USA Today Co. sued OpenAI on Thursday in the Southern District of New York. The complaint is 79 pages, it names fourteen plaintiffs, all owned by the same company, and it asks for damages above $250 million. The filing has a headline number and a lot of pleaded detail, so the useful question is which parts the court would have to decide. I read the complaint itself, which is on the court docket, and the rest of this essay is about what it pleads.
What it claims
The complaint has three counts. Count I is direct copyright infringement, pleaded on five bases: scraping websites and copying third-party datasets, an exchange of copies with Microsoft, building training datasets, storing and reproducing the models themselves, and disseminating outputs through ChatGPT (¶¶ 187–197). Count II is vicarious infringement against the parent and affiliate entities (¶¶ 198–200). Count III is removal of copyright management information under 17 U.S.C. § 1202 (¶¶ 201–209).
That is a narrower set than the complaint the New York Times filed in December 2023. The Times pleaded seven counts, including two of contributory infringement, unfair competition, and trademark dilution. USA Today drops all of those. It also does not plead fair use as a count. The complaint quotes an internal OpenAI statement about “fair use” (¶ 9), but it leaves the defense for OpenAI to raise.
What the outputs show
The complaint’s most concrete material is in ¶¶ 123–141. These paragraphs describe ChatGPT output for articles from the papers, with the prompt quoted in each case: “Please find and summarize the article with this title and give me an in-depth summary.” The complaint says one of the outputs was “largely paraphrasing the original article” (¶ 126), and another was an “extensive multi-section summary” that follows the structure of the original (¶ 123).
These are summaries of specific articles, retrieved and rewritten, and the complaint itself describes most of them as paraphrase (¶¶ 123, 126). That distinction is the one a court will spend time on. The complaint pleads memorization in general terms (¶¶ 8–11, 114–116), and it says some outputs reproduced “substantive content” (¶ 125). The worked examples are the part it can show.
The complaint also quotes material from an earlier court filing that it says shows market substitution. One line is an OpenAI software engineer saying “users won’t click” (¶ 15). The court will have to read the source material, and I have not.
The number that is an estimate
The complaint says OpenAI’s models “have copied hundreds of thousands of articles and other materials from the USA TODAY Publications” (¶ 13). The complaint does not itemize them. It pleads that figure “upon information and belief.” The 19 publications named in ¶ 2 are the ones the complaint lists by title, and the registered copyrights are in an exhibit I did not open. The gap between what is named and what is estimated is the part of the case that discovery will have to close.
Paragraph 13 pleads damages “in excess of $250 million,” alongside per-violation maximums of up to $150,000 for each willful infringement and up to $25,000 for each § 1202 violation. The complaint does not say how it reached the total. Statutory damages are set per registered work, so the number that matters is how many registered works the court counts. The complaint does not give that count.
What the other cases say
The complaint says many related suits have been consolidated in the same court (¶ 1) and cites the consolidated OpenAI docket for its exhibit material. I could not find a fair-use ruling in those cases. On the other side of the industry, the Bartz v. Anthropic settlement received final approval on 20 July 2026 for $1.5 billion. The Authors Guild quotes the order as saying the “past-only release does not apply to any output claims.” I have not read the order itself, so that is the Guild’s quotation, not mine.
An earlier essay on the Times case covers the same publisher-suit thread. In my reading, the outputs are the material the complaint shows rather than asserts, and the narrower count list makes that visible. Judge Mehta’s ruling on the content-for-traffic theory concerns a different defendant and a different claim, but it turns on the same question of what a publisher can prove about lost traffic.
What the complaint has to prove
The complaint asks for a permanent injunction and destruction of “all GPT or other LLM models and training sets that incorporate” its content, under § 503(b) (prayer, item 4). That remedy is the one most likely to draw attention, and it rests on the same training-data theory as the damages claim. The outputs are the part the complaint documents with examples. The copying into training is the part it has to establish with discovery.
Two things would change the reading. The first is the exhibit list: if the registered works are a small share of the 19 publications, the damages ceiling falls well below the headline. The second is the fair-use question in the consolidated cases, which the complaint does not plead. Until a court rules on that, the complaint is a set of claims with the evidence it documents in the summaries and its biggest number in an estimate.
References
- USA Today Co., Inc. v. OpenAI Foundation, No. 1:26-cv-08892 (S.D.N.Y.), Complaint, Doc. 1, filed 8 October 2026. PDF on CourtListener. Paragraph and prayer citations are to this document.
- Roth, E. (2026). USA Today becomes the latest publisher to sue OpenAI. The Verge, 8 October 2026. Secondary; used for the filing date and the $250 million headline.
- New York Times Co. v. Microsoft Corp. and OpenAI, Inc., No. 1:23-cv-11195 (S.D.N.Y.), Complaint, Doc. 1, filed 27 December 2023. PDF. Source for the claim comparison.
- Authors Guild (2026). Court grants final approval of Anthropic copyright settlement. Dated July 2026. Secondary; the order itself was not read.