The Seattle Times and Newsday Want the Models Destroyed, Not Just Damages

Two American newspapers have asked a federal court to destroy AI models.

The Seattle Times and Newsday filed jointly on 4 September 2026 in the US District Court for the Southern District of New York, naming both OpenAI and Microsoft. The complaint describes generative AI as a snake eating its own tail, a system consuming the journalism it depends on to function.

The claims themselves are familiar by now. The remedy is not.

What the complaint alleges

The papers say OpenAI and Microsoft were methodically scraping their articles in a way that bypasses paywalls, and that the resulting material went into datasets used to train and operate ChatGPT, Microsoft Copilot and Bing AI features.

From there the alleged harms stack up. The products can reproduce passages from the reporting verbatim. They closely paraphrase articles. They answer questions in ways that remove any reason to visit the sites or buy a subscription, which the papers tie directly to lost traffic and lost digital advertising revenue.

Two further allegations are worth separating out, because they are not really about copying at all.

The first is hallucination. The complaint says the models attribute false information to the outlets, which is a reputational claim rather than an economic one.

The second is removal of copyright management information, the byline, credit and rights data attached to an article. That is a distinct cause of action under section 1202 of the Digital Millennium Copyright Act, and it does not require proving that the copying itself was infringing.

The unusual part is what they are asking for

Most publisher suits ask for damages and an injunction against future use. This one goes further. The Seattle Times and Newsday want an order requiring destruction of the copies of their works, the training datasets containing them, and any AI models that incorporate them.

Destruction of infringing copies is not an exotic request in copyright law. Courts have that power under section 503 of the Copyright Act, and it is routinely applied to counterfeit inventory and pirated files.

Applying it to model weights is a different matter, and it is untested.

A trained model is not a database with rows you can delete. Once a document has shaped a set of parameters, there is no clean subtraction that removes its contribution while leaving everything else intact. Research into machine unlearning exists, but nothing in it currently supports a court order that says remove these specific articles and prove it.

So the practical version of the remedy is: delete the model, retrain without the material, and absorb the cost. That is why this ask is much heavier than it appears in a headline, and why it is unlikely to survive intact even if the plaintiffs win on liability. It does, however, set a ceiling for settlement talks, which may be the point.

Where it sits in the queue

The New York Times filed on near identical grounds in December 2023, and that case is still running. CNN sued Perplexity earlier this year. There are now dozens of copyright actions in progress against OpenAI, Anthropic and Meta.

Not every publisher has taken the litigation route. The Associated Press signed a licensing deal with OpenAI. So did Vox Media and The Atlantic. The industry has effectively split into two camps, and each is watching what the other one recovers.

What happens after these cases end

The settlements are turning out to be their own fight.

Authors are currently pushing back on how publishers and literary agents propose to divide the Anthropic settlement, arguing over who is actually owed the money that a copyright class action produces. That dispute is instructive for any newsroom eyeing a payout: winning the case establishes a pool, not a distribution.

Why two regional papers matter

The Seattle Times serves one metro area. Newsday covers Long Island. Neither is a national title with an unlimited litigation budget.

That is the detail worth holding on to. Copyright litigation against well funded AI companies used to be the preserve of the largest publishers, because only they could afford the discovery fight. If regional metros can now fund these cases, whether through contingency arrangements or coordinated filings, the number of active suits goes up rather than down.

For OpenAI and Microsoft, the risk was never one adverse judgment. It is the cost of defending against a long queue of them while the underlying legal question about training data stays unresolved.