USA Today OpenAI Lawsuit Seeks $250M Over ChatGPT Training Data

A newspaper chain comes to Manhattan with numbers

USA Today Co., the newspaper company that used to be called Gannett, has sued OpenAI in Manhattan federal court. It wants more than $250 million in damages and an injunction over claims that OpenAI copied hundreds of thousands of its articles to build the models behind ChatGPT.1011 The case was filed on October 8, 2026, in the U.S. District Court for the Southern District of New York. It is captioned USA Today Co. v. OpenAI Foundation, docket No. 1:26-cv-08892, and Steven Lieberman of Rothwell, Figg, Ernst & Manbeck represents the publishers.108 As of the first reports, OpenAI had not commented.113

The complaint's central accusation is blunt. It says OpenAI's commercial success depends on copyright infringement at scale, and that the company never asked permission or offered payment before using the papers' work to build products worth hundreds of billions of dollars.911

Coverage agrees on the main facts. It differs on how to count the plaintiffs. Several outlets describe 19 publications.38 Others count 13 or 14 corporate entities that hold the copyrights.121415 One report says 18 affiliated newspapers.7 The most likely explanation is that 14 legal plaintiffs, USA Today Co. plus 13 subsidiaries, own the rights to 19 newspaper brands.1415 Those brands include The Tennessean, the Detroit Free Press, The Arizona Republic, the Milwaukee Journal Sentinel, The Des Moines Register and Louisville's Courier Journal.3813 The defendants are seven OpenAI entities, and that list follows the company's 2025 recapitalization into a foundation and a public benefit corporation.1415 Microsoft is not named.15

What sets this complaint apart

Most AI copyright complaints say a model was trained on huge amounts of text and stop there. This one counts. It says more than 160,000 entries from the plaintiffs' domains appear in WebText, the dataset OpenAI built for GPT-2, and that 83,266 of them came from usatoday.com alone.17 It also cites C4 as a training source.1 One detailed analysis puts the publishers' C4 count at more than 122 million tokens. The same analysis notes that the itemized domains add up to less than that, so the rest presumably comes from sites the complaint does not list.15

The evidence also reaches today's products. The complaint includes examples of GPT-5.6 output.114 By one account, the publishers gave the model the headline of a real article from each of the 19 titles and asked for an in-depth summary. Each time, they say, it found the piece and produced a summary that followed the article's structure and kept its substance.15 That differs from the verbatim passages shown in earlier cases such as The New York Times' suit. The theory here is substitution: a detailed paraphrase can take a reader's visit away from the publisher even when the wording changes.1512

To support that theory, the publishers quote OpenAI's own staff, using material from discovery in other litigation. The complaint cites Nick Turley, head of ChatGPT, as writing that publishers face an "existential threat" and that OpenAI's products are "largely substitutive."15 It also cites an OpenAI engineer who doubted users would click source links no matter how prominently they were shown.13 It quotes a Microsoft executive, from unredacted New York Times case documents, describing the unauthorized use of content as an "astonishing theft."3

The legal theories and how the $250 million adds up

There are three counts: direct copyright infringement, vicarious infringement, and removal of copyright management information under the DMCA. That last claim covers stripping bylines, titles and copyright notices when articles were scraped.715 Statutory damages for willful infringement can reach $150,000 per work, and the complaint also cites up to $25,000 for each removal of copyright information.715

One analysis worked backward from the headline figure. At the statutory minimum of $750 per work, "hundreds of thousands" of articles comes to roughly $250 million.15 It also flags a catch. If a paper registered whole editions as compilations, each edition may count as a single work, which could sharply cut the number of awards.15 That counting question will likely shape the real exposure far more than the demand in the complaint.

The remedy that matters most to the industry is not the money. The publishers want a jury trial, an order to destroy every model and training set that contains their content, and a permanent bar on future use.315 The New York Times and The Seattle Times have asked for the same thing.15 Courts rarely order a frontier model destroyed. The request is best read as leverage: it turns a damages fight into a threat to the product itself.

