How to Patent AI and Machine Learning Inventions After Ex Parte Desjardins

On December 5, 2025, the United States Patent and Trademark Office issued a memorandum that will reshape how artificial intelligence and machine learning inventions are evaluated for patent eligibility. The precedential decision in Ex Parte Desjardins, now incorporated into the Manual of Patent Examining Procedure, provides clarity that has been sorely lacking for innovators working at the intersection of mathematics and practical technology. For entrepreneurs and businesses developing AI systems, understanding this framework isn't just academic—it's essential to protecting competitive advantage in an increasingly crowded field.

The Patent Eligibility Challenge

The challenge has always been this: patent law requires that inventions fall within eligible subject matter under 35 U.S.C. § 101, yet courts have broadly interpreted "abstract ideas" as categorically ineligible for patent protection. Mathematical algorithms, the bedrock of machine learning, have been particularly vulnerable to rejection under this doctrine. Neural networks perform mathematical transformations. Optimization algorithms solve mathematical problems. Statistical models apply mathematical principles. Standing alone, these would be considered abstract ideas, and abstract ideas cannot be patented.

The question that has bedeviled AI inventors for years is where the line falls between an unpatentable abstract mathematical concept and a patentable technological innovation. Ex Parte Desjardins finally provides an answer, though perhaps not a simple one. The Patent Trial and Appeal Board's Appeals Review Panel reviewed claims directed to a method of training a machine learning model to address "catastrophic forgetting"—a well-documented technical problem in continual learning systems where models lose their ability to perform previously learned tasks as they adapt to new challenges. When a model's performance on earlier work degrades as it learns something new, you have a real technical limitation, not just a mathematical curiosity.

The claimed solution involved a training method that adjusts parameters to optimize performance on new tasks while actively protecting performance on previous tasks. At first glance, this looks exactly like the kind of mathematical process that routinely gets rejected. The Appeals Review Panel even acknowledged that the claims recited an abstract idea—mathematical concepts—at Step 2A Prong One of the eligibility analysis. But they didn't stop there, and that's what makes Desjardins important.

The Improvement to Technology Test

The Panel asked a different question: Do the claims integrate this abstract idea into a practical application by providing an improvement to computer functionality or another technology? The specification disclosed specific technological improvements—the machine learning model learns new tasks while protecting knowledge about previous tasks, the system reduces storage capacity requirements, system complexity decreases, and performance attributes from earlier tasks are preserved during subsequent computational work. Critically, these improvements weren't just conclusions or aspirational statements. The specification explained how the training method achieved these results, and the claims themselves reflected these improvements through specific limitations, particularly the requirement to "adjust the first values of the plurality of parameters to optimize performance of the machine learning model on the second machine learning task while protecting performance of the machine learning model on the first machine learning task."

The Panel held that these claims integrated what would otherwise be a judicial exception into a practical application that improved machine learning technology. Result: patent eligible.

The December 5th memorandum updates the MPEP to incorporate Desjardins throughout the eligibility analysis sections, and the directives to examiners are instructive. First, evaluate the specification: Does it provide sufficient detail that someone skilled in the art would recognize the claimed invention as improving computer functionality or another technical field? The specification doesn't need to explicitly announce "this is an improvement," but the improvement must be apparent from the disclosure—not merely conclusory. Second, evaluate the claims: Do they reflect the disclosed improvement? The claims must include the components or steps that provide the improvement described in the specification. They don't need to explicitly recite the improvement with language like "thereby increasing bandwidth," but the claim as a whole must reflect it. And crucially, examiners are directed to evaluate claims "as a whole" and "as an ordered combination," without dismissing limitations as "generic computer components" without considering whether those components contribute to a technological improvement.

What Makes AI Inventions Patent Eligible

So what does this mean for AI innovators developing actual technology rather than writing academic papers? The framework that emerges from Desjardins and the MPEP updates suggests that AI inventions are more likely to be patent eligible when they solve concrete technical problems, provide measurable improvements, explain how those improvements are achieved, and draft claims that reflect the improvement. None of these requirements is particularly mysterious, but each demands careful attention.

