Copyright Law vs. OpenAI

· LawDiver

USA TODAY and the newspapers it owns sued OpenAI for training ChatGPT on their reporting. The complaint, and where it meets the Third Circuit's ROSS decision.

On October 8, 2026, USA TODAY Co. and the newspapers it owns sued OpenAI in the Southern District of New York, the court where the other OpenAI copyright cases are already consolidated. The complaint treats those papers as one library: USA TODAY and its local titles, registered copyrights attached, copyright notices on the pages. The theory is the same across the group. OpenAI copied the reporting to train and run ChatGPT, stripped the notices on the way in, and now answers the questions those papers exist to answer. The plaintiffs ask for more than $250 million, an injunction, and destruction of the models and training sets that contain their work.

This is a pleading. OpenAI has not answered it.

What the complaint says OpenAI did

The complaint says the models were trained on copyrighted text scraped from the internet without a license, including text behind paywalls. It quotes Sam Altman telling the British House of Lords that today's leading models cannot be trained on public-domain material alone. It quotes a Microsoft executive, from the related multidistrict record, calling the copying an astonishing theft.

The datasets are named. WebText, OpenAI's scrape of links posted on Reddit, allegedly holds more than 160,000 entries from these papers. A filtered English snapshot of Common Crawl, C4, allegedly holds more than 122 million tokens of their text. WebText2, built the same way, was weighted at 22 percent of the GPT-3 training mix while making up less than 4 percent of the tokens. Internal comments, the complaint says, treated news as the high-value slice, and OpenAI's researchers described memorization as the objective of training. By 2022, employees expected GPT-4 to have memorized a ton of data and to be good at regurgitation.

The filters that followed, on this account, hid the copying. They did not suppress output of a publisher that had not yet sued. A Microsoft executive called that an accidental cover-up, because the owner of the work could no longer see that it had been used. The complaint also says copies moved between OpenAI and Microsoft: Bing index content to OpenAI, a joint crawl called Project Mango, and OpenAI's GPT-3 training set delivered to Microsoft.

The injury pleaded is substitution. OpenAI's Head of ChatGPT, Nick Turley, is quoted writing that the products are largely substitutive, and that they will get more substitutive as they improve. An engineer is quoted saying users will not click the links. Internal documents called ChatGPT the modern newsstand. OpenAI has said demand for local news inside ChatGPT already runs at about one million prompts a week. The papers say that is their reporting, answered without a license.

Three claims

Count I is direct infringement under 17 U.S.C. § 501. The complaint places the copies in the scraped datasets, in the training runs, in model weights that memorized the articles, in ChatGPT outputs that repeat them, and in the copies exchanged with Microsoft.

Count II is vicarious infringement. It says the parent entities in the OpenAI group directed the copying and kept the profit.

Count III is a Digital Millennium Copyright Act claim under 17 U.S.C. § 1202. Each article, the papers say, carries copyright management information: the byline, the title, the publisher, and the copyright notice. OpenAI used extractors called Dragnet, Newspaper, and Gutentag that separate the article from copyright notices. The complaint says OpenAI knew those tools drop the notice, and that ChatGPT then distributes the text without it. OpenAI's terms tell the user that the user owns the output. The papers say that combination conceals the infringement.

Statutory damages for willful infringement run up to $150,000 per work. Removing copyright management information runs up to $25,000 per violation. The complaint says the models copied hundreds of thousands of articles. The prayer also asks the court to enjoin the use and, under 17 U.S.C. § 503(b), to order destruction of every model and training set that incorporates the papers' content.

Where this meets ROSS

On September 29, 2026, the Third Circuit affirmed partial summary judgment in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc.. ROSS had used 2,243 West headnotes as the questions in training memos for a legal-research tool. The court held the headnotes original. Editors chose which points of law mattered and how to write each note so it could stand alone. Copying them was not fair use. The use was commercial. It was minimally transformative, because ROSS and Westlaw both use headnotes to help a researcher find the law. ROSS copied more than it needed: the judicial opinions were free, and the headnotes were the easy path. The use harmed Westlaw as a research product and a developing market for licensing headnotes as AI training data.

The opinion says, in terms, that ROSS was not a generative model. It could not write new expression. It returned passages from opinions. In a footnote the court set aside a Department of Justice statement filed on September 1, 2026, in the OpenAI multidistrict litigation in the same New York court. The Department, relying on Bartz v. Anthropic, had argued that training a model which can generate original responses is transformative, and that the OpenAI training did not produce substitutive competition. The Third Circuit said those points do not describe ROSS, which trained a commercial substitute for Westlaw.

That footnote is the overlap. The newspapers filed in the court the footnote is talking about, and they plead the facts the footnote treated as the difference. They say ChatGPT does substitute for the news. They say the models memorize articles and can repeat them, which is a harder story than a model that only writes something new. They say a licensing market for their text already exists, the same derivative market the Third Circuit credited for headnotes. And they add a claim ROSS never had: the training pipeline threw away the byline and the copyright notice.

What ROSS does not decide is this complaint. ROSS lost on headnotes copied to build a retrieval competitor. The Third Circuit left the generative case for the docket this complaint just joined. The pleading is written to meet that case with substitution, regurgitation, and stripped copyright notices.

The complaint is here. The Third Circuit opinion is here.

Copyright Law vs. OpenAI | LawDiver