Artificial intelligence can write computer code, generate realistic images, analyze medical records, and even help you plan dinner.
But apparently, it also has a memory problem. In fact, one of the more interesting recent patent decisions involves an AI system that has trouble remembering what it previously learned.
More on that shortly.
But can you patent AI? First, there is some good news for software developers, technology startups, and anyone working on artificial intelligence inventions. The United States Patent and Trademark Office (USPTO) appears to be taking a more inviting approach to patents involving AI and software. During a September 29, 2026 Senate hearing, USPTO Director John A. Squires encouraged inventors to submit their AI inventions for patent protection rather than assume they would be rejected. After agreeing that AI innovation is an area of cricital importance he issued a simple invitation. “Bring them to the US patent office.”
That may not sound particularly revolutionary. But for those patent attorneys who have spent years navigating the uncertain world of software patents, it is a noteworthy change at the Patent Office.
When Is a Computer Invention More Than Just an Idea?
For years, software patents have faced a rather unusual problem. An inventor might develop software that performs a complicated task, improves efficiency, or saves a company millions of dollars, only to be told that the invention is not eligible for patent protection because it is directed to an “abstract idea.”
The problem traces back in part to the Supreme Court’s 2014 decision in Alice Corp. v. CLS Bank. Under Alice, simply taking an existing idea and performing it on a computer generally is not enough to make it patent eligible. Consider the following example.
Suppose you develop software that uses AI to predict which products customers are most likely to purchase. That might be a valuable business tool. But using AI to make better sales predictions does not necessarily qualify as a patentable technological invention. Now imagine you develop a new technique that allows the AI system to perform those predictions using half the computing resources, without sacrificing accuracy. Based on the recent testimony and patent rulings, that may make a difference. You may have improved the underlying technology rather than simply used existing technology to accomplish a business objective. And that brings us back to an AI system with a memory problem.
Apparently, AI Can Be Forgetful
Researchers working with artificial intelligence have identified a phenomenon called “catastrophic forgetting.” The basic problem is based on the fact that when an AI model learns how to perform one task and then learns another task, it may forget or lose some of what it previously learned.
Imagine spending months teaching someone to play the piano, only to discover that after learning the guitar, they can no longer play the piano. That is roughly the same issue AI is having.
Google DeepMind developed a machine learning technique designed to address catastrophic forgetting. The technology became the subject of a patent application considered by the USPTO in Ex parte Desjardins. The USPTO’s Appeals Review Panel concluded that the claimed invention was directed to a technological improvement and therefore qualified as patent-eligible subject matter. The decision was subsequently designated precedential, meaning USPTO patent examiners and the Patent Trial and Appeal Board must follow its ruling when considering patent eligibility.
One of the key findings is that: The invention was not simply using AI to do something useful. It involved an improvement to the AI technology. That distinction will be helpful for other patent applicants seeking to patent inventions involving machine learning, computer systems, and software.
Of course, eligibility is only one hurdle. As the claims in Desjardins still faced an obviousness rejection.
The Patent Office Is Changing Its Approach. The Courts? Not So Fast.
Following the Desjardins decision, the USPTO has provided additional guidance for patent examiners and a new procedure allowing applicants to submit evidence explaining why their inventions represent technological improvements.
The procedure, called a Subject Matter Eligibility Declaration (or SMEDs), are voluntary declarations that allows inventors to provide additional evidence about how their invention improves the technology. For example, an applicant might provide evidence showing that an invention makes a computer faster by improving the computer processing speed, reduces memory requirements, or overcomes a recognized technical limitation. Think of it as another opportunity to explain why your invention is more than just an abstract idea.
That is encouraging. But that is not the end. As Director Squires stated, elligibility is different from patentability.
But even if the USPTO finds something patentable, they do not get the final word on whether a patent is valid. Federal courts still have the right to review patent validity and when they do, they apply the Supreme Court’s Alice framework in its review. As we discussed in our earlier article, Software Patent Eligibility in 2026: Ice Cream, Online Shopping, and AI, courts continue to scrutinize software patents, particularly those directed primarily to business methods or abstract processes.
In other words, the Patent Office may be opening the door wider, but there may still be legal obstacles waiting on the other side.
So, Can You Patent an AI Invention?
Perhaps. It certainly may deserve a closer look. If you are developing AI software, consider what makes your invention different. Are you simply using an existing AI model to perform a task that people already perform? Or have you developed something that improves how computers, software, or AI systems actually function?
For example, does your technology:
- Train an AI model more efficiently or with less data?
- Improve processing speed or reduce computing requirements?
- Solve a technical problem that existing software cannot adequately address?
- Introduce a new architecture or method for operating a computer system?
- Improve the performance or operation of another technology through a specific technical solution?
None of these automatically establish patentability. But they are the type of improvements the patent office may consider when reviewing your patent application. The other important consideration is how the invention is described in the patent application. A patent application that simply says “use artificial intelligence to accomplish X” may face substantial difficulties. On the other hand, an application that explains how a new technology works, why it matters, and what makes it different has a better chance of success.
Maybe It Is Time to Give Your AI Invention a Second Look
For years, many software developers have assumed their inventions were not worth patenting because of the uncertainty surrounding software patent eligibility. The USPTO’s recent rulings suggest those assumptions may be incorrect. While not every AI invention is going to be patentable, an inventor should not necessarily dismiss the idea automatically. If you have a computer technology that improves how computers, software or AI systems actually perform, then it may warrant a closer investigation. After all, if a computer can be taught how to remember what it has forgotten, perhaps we should remember that technological innovation does not always come with gears, motors, or moving parts.
Sometimes the invention is in the code.
At the IP Center, we work with inventors, startups, software developers, and established companies in Kansas City and throughout the U.S. to evaluate and protect innovations involving artificial intelligence, software, electronics, and other technologies. If you have developed something new, contact The IP Center to discuss whether patent protection makes sense for your invention.
This article is for general informational purposes and does not constitute legal advice.