USA TODAY Co. asks a federal court to make OpenAI pay for training on 19 publications

The complaint asks for two remedies, and they reach different places ^[4]. Damages would move money from OpenAI to the publishers. The second remedy is an order blocking the alleged infringement ^[4]. Reporting on the complaint does not say whether that order would cover a model already trained or only future copying ^[4].

Divided across the 19 named publications, the demand comes to more than $13 million a title ^[18]. Treat that figure like someone else's benchmark table. The first page of a complaint is written by the side asking to be paid, and the $250 million is the plaintiffs' demand, not a court finding or an award ^[8].

OpenAI's public position has three parts. In a January 2024 statement it said it viewed training on publicly available internet material as fair use ^[16]. The same statement described publisher opt-outs and licensing partnerships as parts of how it works with news organizations ^[17]. The licensing part has a large example. OpenAI's 2024 deal with News Corp covers both current and archived material from News Corp titles, among them the New York Post and The Wall Street Journal ^[12].

That deal shows a licensed route for news archives exists and that OpenAI uses it ^[12]. It does not establish whether the material in this case was covered by any permission. It also does not settle whether training without a license is lawful ^[13].

The case joins dozens Reuters has described against OpenAI, Anthropic and Meta ^[9]. In September, Reuters reported arguments by authors and news outlets in a separate Manhattan case that statements by OpenAI and Microsoft executives could undermine the companies' copyright defenses ^[10]. This filing puts local papers next to a national title ^[5]. The 19 titles come from a company that says it owns more than 200 local media organizations ^[15].

For this dispute to become a supply problem for teams building on these models, a court would first have to reject fair use for this material. An injunction would then have to reach trained models or their outputs. The record so far is at the first step. The complaint was filed on October 8th ^[1], the court has not ruled on its claims ^[11], and Reuters reported that OpenAI did not immediately comment ^[7].