Consider the technical problem first. Generic statements about "improving accuracy" or "processing data faster" won't satisfy the test. You need to identify the specific technical limitation or challenge in existing systems and explain why it matters. In Desjardins, the problem was catastrophic forgetting in continual learning systems—a well-documented technical issue that prevents machine learning models from maintaining performance across sequential tasks. This isn't a business problem or a user experience problem; it's a technical limitation inherent in how these systems function. That specificity matters.

The improvements need to be concrete and, where possible, quantifiable. Reduced computational requirements matter—storage capacity, processing power, memory usage. Improved performance metrics matter—accuracy gains, speed increases, reliability improvements. Enabled capabilities that were previously impossible matter. Technical challenges that prevented practical implementation being solved matters. In Desjardins, the specification disclosed reduced storage capacity, decreased system complexity, and preserved performance across tasks. These are measurable, verifiable claims about technological improvement, not marketing language about user satisfaction.

But here's where many patent applications fail: they state the improvement without explaining how it's achieved. This is critical and non-obvious to many inventors. Your specification must provide enough technical detail that a person of ordinary skill in the art would understand how your invention achieves the stated improvements. Conclusory statements don't work. Simply declaring "the system improves performance" without explaining the technical mechanism will not satisfy the test. You must describe the specific approach, methodology, or structure that delivers the improvement. The specification needs to enable someone reading it to understand not just what you claim works better, but why your particular technical approach produces that result.

Finally, your claims must actually reflect the improvement. In Desjardins, the claim limitation "adjust the first values of the plurality of parameters to optimize performance of the machine learning model on the second machine learning task while protecting performance of the machine learning model on the first machine learning task" captured the essence of the technical solution. The claim doesn't say "thereby solving catastrophic forgetting," but it includes the specific functional requirement that delivers the benefit. Claims that could cover both improved and non-improved implementations may not reflect the disclosed technological advancement. Your claims should be tailored to capture the specific technical approach that delivers the benefit, even if you don't explicitly recite the benefit itself.

Common Pitfalls in AI Patent Applications

There are predictable ways that patent applications fail the eligibility test, and understanding them helps avoid wasted effort and expense. Many applications focus exclusively on what the algorithm does mathematically without explaining what technical problem is being solved or how the approach improves upon existing methods. This leads to rejections under § 101 because the claims appear to merely implement abstract mathematical concepts with no practical application. Adding "on a computer" or "using a processor" to otherwise abstract claims doesn't cure the problem—examiners are directed not to dismiss limitations as "generic computer components" without analysis, but they still must evaluate whether those components contribute to a technological improvement.

Another common failure mode is the conclusory improvement statement. Declaring "this improves performance" or "this is more efficient" without technical explanation won't satisfy the eligibility test. The specification must explain how the improvement is achieved with enough detail that the improvement would be apparent to someone skilled in the field. And claims that are too broad to reflect the improvement create their own problems—if your claims could cover both improved and non-improved implementations, they may not reflect the disclosed technological advancement. Your claims should be tailored to capture the specific technical approach that delivers the benefit.

Strategic Approach to Drafting

Drafting a patent application that will survive eligibility scrutiny requires thinking about the invention in a particular way from the outset. Before writing anything, you need to clearly articulate what technical problem exists in current AI and machine learning systems, why that problem is significant, what others have tried that didn't work or had limitations, and how your approach solves or mitigates the problem. This isn't marketing language—it's technical problem definition that will form the foundation of your eligibility argument.

Performance data matters more than many inventors realize. Benchmark comparisons, resource utilization metrics, performance measurements, user studies or technical evaluations—this data strengthens both the specification and any arguments you'll need to make during prosecution. If you claim reduced storage requirements, can you quantify the reduction? If you claim improved accuracy, what are the numbers? If you claim faster processing, compared to what? These measurements don't need to appear in the claims, but they should appear in the specification to demonstrate that the improvements are real and measurable.

The specification itself should provide genuine technical depth. Don't assume the examiner understands your technical field at the level you do. Explain the technical problem in detail, describe why existing approaches are inadequate, articulate how your invention addresses the limitation, and explain why your approach provides the stated benefits. The specification should enable someone skilled in the art to understand both what you invented and why it's an improvement. This often means the specification will be longer and more detailed than you initially think necessary, but that detail pays dividends when you face eligibility challenges.

