Technology / Inventions & IP

OpenAI and Anthropic face five attention-patent claims—but a 2015 priority date proves little

Parallel Delaware complaints say ChatGPT and Claude use patented iterative-attention methods first filed in 2015 and 2016. The defendants have not answered, and no court has tested the claims, their validity or the alleged infringement.

INNOVOX News DeskNov 14, 2026 · 6 min read
Representative drawing from U.S. Patent 12,511,580 showing users connected to an adaptive system with content, structural and usage networks
USPTO · U.S. Patent 12,511,580 drawing · typically unrestricted under USPTO terms

The story

A newly formed patent fight is putting one of generative AI’s central technical ideas before a federal court. Flinn Artificial Intelligence Technologies filed related complaints against Anthropic and OpenAI in Delaware on November 12, alleging that Claude and ChatGPT products infringe five U.S. patents titled “Iterative attention-based neural network training and processing.” The defendants had not publicly answered the allegations when the cases were reported, and no judge has decided that any product infringes a valid claim.

The asserted patents are U.S. 12,327,166; 12,223,404; 12,511,580; 12,462,197; and 12,462,198. Court records list the same portfolio in both cases. The patents name Steven Dennis Flinn and Naomi Felina Moneypenny as inventors, and the complaints say their work reaches back to applications filed in August 2015 and January 2016—before the influential 2017 paper “Attention Is All You Need” introduced the transformer architecture that helped reshape modern AI.

That chronology is important but not decisive. A patent does not confer ownership of an entire research field or a word used to describe it. Infringement is tested claim by claim: the court first determines what the patent language means, then asks whether the accused system performs every required element of at least one asserted claim. Validity can also turn on whether the claimed subject matter was new and non-obvious, adequately described and enabled, and eligible for patent protection.

The prosecution history makes that distinction especially relevant. The USPTO’s Official Gazette shows that patent 12,511,580 issued on December 30, 2025 with 65 claims. It sits in a continuation chain that includes a January 2016 application and claims the benefit of a continuation-in-part filed in August 2015. Continuations can preserve earlier disclosure while pursuing different claim sets over time, but each claim must be entitled to the priority date it invokes. The dates therefore open the inquiry; they do not end it.

Claim 1 of the ’580 patent illustrates the portfolio’s focus. It describes accessing syntactical elements, generating a vector representation with a multiple-attention neural network, searching stored vector representations for a similar vector, identifying associated content and generating a response for the user. Other claims add iterative training, probability generation and repeated attention or processing stages. That language could matter to retrieval, memory and agent-like workflows, yet a high-level resemblance is not enough—the accused implementation must match the legally construed limitations.

Flinn’s complaints cast a wider technical theory. They allege that Claude and ChatGPT systems use iterative attention architectures across model training and inference, and they identify products and functions including conversational responses, coding, reasoning, retrieval, memory, tools and agents. Those descriptions are the plaintiff’s allegations, based substantially on public product material. The cases have not reached discovery, where source code, architecture documents and expert analysis could test whether the asserted steps actually occur in the claimed combinations.

The requested remedy is primarily financial. The complaints seek a reasonable royalty, damages for alleged past infringement, enhanced damages if infringement is found willful, interest and fees, along with compensation for continued use. Flinn says the inventors developed the technology years before today’s commercial systems; OpenAI and Anthropic will have opportunities to dispute infringement, priority, validity, damages and any claim that they had legally relevant notice.

The suits are notable because patent litigation against frontier AI developers remains less common than copyright disputes over training data and generated output. A serious patent case can reach deeper into how a model and its surrounding product stack operate. It may also encourage other owners of pre-transformer neural-network portfolios to reassess whether later continuation claims can be read on current systems.

INNOVOX analysis: the strongest part of Flinn’s opening position is the documented filing chain; the weakest shortcut would be treating that history as proof that the patents cover transformers or attention generally. Modern AI products combine model architectures, retrieval systems, orchestration software and product interfaces. The litigation’s value as a precedent will depend on whether the asserted claims map cleanly onto those layers without stretching their language beyond the inventors’ original disclosure.

What to watch next is the precision of the parties’ technical positions. Early pleadings may challenge the complaints’ specificity. Later infringement contentions should identify the exact claims and accused functions. Patent-office review could test validity before trial, while claim construction will define terms such as “multiple-attention neural network” and “vector representation.” Until those steps occur, the cases establish a consequential allegation—not a verdict on who invented modern AI attention.

INNOVOX analysis

The consequential question is not whether the inventors used the word ‘attention’ before the 2017 transformer paper. It is whether the patents disclose and validly claim the particular multi-stage operations Flinn identifies—and whether current Claude and ChatGPT implementations perform every required step. That element-by-element inquiry will be far narrower than the complaints’ headline chronology.

What to watch

Watch for the defendants’ answers or early dismissal motions, detailed infringement contentions, challenges at the Patent Trial and Appeal Board, and the court’s eventual claim construction. Technical discovery about model architecture, retrieval and agent workflows could determine whether the cases remain broad portfolio disputes or narrow to a few product functions.