Publisher litigation against AI companies is really three separate fights that get reported as one. The first is whether training a model on copyrighted articles is infringement. The second is whether a lawfully trained model that retrieves and summarises your content at query time causes a different, compensable harm. The third is whether a platform's dominance gives publishers a competition claim. In October 2026 the first is unresolved for news, the second is only now being pleaded directly, and the third has just lost its first two cases in a US federal court.
What is the latest lawsuit, and what does it allege?
On 9 October 2026 USA Today Co sued OpenAI, seeking $250 million in damages, according to Press Gazette's tracker of publisher AI cases. The complaint alleges that OpenAI used the company's articles without permission or payment. It cites more than 160,000 entries from USA Today Co publications in OpenAI's WebText dataset, 83,266 of them from USA Today itself, and more than 122 million tokens from its publications in a 2019 Common Crawl snapshot. The coverage we reviewed did not state the court or the precise causes of action, so check the filing before relying on either.
Two features make it worth watching. First, it leans on dataset evidence that the company can trace back to specific sites, which is the same crawl pipeline we describe in our explainer on Common Crawl and CCBot. Second, press reports of the complaint say it also argues that ChatGPT's summaries substitute for visits to publisher sites, quoting an OpenAI executive who allegedly said such summaries could reduce traffic to other websites. That is a harm theory about the product in use, not only about how the model was built.
What have the courts actually decided so far?
Very little that settles the news question. The decisions that exist fall into four groups.
Books and training. In June 2025 Judge William Alsup held in Bartz v Anthropic that training on lawfully obtained books can qualify as fair use, but that Anthropic could not rely on fair use for building a central library from pirated copies. Anthropic agreed a settlement on 5 September 2025. A federal judge in San Francisco gave final approval on 20 July 2026: $1.5 billion plus interest, covering 482,460 books, which works out at roughly $3,000 per work before fees and expenses, according to BetaNews.
News and survival of the claims. The New York Times sued OpenAI and Microsoft in December 2023. In April 2025 Judge Sidney Stein largely denied the motion to dismiss, so the claims proceed. The case now sits inside a consolidated proceeding in the Southern District of New York, In re OpenAI Copyright Infringement Litigation. We found no published trial date. The Times also has a pending suit against Perplexity, and Dow Jones has its own pending suit against Perplexity.
Government position. According to the AI Vortex case tracker, the US Department of Justice filed a statement of interest in that proceeding on 1 September 2026 arguing that training on copyrighted text is not categorically infringing and can be fair use. That is advocacy, not a ruling, and we have only seen it described secondhand. Our earlier analysis of the DOJ filing covers what it could mean.
Competition law. Around 1 October 2026 Judge Amit Mehta granted Google's motions to dismiss amended antitrust complaints from Penske Media and Chegg over AI Overviews. He found no plausible agreement for the reciprocal dealing claim, writing that "an expectation is not an agreement". He also found the tying, monopolisation and standing arguments unpersuasive, and declined to rule on the state-law unjust enrichment claim. The Penske dismissal was without prejudice, so a new complaint or an appeal remains possible. The judge also said the court did not treat the alleged harms lightly.
So are publishers winning?
On the evidence, publishers are surviving rather than winning. Surviving a motion to dismiss means the claim is legally plausible, nothing more. The one nine-figure outcome so far came from authors, and it turned on the way the books were acquired, not on training itself. If the courts keep drawing a line between lawful acquisition and unlawful acquisition, news publishers with paywalled content have a possible angle, because the USA Today Co complaint reportedly alleges the use of paywalled articles, but that is an inference from the pleadings, not a result.
Why does the training-versus-retrieval distinction matter commercially?
Because it determines which tool can fix the problem. Litigation looks backwards at what a model learned. It cannot easily price the traffic that arrives every day when an assistant fetches your page to answer a live question. A fair use ruling for training would still leave retrieval, summarisation and substitution open, which is why the USA Today Co theory about summaries replacing visits is the more interesting development for revenue.
For that live-retrieval layer, publishers have operational options that do not depend on a court: block or rate-limit the agent, charge for access, or monetise the retrieval itself. We cover the trade-offs in should publishers block or monetise AI crawlers. At the edge, a layer such as blankspace's can recognise a Live Search Agent request and respond to it differently, which is a commercial route that works whatever happens in the training cases.
What should publishers do while the cases run?
Keep a record. Dated server logs showing which AI user agents fetched which pages are the evidence that turns a grievance into a claim, and our guide to finding AI crawler traffic in server logs shows how to build it. Make access terms explicit through robots.txt, licensing signals and terms of use, so that "never asked permission" is a fact you can prove. Keep licensing conversations open, because Press Gazette's tracker shows publishers suing and signing in the same month. And do not plan revenue around a damages award: the one large precedent paid roughly $3,000 per book across hundreds of thousands of works.
Frequently asked questions
Has any court ruled that AI training on news articles is infringement?
Not in the cases we reviewed. The closest rulings are Judge Alsup's 2025 decision on books, which treated training on lawfully obtained works as potentially fair use, and Judge Stein's 2025 refusal to dismiss the New York Times claims, which only allows them to proceed.
How much did the Anthropic settlement pay?
The final approved settlement is $1.5 billion plus interest across 482,460 books, which is about $3,000 per work before fees and expenses. It was a class action by authors, so it is a reference point for book licensing, not a measure of what news content is worth.
Why did the Penske Media suit against Google fail?
Judge Mehta found the plaintiffs had not plausibly alleged an agreement, a distinct tied product, antitrust standing or a properly defined market for the claims he dismissed. Penske's dismissal was without prejudice, so it can try again or appeal.
What is USA Today Co asking for?
It is seeking $250 million in damages from OpenAI. Public reports of the complaint do not describe other remedies, and we could not confirm the court from the coverage we read.
Should a publisher sue, license or block?
Those are not mutually exclusive, and the right mix depends on size, evidence and appetite for cost. Large publishers with strong evidence may litigate; most will rely on access controls and commercial terms. This article is general information, not legal advice, so take advice from a qualified lawyer before acting.