USPTO Director Gives Adress on Subject Matter Eligibility for AI

Artificial intelligence and U.S. patent law have been doing an awkward dance for years. One partner says, “This is groundbreaking technology.” The other squints at the claims and mutters, “Looks a bit abstract to me.” That tension is exactly why the recent USPTO director’s address on subject matter eligibility for AI landed with such force. It was not just another speech stuffed with polite applause lines and coffee-fueled optimism. It was a message about where the U.S. Patent and Trademark Office wants to steer the conversation on AI patent eligibility, and why Section 101 should not become the legal equivalent of a trapdoor every time machine learning shows up in a patent application.

The big takeaway is simple: the USPTO is signaling that AI inventions should not be treated as automatically suspicious, automatically abstract, or automatically doomed. Instead, they should be evaluated using the same legal framework applied to other technologies, with careful attention to whether the claims recite a real technological improvement, a practical application, or a specific solution to a technical problem. In other words, the Office is trying to move the discussion away from “AI is scary” and toward “What exactly did the inventor improve?” That may sound obvious, but in patent law, obvious things often require several memos, a landmark case, a panel decision, and at least one room full of lawyers nodding very seriously.

Why the Director’s Address Matters

Director John A. Squires’s remarks stood out because they framed patent eligibility as a threshold issue that has too often been stretched beyond its intended role. His message was that Section 101 is supposed to identify whether an invention fits within the statutory categories of patentable subject matter, while the heavier lifting on novelty, nonobviousness, and claim scope belongs to Sections 102, 103, and 112. For AI applicants, that is a meaningful distinction. It suggests that the question should not be whether an invention uses math, models, or algorithms somewhere in the background. The real question is whether the claimed invention applies those tools in a concrete, technologically meaningful way.

That theme matters because AI claims are especially vulnerable to being summarized at too high a level. Once an examiner or tribunal paraphrases a claim as “using machine learning to do X,” the invention can start to look like a dressed-up abstract idea. But a claim drafted with enough technical substance may actually describe a new training approach, a better model architecture, an improved data structure, reduced memory usage, better signal processing, stronger security detection, or more accurate treatment personalization. The director’s address effectively said: stop sanding down the details until everything looks abstract.

That is why the speech hit a nerve. It did not announce that every AI claim is patent-eligible. Far from it. Instead, it argued that AI should be analyzed through the normal legal lens, and that the Office should do a better job of distinguishing genuine technical advances from claims that merely chase a result on a generic computer. That is a much more sensible place to be.

The Foundation: USPTO’s 2024 AI Eligibility Guidance

The director’s remarks did not appear out of thin air. They build on the USPTO’s 2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence. That update was important because it translated existing Section 101 doctrine into practical instructions for AI-related examination. It did not invent a special AI test. Instead, it clarified how the existing framework applies to inventions that happen to involve artificial intelligence.

The update focused heavily on Step 2A of the eligibility analysis, especially the distinction between merely reciting a judicial exception and integrating that exception into a practical application. That sounds like classic patent-law soup, but the basic idea is user-friendly enough: if a claim just states a mathematical concept or abstract process and tells a generic computer to run it, the claim is in trouble. If the claim applies that concept in a way that improves computer functionality or another technical field, the story changes.

The USPTO also released new examples that became must-read material for practitioners. Example 47 dealt with anomaly detection using an artificial neural network. Example 48 covered AI-based speech signal analysis and separating desired speech from background noise. Example 49 addressed AI-assisted personalization of medical treatment for a particular patient. These examples were not window dressing. They showed applicants and examiners where the line might fall between an abstract idea and a specific technical application.

The practical lesson from those examples is that AI patent eligibility often turns on specificity. A claim that merely says “use AI to detect anomalies” may collapse into abstraction. A claim that explains how the model interacts with a technical environment, how it improves signal processing, how it changes machine operation, or how it supports a particular treatment decision starts to look much more like eligible subject matter. In patent prosecution, details are not garnish. They are dinner.

What the Director Was Really Saying About Section 101

The address on subject matter eligibility for AI was notable because it treated Section 101 as a gatekeeper, not a wrecking ball. That distinction is crucial. For years, AI, software, diagnostics, and fintech inventions have all struggled with the same problem: courts and examiners sometimes use eligibility doctrine to do work that is arguably better handled elsewhere in the statute. When that happens, nuanced questions about claim scope, technical contribution, and supporting disclosure get flattened into a blunt “abstract idea” label.

Squires’s comments suggested a different approach. He emphasized that Congress wrote Section 101 in broad terms for a reason. Patent law was not designed only for gears, levers, and 19th-century machinery. It was written to cover future technologies too. That means the existence of AI should not itself trigger a special skepticism tax. The Office’s task is to identify whether the claims reflect a practical technological improvement, not to panic every time a neural network enters the chat.

