USPTO ARP Affirms Patent Eligibility for DeepMind AI Training Met

Patent law and artificial intelligence have spent the last few years acting like two brilliant people at a dinner party who keep talking past each other. One says, “This is a real technical advance.” The other says, “Cool story, still looks abstract.” In Ex parte Desjardins, the USPTO’s Appeals Review Panel, or ARP, delivered one of the clearest recent signals that AI-related inventions are not automatically doomed under Section 101 just because math shows up to the party wearing a name tag.

The case involved a DeepMind patent application directed to training machine learning models across multiple tasks while preserving performance on earlier tasks. In plain English, the invention tackled a classic headache in AI: teaching a model something new without making it forget what it already knew. If that sounds suspiciously like how humans feel after opening one too many browser tabs, welcome to continual learning.

The ARP’s decision matters because it did not say all AI claims are patentable. It said something more useful and more legally important: when AI claims are framed as a concrete technical improvement to how the model itself operates, those claims can clear the patent-eligibility hurdle. That is a big deal for companies building model-training systems, infrastructure, and optimization techniques in the United States.

What Happened in the DeepMind Appeal?

The underlying patent application described a computer-implemented method for training a machine learning model on multiple machine learning tasks. The claims focused on determining how important certain parameters were to an earlier task, then adjusting those parameters during training on a later task in a way that protected prior performance. In other words, the method tried to reduce catastrophic forgetting while still allowing the model to learn something new.

That sounds technical because it is technical. But the Board had previously added a new Section 101 rejection, reasoning that part of the claimed method involved a mathematical calculation and therefore fell into abstract-idea territory. The ARP agreed that some claim language did recite an abstract idea at a high level. Then it took the next, crucial step: it looked at the claims as a whole and asked whether the supposed abstraction had been integrated into a practical application.

Its answer was yes. The ARP concluded that the claims reflected an improvement in the training of the machine learning model itself. That point is the star of the show. The panel did not treat the invention as “math in a trench coat pretending to be software.” It treated the invention as a technical way to improve model behavior across sequential training tasks.

Why the Technical Problem Mattered

The DeepMind application did not describe a generic AI system sprinkled over a business workflow like parsley on a fancy plate. It targeted a recognized technical problem in machine learning: catastrophic forgetting. In neural networks, training on new tasks can overwrite knowledge learned from old ones. That makes continual learning difficult, especially when one model needs to handle multiple tasks over time without exploding storage demands.

DeepMind’s own research has long focused on this problem. The core idea is intuitive even if the implementation is not: some parameters matter more than others for earlier tasks, so later training should be more careful with those parameters. The patent application translated that idea into claim language centered on parameter importance, constrained updates, preserved performance, and improved efficiency.

That framing helped the invention look less like an abstract equation floating in outer space and more like a technical mechanism for improving machine learning systems. The application also tied the method to practical benefits, including reduced storage requirements and reduced system complexity because a single model could retain acceptable performance across multiple tasks instead of requiring multiple separate model instances.

Why the ARP Said the Claims Were Patent-Eligible

The ARP’s reasoning turned on a familiar but often messy distinction in patent law: a claim that merely uses an abstract idea is different from a claim that integrates that idea into a practical application. The panel said DeepMind’s claims fell into the second bucket.

The claims improved the model itself

This was the headline point. The ARP found that the claims were directed to improving how the machine learning model itself operates. The invention was not just using AI as a black box to get a business result faster, cheaper, or with more buzzwords. It was aimed at improving sequential learning, preserving prior-task performance, reducing storage use, and lowering system complexity.

That distinction is pure gold in AI patent drafting. U.S. patent law has become much more skeptical of claims that simply apply generic computing or generic machine learning to a new field. By contrast, claims that change model behavior, training dynamics, system architecture, or computational efficiency stand on much stronger ground.

The decision leaned on software patent precedent

The ARP relied on familiar Federal Circuit logic from cases such as Enfish and McRO. Those cases are important because they recognize that software innovations can be patent-eligible when they improve computer functionality or another technical field. The ARP effectively said that AI should not be shoved into a separate penalty box just because the claimed invention includes mathematical concepts.

That matters because nearly every serious machine learning invention contains math somewhere in the engine room. Treating math as an automatic deal-breaker would be like denying patent protection to a jet engine because physics was involved. That approach may be dramatic, but it is not especially helpful.

Section 101 is not supposed to do every job in the building

One of the most important parts of the ARP decision was its warning against using Section 101 as a catch-all weapon. The panel emphasized that Sections 102, 103, and 112 remain the traditional tools for testing novelty, nonobviousness, and adequate disclosure. That is a meaningful signal to applicants and examiners alike: if the real complaint is that a claim is obvious, unsupported, or too broad, use the right statutory tool.

That does not make AI patents easy. It just means the door should not be slammed shut at the eligibility stage when the claims describe a real technical improvement.

This Was Not a Total Victory for DeepMind

Here is the nuance that too many hot takes skip over: the ARP did not say the application must issue as a patent tomorrow with confetti and a brass band. The panel vacated the new Section 101 rejection, but it did not disturb the separate Section 103 rejection. So the eligibility win was real, but it was not the same thing as a final allowance.

That distinction matters for anyone working in AI patent strategy. Passing Section 101 is only one checkpoint. An applicant still needs claims that are novel, nonobvious, and adequately supported. In other words, clearing eligibility gets you into the stadium. You still have to win the game.

Why This Decision Matters for AI Patent Drafting

The ARP decision offers a drafting playbook for applicants who want stronger AI and machine learning patent claims in the United States.

