallel claim positions tied to a defined technical environment, data-flow control, device behaviour or a downstream technical action. That gives the “practical application” analysis something concrete to attach to.

For pending cases already facing a Section 101 rejection, responses usually improve when they do less high-level advocacy and more element-by-element mapping. Which claim limitation reflects which technical improvement in the specification? How does that improvement change the operation of the model or the relevant technical system? Where the evidentiary record supports it, a Rule 132 subject matter eligibility declaration may help explain how a skilled person would understand the improvement, but it cannot repair missing technical disclosure after the fact. The USPTO’s recent direction is not rewarding polished AI rhetoric. It is rewarding applications that identify a technical improvement, explain it with specificity, and actually claim it.