Six Newspapers Sue Microsoft and OpenAI Over Paywalled Articles in the Training Set

On 2 October, six publishing companies filed suit against Microsoft and ten OpenAI entities in the US District Court for the Southern District of Mississippi. The complaint alleges copyright infringement and the removal of copyright management information in the development of GPT models.

The plaintiffs include Emmerich Newspapers, Ojai Media, and several Coopwood companies, which publish newspapers and magazines across the South. The defendants are Microsoft plus a set of interrelated OpenAI entities. The case is assigned to Judge Henry Travillion Wingate, with the docket last updated on 5 October.

What the complaint actually alleges

Three allegations carry the weight.

First, the complaint says the defendants systematically crawled the plaintiffs' websites, including content behind paywalls, and copied articles onto their servers repeatedly as the models were updated. Repeated copying is the part that matters legally. A single crawl is one act. Re-copying the corpus at every training run multiplies the number of allegedly infringing copies, which multiplies the exposure.

Second, the plaintiffs argue the models exhibit memorization, encoding retrievable copies of the training works. That claim connects to a body of litigation where the question is not whether training happened but whether the resulting model can reproduce protected expression on demand.

Third, the complaint alleges that products like ChatGPT Search and Deep Research retrieve content from the plaintiffs' websites in real time, incorporating it into responses. This is the retrieval layer, separate from training. Even a model trained on licensed data can, at answer time, pull text from a site it was not licensed to read.

The plaintiffs also argue the defendants stripped author credits, publication names, copyright notices, and terms-of-use information from the articles. That maps to the copyright-management-information theory, which does not require proving that the underlying copying was unlawful. Removing a copyright notice from a referenced work is its own violation under US law.

The three counts are direct copyright infringement, removal of copyright management information, and willful theft of the articles.

Why the paywall detail matters

The paywall allegation is the sharpest part of the complaint, and it is a detail worth separating from the general question of whether training on copyrighted text is fair use.

Content behind a paywall sits outside the open web. It is published under terms that condition access, often on payment, and the terms typically prohibit automated collection. A crawl that reaches past a paywall circumvents the access condition itself rather than making a mistake about which pages were public.

That framing lines up with a broader shift in how data owners are litigating. The fight is moving from whether copying is protected to whether the method of collection respected the access conditions the publisher set. Contract and access terms are doing work that copyright alone could not do.

Where this fits in the wider pattern

This case joins a dense stretch of copyright and access litigation. Reddit's contract claims against Anthropic are proceeding, giving platforms a route to challenge data collection outside copyright law. Anthropic, in its submission to Australia's parliament, proposed a "narrow form of conditional approval" for training that would let rights holders opt out through robots.txt, a mechanism Australia's public broadcasters rejected as routinely bypassed.

Across these cases, the same technical facts keep appearing. Robots.txt signals were ignored by major crawlers in documented testing. Paywalled content appears to have been collected. Retrieval features surface publisher text in answers even when the publisher never licensed it.

The publishers here are regional and local outfits whose archives carry real value and whose legal budgets are modest. That is precisely why the case is worth watching. A win for large publishers with large legal teams tests the law. A case brought by regional newspapers tests whether the remedies are usable by the rights holders who cannot afford years of discovery.

What a ruling could change

The outcome will hinge on questions the courts have been circling for two years without settling. Is training on copyrighted text fair use, and does the answer change when the text was behind a paywall. Does memorization turn a training use into an infringing reproduction. Does real-time retrieval create liability separate from training. And does removing copyright notices from works that appear in a dataset count as a violation even if the copying itself is permitted.

None of these is resolved. What the complaint does is attach all four questions to a single set of facts and put them in front of a district court. For anyone building products on top of public web content, that combination is the thing to track. The answers will not arrive quickly, and the industry's current practice of crawling first and negotiating later is exactly what this lawsuit is testing.

Why regional publishers bring a different case

The identity of the plaintiffs shapes the litigation in ways that the large-publisher cases of the past few years did not.

The New York Times case, and the music-industry suits before it, drew on corporate archives with decades of commercial licensing revenue to point to. They could argue that a market existed and was displaced. A group of regional newspapers brings a different evidentiary posture. Their archives are smaller, their licensing history is thinner, and their claim rests more directly on the copying itself and on the removal of copyright information.

That matters for the remedy. If the plaintiffs prevail on the copyright-management-information count, the remedy is available even where the underlying copying is disputed, because stripping a notice is a standalone violation. For a plaintiff without a long licensing history, that count is the more tractable path.

The retrieval layer as a separate exposure

The allegation about ChatGPT Search and Deep Research is worth separating from the training question, because it points at a different kind of liability.

Training is a one-time act, however many times the corpus is rebuilt. Retrieval happens on every query, at the moment a user asks. If a product retrieves text from a publisher's site in real time and incorporates it into a response, the allegedly infringing act is continuous and user-triggered.

That distinction gives plaintiffs a claim that does not depend on winning the fair-use argument about training. Even a model trained entirely on licensed data can, through a retrieval feature, surface text from a source it never licensed. The complaint's framing treats the retrieval and the training as two separate wrongs, which is a stronger structure than tying everything to the training corpus.

What publishers are watching

Two signals will tell the industry whether this kind of case is worth bringing at scale.

The first is whether the court allows the copyright-management-information claim to survive a motion to dismiss. That count does not require resolving fair use, so a ruling for the plaintiffs on it would give publishers a faster route to relief than a full infringement trial.

The second is whether discovery reaches the training pipeline. If the plaintiffs can compel records of what was crawled and when, the case changes from a legal argument about fair use into a factual argument about what the crawler actually did, including whether it reached past paywalls. Practitioners on the publishing side are watching the docket for exactly that kind of order, because it would tell them what evidence is obtainable before they commit to their own suits.