1. Describe the technical problem clearly

Do not just say your invention uses AI. That is about as persuasive as saying your restaurant serves food. Explain the engineering problem. Is the invention reducing catastrophic forgetting, improving inference latency, lowering storage overhead, increasing robustness, enhancing model calibration, or improving resource allocation during training? The more concrete the problem, the better.

2. Tie the benefits to the claimed steps

The specification can sing beautifully about technical benefits, but the claims need to carry the tune. In Desjardins, the ARP was persuaded because the claims themselves reflected the disclosed improvement. That is a crucial lesson. If the specification talks about better efficiency and preserved performance, the claim language should not wander off and talk only in generic functional terms.

3. Show how the model or system changes

Patent eligibility tends to improve when claims explain how the model is trained, constrained, updated, structured, or deployed in a technically meaningful way. Claims are weakest when they simply announce a result and let a generic model do the magic offstage.

4. Expect tougher fights under Sections 103 and 112

Even if the USPTO becomes more careful about overusing Section 101, that does not mean applicants are in for a free buffet. Examiners can still press hard on obviousness and written description. The ARP itself basically reminded everyone that those provisions are where many AI battles should happen.

The Broader Patent Landscape: Desjardins vs. Recentive

The DeepMind decision also stands out because it arrived in a legal environment where courts have sometimes been less enthusiastic about machine learning claims. Earlier Federal Circuit decisions in this area have stressed that merely applying existing machine learning techniques to a new data environment or industry is not enough. That contrast makes Desjardins especially important.

Together, the emerging message is surprisingly coherent: if an AI claim merely uses machine learning as a tool in a familiar way, expect trouble. If it improves the model, training method, system architecture, or another technical aspect of computing, the claim stands a better chance. That does not solve every Section 101 problem, but it gives applicants a roadmap that is far more practical than crossing fingers and sacrificing a keyboard to the patent gods.

What This Means for Startups, Big Tech, and R&D Teams

For startups, the decision is encouraging because patent eligibility can affect funding, valuation, and defensive strategy. Investors tend to like intellectual property that is more than decorative wall art. A clearer path for technical AI claims can help early-stage companies justify prosecution budgets and protect core training innovations.

For larger AI companies, the case reinforces the value of patenting infrastructure-level advances, not just flashy end-user features. Training pipelines, parameter optimization methods, continual-learning systems, model efficiency techniques, and deployment improvements may all deserve closer patent review when they are tied to specific technical gains.

For in-house counsel and outside patent prosecutors, the lesson is even more practical: write AI applications as if every sentence may someday need to explain why the invention improves computer technology itself. Because, frankly, it might.

Experiences From the AI Patent Front Lines

One of the most interesting things about the reaction to the DeepMind decision is how familiar it feels to people who work around AI patents every day. For years, applicants have experienced a strange sort of legal whiplash. On Monday, everyone agrees that model training, parameter control, and system efficiency are hard engineering problems. On Tuesday, those same ideas are sometimes described in prosecution as if they were just fancy algebra in a nicer blazer. The Desjardins ruling speaks directly to that disconnect.

In practice, many AI inventors have learned the hard way that saying “our model performs better” is not enough. Examiners, boards, and later courts usually want to know why it performs better, how the training process changes, and what technical tradeoff the invention improves. Teams that document those points early tend to be in a much stronger position. Teams that wait until prosecution to reverse-engineer a technical story often discover that the record is thinner than a budget airline pillow.

Another common experience is that AI innovation often happens in infrastructure layers that non-specialists never see. A consumer may notice a chatbot answer faster or a vision system miss fewer objects, but the real invention may live in training constraints, parameter regularization, model retention strategies, memory handling, or deployment architecture. The DeepMind case is a reminder that these quieter layers of innovation are often where the strongest patent arguments live. The magic is not always in the app demo. Sometimes it is in the plumbing, and yes, the plumbing deserves legal respect too.

There is also a practical drafting lesson that many patent professionals will recognize immediately: the specification and the claims need to act like old friends, not distant cousins who only meet at weddings. In AI applications, it is common to see a detailed technical description paired with claims that become overly abstract in the name of breadth. That mismatch can be fatal. What made the ARP decision notable was that the panel found the claims actually reflected the disclosed technical improvement. That is the kind of alignment prosecutors dream about and litigators mention with unusual tenderness.

From a business perspective, the experience around AI patents has also been shaped by uncertainty in investment conversations. Founders and R&D leaders regularly ask whether it is still worth filing on machine learning inventions in the United States. The most honest answer has been, “Yes, but only if you can tell a real technical story.” Desjardins does not turn every AI idea into a patent winner, and it does not erase the risk of obviousness rejections. But it does make that technical story more meaningful at the eligibility stage.

Finally, the decision reflects something many people in the field have felt for a while: AI should not be treated as categorically suspicious just because it involves algorithms. Software, models, and training methods can improve computer functionality in the same way other technologies do. When patent law recognizes that point, the system feels a little less like a trapdoor and a little more like what it is supposed to be: a framework for rewarding real innovation.

Conclusion

The USPTO ARP’s decision in the DeepMind appeal is not a blank check for every machine learning claim with the words “neural network” taped to it. But it is an important marker. It confirms that AI training methods can be patent-eligible when they solve a concrete technical problem and improve how a model or computer system actually works.

That is the real takeaway. Patent eligibility for AI in the United States is not dead, not automatic, and not random. It depends heavily on whether the invention is framed as a specific technological improvement rather than a generic use of machine learning in a new setting. For applicants, that means careful drafting. For innovators, it means there is still room to protect real advances in training methods, model efficiency, and system design. And for Section 101, it means the bouncer may finally be checking the guest list with a little more precision.