USPTO And Federal Circuit Decisions Reveal Patent Eligibility Parameters For AI Use

The use of artificial intelligence (AI), machine learning, and predictive analytics is growing rapidly across industries. From health care and financing to marketing and logistics, many companies are implementing these solutions into their business processes and products. In this way, companies are creating unique applications for AI that improve their business and products.

When implementing AI, companies need to keep in mind a variety of considerations: cost, employee training, licensing, manufacturing contracts, cybersecurity, etc. A few considerations that can easily be overlooked are the creation and protection of intellectual property rights. The use of AI or predictive analytics may just be a business practice, but it could be copyrightable, a trade secret (or use trade secrets), or patentable. Keeping the underlying model or method confidential may preserve trade secret protection, but if the use of AI is patentable, then, alternatively, a company may be able to use a patent to prevent its competitors from using AI in the same way or may obtain additional monetary value from its use of AI, such as through licensing the patented process or product.

However, the application of AI to a product or process does not automatically create patentable technology. Some recent cases from the U.S. Patent and Trademark Office (USPTO) and U.S. Court of Appeals for the Federal Circuit shed some light on how to determine patentability for AI, machine learning, and predictive analytics software.

Simply Using Software for Something Known Is Not Patentable

In the case of In re Brian McFadden (2026), the Federal Circuit considered whether a patent that claimed methods and devices used “for operating, regulating, and controlling” an information exchange, was patentable under 35 U.S.C. §101. In general, §101 states that a patent may be granted to “[w]hoever invents or discovers” new technology. As such, it prevents patenting abstract ideas, laws of nature, and natural phenomena (including products of nature). In this way, U.S. patent law distinguishes between the idea itself and applications of those ideas.

In McFadden, the USPTO rejected the claimed subject matter as an abstract idea. McFadden’s patent described an information exchange from producers to consumers (e.g., marketing activity) accomplished by generating information distributions, comparing them, calculating differences, and determining an exchange value. The patent further explained that the process could include a variety of computed coded software and hardware systems. As such, the USPTO denied McFadden’s patent claims because they found the patent was only “software running on generic computer elements” to accomplish an abstract idea.

McFadden appealed this decision, but the Federal Circuit agreed with the USPTO. The Federal Circuit found that the patent only claimed mathematical operations and determinations implemented via software on a standard computer. Although, a patent can still be granted if something more transforms the abstract idea, such as an innovation to the hardware that performed the abstract idea, or a change in the overall process itself. Here, the court concluded that nothing more was added to the way McFadden organized and compared data, stating that “improving abstract and generic calculations on a standard computer, without more, is still abstract.” Thus, the Federal Circuit affirmed the decision of the USPTO.

“[I]mproving abstract and generic calculations on a standard computer, without more, is still abstract.”

Using AI Can Be Patentable with “Something More”

The key takeaway from In re Brian McFadden is that when an idea or known process is implemented using software, then U.S. patent law may not view it as patentable, without something more. For example, taking a known workflow or business practice and automating it using AI may not be patentable on its own. However, using AI can still be patented if the AI, itself, is functioning in an improved way, or the hardware or product on which the AI is implemented has been improved.

Notably, the USPTO has granted patents to improvements accomplished via software or improved processing power. Where AI, machine learning, or software is used to create new benefits (like reduced storage, reduced system complexity, and avoiding prior problems not previously addressed), then the use of AI or software may be patentable.

In late 2025, the USPTO highlighted the importance of Ex Parte Desjardins (2024) and designated the opinion precedential. In Desjardins, a method of training a machine learning model on a series of tasks was found patentable. Although the USPTO initially found that the patent was directed towards the abstract idea of processing data and mathematical calculations (similar to McFadden), the patent improved how the machine learning model operated, including addressing the problem of “catastrophic forgetting,” which was not previously addressed by other machine learning models. Thus, the patent accomplished the abstract idea not only by training the machine learning model on known tasks but also by improving the operation of the model while it learned new tasks. This improvement was the “something more” needed to have patentable subject matter.

Accordingly, there is a difference between using AI or software for a specific purpose (like in McFadden) and improving the AI, software, or hardware on which it is used (like in Desjardins).

Key Considerations

As businesses implement AI, machine learning, and predictive analytics software into their existing workflow or products, the following considerations should be kept in mind to assist the business in determining if any intellectual property rights are implicated:

• Has AI changed how work gets performed or is it simply automating an existing process?
• Was the underlying process or product improved, separate from the use of AI?
• Did AI solve a technical limitation that previously existed?
• Would the invention still be valuable if the AI model was replaced with a different model?
• Did the innovation improve how the AI model itself operates?