The pile of cases

This is not a new kind of lawsuit, and that is part of why it matters. OpenAI already faces suits from The New York Times, The Intercept, Ziff Davis, CBC/Radio-Canada, Encyclopaedia Britannica and Merriam-Webster, and The Seattle Times.16 In June 2026, more than 30 U.S. local newspaper owners jointly sued OpenAI and Microsoft.9 Across the industry, copyright owners have filed dozens of suits against OpenAI, Anthropic and Meta.10

The flagship case is the Times suit, consolidated before Judge Sidney Stein. In March 2025, Stein denied most of OpenAI's motions to dismiss but narrowed some DMCA claims. The USA Today plaintiffs asked the court to relate their case to the broader OpenAI multidistrict litigation.14 By one account, summary judgment motions are already pending there.15 Arriving late has an advantage: the new plaintiffs can use discovery that is already done and still raise fresh claims about GPT-5.6 and ChatGPT search.15

The policy backdrop has also shifted. In September 2026, the Justice Department filed a brief supporting OpenAI in the Times case. Reuters described it as the government's first stated position on AI training copyright litigation. One newsletter noted that the USA Today suit came a day after a court approved Anthropic's $1.5 billion copyright settlement, which it called a likely price anchor for publisher negotiations. The two signals pull in opposite directions. Washington is leaning toward fair use, while settlement values show what defendants will pay to avoid finding out.

Lawsuit as negotiating tactic

The strongest context comes from licensing. USA Today Co. is not against AI. It licensed its content to Perplexity in 2025 and signed a multi-year AI deal with Meta in December 2025.1516 The complaint argues the copying was willful, pointing to the paid deals OpenAI has signed with other publishers.16 OpenAI was reported in 2024 to be offering news organizations $1 million to $5 million a year, while News Corp's deal reportedly topped $250 million over five years.9

That comparison is the most telling detail in the coverage. The damages demand roughly matches the reported value of OpenAI's largest news licence, which suggests the publisher is pricing its archive at the top of the market. Several outlets read the filing the same way, as a move that sets a price and starts discovery while leaving room for a deal.65 Commentators disagree on emphasis. Some treat the suit as a test of fair use in principle.25 Others treat it as a business negotiation carried out in court.6 The second reading fits the evidence better: a company that licenses to rivals is suing the one company that has not paid it.

What it means for AI liability and enterprise buyers

For businesses using generative AI, including legal tech vendors building on OpenAI's models, the immediate impact is small. The suit does not target users, and no order changes how ChatGPT works.614 The risk arrives indirectly, through higher licensing costs, changes in how models handle news content, or limits imposed by a settlement or ruling.6 One analysis advises buyers to ask vendors what their models were trained on and whether retrieval respects site terms. It also suggests reading indemnity terms such as OpenAI's Copyright Shield, launched in 2023, carefully, since such protection usually depends on customers using built-in filters.15

The retrieval angle matters more than it first appears. A fair use defense for training a model once is a different question from repeatedly fetching and summarizing a live, possibly paywalled article on request. The publishers say ChatGPT's summaries cut their traffic and ad revenue, and that some content came from behind paywalls.13 Britannica's suit also targets this kind of real-time retrieval, known as RAG.9 OpenAI is expected to argue that training is transformative, that GPT-2-era data says little about current models, that summaries convey unprotected facts, and that some claims are too old.15 The complaint answers the timing defense by calling the infringement ongoing, since the models need a constant supply of new material.15

The verdict for now

Nothing has been decided. A complaint is a list of allegations, and $250 million is a demand, not an award.814 Still, this filing shows where publisher litigation is heading. It brings dataset counts, internal documents and live product tests where earlier suits relied on general claims of mass scraping. Even if courts eventually rule that training is fair use, the summary-and-retrieval features at the center of this case are more vulnerable. The likeliest result is a licence, and this complaint is how USA Today Co. is setting its price.