Claims require strategic thinking. Independent claims should include the specific limitations that deliver the improvement, reflect the technical solution described in the specification, and be narrow enough to clearly show the improvement while remaining broad enough to have commercial value. That's a difficult balance, which is why dependent claims matter—they provide fallback positions. Consider including dependent claims that explicitly recite the improvements, such as "wherein the method reduces storage requirements by at least X%" or "wherein catastrophic forgetting is reduced by at least Y%." These may seem redundant when the independent claim already implies the benefit, but they provide concrete positions to fall back to if the independent claim faces eligibility challenges.

Business Implications

For businesses developing AI technology, Desjardins makes patent protection more accessible in ways that have practical commercial implications. Venture capital investors favor patent-protected technology because it creates defensible market positions. Patents establish barriers to competition that matter in crowded markets. They enable monetization through licensing or sale. They add tangible value to company valuations in ways that unprotected trade secrets do not. If your technology solves real technical problems and improves system performance, you now have a clearer path to patent eligibility, and that clarity has value.

Timing matters in ways that surprise many inventors. Patent applications can be filed while technology is still in development—in fact, filing earlier often provides strategic advantages. You establish a priority date, which matters significantly in a first-to-file system. You gain the ability to disclose and publish your work without losing patent rights. You enable continuation practice that allows you to capture improvements and variations as development continues. Waiting for technology to be "finished" before considering patent protection often means losing priority to competitors or being forced to keep valuable innovations secret when publication would serve business interests.

That said, patents aren't appropriate for every AI innovation. They're expensive, requiring both upfront costs for preparation and filing, and ongoing maintenance fees throughout the patent's life. The strategic question is whether patent protection makes sense for your particular technology and business model. If the technology can't be reverse-engineered and will remain an effective trade secret, patents may not be necessary. If patent protection doesn't provide meaningful competitive advantage in your market, the cost may not be justified. If you're unlikely to enforce patents against infringers, they provide little value. And for some business models, first-mover advantage or network effects provide better protection than patents ever could.

Working With Patent Counsel

Working with patent counsel who understand both the technology and the legal landscape makes a substantial difference in outcomes. Your patent attorney needs to grasp not just what your algorithm does, but how it works and why it matters—this enables them to identify the improvements that support eligibility arguments. They need to know the body of case law around patent eligibility, understanding how Desjardins fits within the larger framework established by Enfish, McRO, Alice, and the other precedential decisions that shape this area of law. They need both technical writing skill and legal strategy capability to draft specifications and claims that work together to establish eligibility while providing commercially valuable protection. And when you receive an eligibility rejection—and many AI patent applications still will—experienced counsel can distinguish your invention from unfavorable precedent and analogize to favorable decisions in ways that move prosecution forward rather than ending in abandonment.

The Desjardins decision doesn't guarantee that every AI invention will be patentable, and it shouldn't. Mathematical concepts applied in conventional ways should face rejection—patent law shouldn't reward incremental applications of known techniques just because they happen to involve machine learning. But for businesses developing AI technology that solves genuine technical problems, that reduces computational requirements, that improves system performance, that enables capabilities previously not possible, patent protection is achievable. The framework is clearer now than it has been in years.

The key is approaching patent strategy with appropriate sophistication. Audit your technology for patent-eligible improvements by asking what technical problems you're solving and what measurable benefits your approach provides. Document your innovation as you develop it because technical notes, benchmark data, and performance comparisons become valuable evidence of improvement when it's time to file. And consult with a patent attorney who understands both AI technology and the evolving legal framework for software patent eligibility before you've built so much momentum in one direction that changing course becomes impractical.

Patent protection for AI inventions is more accessible after Desjardins, but accessibility doesn't mean simplicity. The path is clearer, but it still requires careful strategy and skilled execution. For businesses building AI technology that solves real problems, that clarity is worth the effort.

About Tinch Law

Tinch Law provides intellectual property counsel to entrepreneurs, startups, and small businesses. Helping innovators develop patent strategies that align with business goals and protect competitive advantages.

Located in Maryland, working with clients throughout the Mid-Atlantic region and nationally on patent, trademark, and IP strategy matters.

---

This article is for informational purposes only and does not constitute legal advice. Patent law is complex and fact-specific. Consult with a licensed patent attorney regarding your specific situation.

‍