Two more names went on the list this week. The Seattle Times and Newsday sued OpenAI and Microsoft, claiming their journalism was scraped into training data without permission and that the models spit chunks of their reporting back at users who ask. The legal theory is familiar by now. The plaintiffs are not.
Up to this point the marquee cases came from outlets with national brands and litigation budgets to match. A metro daily in Seattle and a Long Island tabloid are a different kind of plaintiff. They cover school boards, county budgets, port authorities, storm damage. That work is expensive, locally specific, and almost impossible to replace, which is exactly why it’s valuable as training data and exactly why nobody has been lining up to pay for it.
Two doors, one of them locked
There have really only been two paths available to publishers since the training-data fight started. Sign a licensing agreement, or sue. The licensing door has opened for a handful of large media companies with enough traffic, archives, and lawyers to make a deal worth doing. It has not opened for the roughly 200-person newsroom covering a single metro area. Suing is the only way for those outlets to force a number onto the table.
Which is why I read this filing less as a bid to stop OpenAI and more as an invoice sent through a courthouse. Nobody at Newsday thinks a judge is going to make GPT unlearn Long Island. They think a settlement, or the credible threat of one, is worth more than another year of being ignored.
The regurgitation claim is the load-bearing one
The strongest part of these complaints has never been “you trained on us.” That argument runs straight into fair use and a defense budget the size of a small country. The sharper allegation is the second one: that the models reproduce passages from the reporting verbatim when prompted. Output, not intake. It’s more concrete, easier to demonstrate to a court, and harder to wave off as transformative.
It’s also the claim most vulnerable to being engineered away. Filters, deduplication, retrieval that cites rather than recites. If the copying stops, the evidence gets thinner, and the underlying question of whether training itself required a license goes right back to being unresolved.
What a win would actually buy
Assume the papers prevail or settle well. A per-outlet payment lands, the lawyers take a cut, and a regional publisher gets a year or two of runway. That is not a business model. It’s a one-time transfer that does nothing for the next thousand local newsrooms with no archive worth licensing and no budget to file.
The cases keep stacking anyway, because the alternative is accepting zero. OpenAI and Microsoft can absorb a long docket; individually, none of these suits threatens them. Collectively they’re building a record, and courts read records.
Seattle Times and Newsday filed in early September 2026. The New York Times case, filed at the end of 2023, still hasn’t reached a verdict.