A Second Front Opens in Manhattan
On September 4, 2026, the Seattle Times Co. and Newsday filed a copyright and trademark complaint against OpenAI and Microsoft in the U.S. District Court for the Southern District of New York, joining a docket that already houses the New York Times' long-running suit against the same two companies. The complaint accuses OpenAI and Microsoft of scraping paywalled journalism to train and operate ChatGPT, Copilot and Bing's AI features, then generating hallucinated content falsely attributed to both outlets — a trademark dilution claim layered on top of the copyright count. Seattle Times President and CEO Alan Fisco said the company "must defend our content — which we spend millions of dollars a year to produce." The plaintiffs are seeking damages and, notably, court orders for the "impoundment and/or destruction" of the training datasets and any models built from them — a remedy that, if granted, would reach far beyond a licensing fee.
Timing That Cuts Against the Plaintiffs — On Paper
The suit lands three days after the U.S. Department of Justice filed a Statement of Interest in the same consolidated litigation, arguing that training large language models on copyrighted text is "exceedingly transformative" fair use. The DOJ's September 1 filing — the federal government's first formal position on AI-training copyright litigation — told the court that training should be evaluated separately from output generation, that commercial purpose carries little weight against a highly transformative use, and that training alone doesn't create market substitution. It went so far as to call a rival ruling, Kadrey v. Meta, "deeply flawed" for blending training and output into one continuous use. That's a real headwind for the Seattle Times and Newsday, and it would be dishonest to pretend otherwise: the government now officially favors OpenAI's core defense.
But Acquisition, Not Use, Is Where Publishers Have Actually Won
The complaint's real leverage isn't its fair-use argument — it's the paywall-circumvention allegation underneath it. That distinction is the one that decided the only AI-training case to reach a full ruling and settlement so far. In Bartz v. Anthropic, Judge William Alsup held in June 2025 that training on lawfully acquired books was "quintessentially transformative" fair use — but that assembling a permanent library from pirated copies downloaded from shadow sites was not, because Anthropic "could have purchased or otherwise accessed lawfully" the same material. That distinction cost Anthropic $1.5 billion: a settlement approved July 22, 2026, paying roughly $3,000 apiece across more than 400,000 pirated titles. The U.S. Copyright Office's own 2025 report on generative AI training reached the same conclusion in the abstract, warning that knowingly using pirated or unlawfully accessed works should weigh against fair use even where the training itself is transformative.
Seattle Times and Newsday are, in effect, pressing the Bartz theory against a news publisher rather than a book-pirating site: if the underlying content came from behind a paywall rather than the open web, the DOJ's transformative-use argument — built for lawfully obtained training data — doesn't obviously reach it. Whether scraping a paywall counts as "unlawful acquisition" in the way pirating a book does is untested. But it's the correct fight for the plaintiffs to pick, and it's why this suit is more dangerous to OpenAI than a straightforward fair-use claim would be.
The Strongest Case for the Publishers
It's worth stating plainly: local and regional newsrooms are not abstractions in this fight. The complaint cites industry data showing search referral traffic to midsize publishers fell roughly 47% year-over-year in December 2025 — a real, measurable decline in the audience that used to fund reporting. If a chatbot answers a reader's question using a paywalled Seattle Times investigation without ever sending that reader to seattletimes.com, the newsroom loses both the subscription and the ad impression that paid for the investigation in the first place. A system that lets AI companies free-ride on that reporting indefinitely would be shrinking the very corpus of original journalism that makes any AI answer engine useful in five years. That's a legitimate market-structure concern, not just an incumbent's grievance.
Why Impoundment Is the Wrong Remedy
Even so, the relief the plaintiffs are asking for — destruction of trained models and datasets — is disproportionate to the harm and would set a genuinely bad precedent. Courts have alternatives that actually match the injury: damages calibrated to demonstrated market harm, forward-looking licensing along the lines of what AP and Vox Media have already negotiated with OpenAI, or narrowly tailored data-deletion orders limited to unlawfully acquired material, as the Bartz settlement did. Ordering wholesale model destruction over a subset of paywalled inputs would be a sledgehammer applied to a targeted problem, chilling investment in an industry the DOJ itself now calls a national-security priority. The better outcome — and the one courts are trending toward — is compensation and clean sourcing going forward, not retroactive destruction of general-purpose models trained overwhelmingly on lawfully obtained data.
The real question this suit poses isn't whether AI training is fair use in the abstract — the DOJ has now weighed in, and Bartz already answered it for lawfully acquired material. It's whether paywalled journalism gets treated like the open web or like a locked library.
Judge Stein's docket is about to find out.