He also leaned on familiar guideposts: the statutory definition of “process,” the logic of cases like Enfish, and the idea that eligibility must look for “something more” than a bare abstract concept. Read together, that framework sends a clear signal. A claim directed to an improved way of training a machine learning model, preserving prior knowledge, reducing system complexity, improving computational efficiency, or solving a concrete technical bottleneck should not be dismissed merely because mathematics sits under the hood. Most advanced technology uses mathematics. If math alone were disqualifying, half of modern engineering would need a support group.

The Case That Put Real Weight Behind the Message: Ex parte Desjardins

The speech mattered even more because it lined up with concrete USPTO action. In Ex parte Desjardins, the Appeals Review Panel addressed claims directed to training a machine learning model in a way that allowed the system to learn new tasks while preserving performance on earlier tasks. That is not just a fancy sentence for investor decks. It goes to the technical problem known as catastrophic forgetting, which is a genuine challenge in machine learning systems.

The panel treated the invention as more than a generalized algorithm. It focused on whether the claims reflected a technological improvement in the operation of the machine learning model itself. That was a big deal. The decision rejected the sort of overgeneralization that turns every model-related invention into “just an algorithm on a computer.” Later, when the USPTO designated the decision as precedential, the message became harder to ignore: improvements in machine learning technology can qualify as practical applications under the Office’s Section 101 framework.

That does not mean applicants can write vague AI claims and hope for a warm hug from the Patent Office. The claims still need to reflect the improvement. The specification still needs to explain the technical problem and how the invention solves it. Conclusory language will not save weak drafting. But Desjardins gave applicants something valuable: a clear, modern example showing that the Office is willing to recognize machine learning improvements as patent-eligible when the claims are framed properly.

The Other Side of the Coin: The Federal Circuit’s Recentive Decision

Of course, the AI eligibility story is not all sunshine, polished conference lecterns, and optimistic administrative guidance. The Federal Circuit’s decision in Recentive Analytics, Inc. v. Fox Corp. is the chilly gust reminding everyone that Section 101 is still a live hazard. In that case, the court held that claims applying existing machine learning models to new data environments, without disclosing improvements to the models themselves, were patent-ineligible.

That ruling matters because it sharpens the difference between two categories of AI inventions. The first category is “we used machine learning here.” The second is “we improved machine learning here” or “we applied machine learning in a concrete technical way that changes how a system functions.” The first category faces real risk. The second has a much better shot.

So while the USPTO director’s address was encouraging, it was not a magic eraser for case law. Applicants still need to draft with the courts in mind. That means claiming technical architecture, technical steps, model behavior, parameter management, signal transformations, hardware interactions, system constraints, and measurable performance improvements where possible. Patent eligibility for AI is stronger when the invention feels like engineering rather than aspiration.

What Patent Applicants Should Do Now

1. Describe the technical problem like an engineer, not a marketer

If the application reads like a startup pitch deck promising to “leverage AI for smarter outcomes,” expect turbulence. A strong AI patent application should explain the bottleneck, the prior limitations, and the technical reason the invention improves system performance. Talk about latency, memory usage, noise reduction, task transfer, training instability, packet analysis, false positives, model drift, or treatment calibration. Give the examiner something concrete to hold onto.

2. Make the claims reflect the improvement

This is where many otherwise smart applications get into trouble. The specification may contain a lovely explanation of the technical advance, but the claims drift upward into high-altitude vagueness. If the claims do not recite the meaningful features that produce the improvement, the invention can still be characterized as abstract. The claim language needs to carry the technical story, not merely wave at it from across the room.

3. Avoid result-only claiming

Claims that say “classify data more accurately,” “optimize performance,” or “generate recommendations” without specifying how the invention achieves that result are easy targets under Section 101. The better approach is to recite a particular way of getting there. Patent law likes mechanisms more than wishes.

4. Separate eligibility from inventorship

Another recurring mistake is blending Section 101 issues with AI inventorship issues. These are related only in the sense that they both involve AI and both can ruin someone’s afternoon. Subject matter eligibility asks whether the claimed invention is the kind of thing patent law protects. Inventorship asks whether a human made the legally significant inventive contribution. A well-prepared filing strategy addresses both, but they are not the same question.

5. Use evidence where it helps

The USPTO’s later guidance on Subject Matter Eligibility Declarations shows that applicants may support the record with evidence tailored to eligibility issues. That can matter where the technical improvement is real but not immediately obvious from a cold reading of the claims. No, evidence does not replace careful drafting. But it can help prevent an invention from being reduced to a caricature.

