University of Tennessee sues Anthropic over neural network patents

The University of Tennessee Research Foundation sued Anthropic on Monday in the US District Court for the District of Delaware, accusing the artificial intelligence company of building its models on patented neural network methods it never licensed.

The complaint, unsealed on Tuesday, is thought to be the first patent infringement case brought against Anthropic, and it arrives days after a California judge approved the company’s $1.5bn book piracy settlement with a class of authors.

The foundation, which manages patents for the Knoxville campus, cast the case as part of a wider habit. Anthropic’s “cavalier approach to others’ intellectual property rights in the development of its products extends beyond the use of copyrighted material,” it said in the filing.

The framing deliberately ties the patent claim to the copyright dispute that produced that settlement, and to a growing docket of complaints against the company, among them a lawsuit over usage limits on its Claude Max plans.

At issue are two patents covering what the foundation describes as contributions to machine learning, neuromorphic computing, and neuroscience-inspired computing, all invented by University of Tennessee professors.

Bloomberg Law reported that the asserted patents are numbered 10,019,470 and 10,095,718, and that the complaint singles out Claude Code, Anthropic’s agentic coding tool, along with its underlying software architecture.

The technology traces to TENNLab, a University of Tennessee group that has worked on brain-inspired computing since 2014. The lab, led by professors Garrett Rose, James Plank, Catherine Schuman, and Ahmedullah Aziz, holds seven issued patents and several pending applications, with work spanning spiking neural networks and custom neuromorphic hardware.

Its early frameworks, known as NIDA and DANNA, underpin much of that portfolio. Schuman has spoken publicly about how collaborative the field has become, once remarking that it is “basically impossible to do research in this field in isolation.”

The copyright case sets the backdrop. A federal judge in San Francisco had initially held back on the $1.5bn deal over concerns about how authors would be paid, before signing off on it in a decision described as the largest of its kind in US copyright history.

The patent suit stakes out different ground, shifting the argument from the training data to the design of the models.

The foundation is seeking unspecified monetary damages and an injunction that would bar Anthropic from further infringement.

It has not put a figure on what it believes the technology is worth, and the complaint does not disclose whether the two sides discussed a licence before the suit was filed.

According to Bloomberg Law’s reading of the complaint, the foundation alleges that Anthropic’s products implement patented methods for constructing neuromorphic networks, including a background execution scheduling system and a memory consolidation engine tied to one of the two patents.

That level of detail could not be independently confirmed from the public reporting, and the precise mapping of patent claims to Anthropic’s code will be contested as the case proceeds.

Spokespeople for Anthropic and the foundation did not immediately respond to requests for comment. Anthropic has not filed a formal response, and the company has previously said little in public about the specifics of pending litigation.

What happens next is procedural and slow. Anthropic will have to answer the complaint in Delaware, a court that hears a large share of the country’s patent disputes, and the case is likely to turn on claim construction, the painstaking process by which a judge decides what the patents actually cover. Only after that will the question of whether Claude infringes them come into focus.

For a company racing to build out its Claude partner network and spending the past year settling one intellectual property fight, the suit is a reminder that the books used to train a model are only one kind of property a rival can claim. The patents here describe how a machine learns, not what it read.