Why Stakeholders Wanted More Clarity

Industry groups and legal organizations generally welcomed the USPTO’s effort to clarify AI subject matter eligibility, but they also asked for more. That was a predictable reaction. AI develops fast, and Section 101 doctrine does not exactly move like a Formula 1 car. Groups such as AIPLA and IPO pushed for iterative guidance, additional examples, and clearer explanations of how examiners should distinguish genuine technological improvements from claims that merely invoke AI at a high level.

That pushback was healthy. It reflected a practical reality: examiners, applicants, investors, and courts all need a framework that is predictable enough to support real innovation decisions. If nobody can tell whether an AI invention is patent-eligible until after years of prosecution and appeals, the system stops functioning like an incentive structure and starts behaving like an expensive weather forecast.

What This Means for the Future of AI Patent Eligibility

The director’s address did not settle the law. Courts still matter, and judicial precedent remains the final boss in Section 101 fights. But the speech did something important: it reframed the discussion inside the USPTO and signaled a more grounded, technology-aware approach to AI. Together with the 2024 guidance, Desjardins, later MPEP updates, and the availability of targeted evidentiary submissions, the Office appears to be building a more coherent path for applicants whose inventions genuinely improve computing or another technical field.

The biggest lesson is that patent eligibility for AI is neither dead nor automatic. It is claim-driven, fact-sensitive, and deeply dependent on whether the invention is presented as a specific technological solution rather than a broad functional ambition. That may frustrate people who want a one-sentence answer. But honestly, one-sentence answers are what got everyone into this mess in the first place.

For companies developing AI tools, infrastructure, diagnostics, security systems, and machine learning platforms, the moment calls for disciplined optimism. The door is not nailed shut. But you still need the right key.

Experience From the Field: What This Topic Looks Like in Real Life

Talk to enough patent prosecutors, founders, and in-house counsel, and you will hear the same theme in different accents: AI patent eligibility is no longer a pure “yes or no” debate. It feels more like a drafting discipline. The companies that do well are usually the ones that stop talking about AI as magic and start talking about it as engineering.

One common experience comes from early-stage startups. A founder will walk into a patent strategy meeting excited about a model that predicts something valuable, whether that is equipment failure, credit risk, fraud, disease progression, or customer churn. The first instinct is often to claim the business value: better predictions, faster decisions, lower cost. But when those ideas get translated into patent language, they can sound like an abstract goal sitting on generic computing infrastructure. The conversation changes once the team focuses on the mechanism: How is the model trained differently? What data pipeline changed? What technical bottleneck is solved? Did memory usage drop? Did signal quality improve? Did the system preserve prior learning while adding new tasks? Suddenly the invention starts to look patent-ready instead of presentation-ready.

In-house patent teams report a similar shift. They are increasingly training product and research teams to document technical improvements as they happen, rather than reconstructing them months later from Slack messages, whiteboard photos, and heroic memory. That means preserving benchmarking data, architectural diagrams, ablation results, and design rationales. It also means recording why a particular model configuration or training sequence solved a technical problem better than prior approaches. When prosecution starts, that record can make the difference between “this sounds abstract” and “this is a concrete improvement to computer technology.”

Outside counsel see another pattern: many AI applications rise or fall on the specification. If the specification gives only a broad overview with generic references to processors, models, inputs, and outputs, then even a clever claim set may wobble under Section 101. But when the specification explains a precise technical problem and a detailed solution, it becomes easier to draft claims that survive scrutiny and easier to argue that the invention fits within the practical-application side of the analysis. In plain English, the best Section 101 arguments are usually written six months before the first rejection ever arrives.

There is also a human side to all of this. Inventors often feel confused when they hear that a sophisticated AI system might be labeled “abstract.” From their perspective, they built a real product, solved a real problem, and spent real money doing it. That reaction is understandable. The challenge is that patent law is not judging whether the invention is impressive. It is judging whether the claims capture the technical contribution in a legally recognizable way. Once inventors understand that difference, the drafting process gets better, the claims get sharper, and the odds improve. It is not glamorous, but neither is debugging a training pipeline at 2:00 a.m., and innovation survives that too.

Conclusion

The USPTO director’s address on subject matter eligibility for AI matters because it confirms a broader shift in tone and practice: AI inventions should be evaluated with precision, not prejudice. The Office is signaling that real technical improvements in machine learning, model training, signal processing, data structures, and applied systems deserve a fair hearing under Section 101. At the same time, the courts continue to warn that broad claims to using machine learning in a new context will not sail through just because the acronym is trendy.

The smartest takeaway is not blind optimism or blanket pessimism. It is disciplined drafting, evidence-backed disclosure, and a claim strategy that shows how the invention changes technology rather than merely announcing a desired outcome. In the current landscape, AI patent eligibility is alive, evolving, and still allergic to fluff. Frankly, that may be the healthiest thing anyone has said about Section 101 in years.

SEO Tags