The demand to destroy trained AI models now appears in ongoing US lawsuits, but the outcome is uncertain because the US Department of Justice defends training as fair use. So record now, per workflow, which models you use and what you know about their training data, so you can switch quickly if a provider comes under a ban.
On 5 September 2026 the news agency Reuters reported that the Seattle Times and Newsday brought a copyright case against OpenAI and Microsoft in the federal court for the Southern District of New York. According to the reporting, the newspapers accuse the companies of collecting and using their journalism, including paid articles behind a paywall, without permission to train models such as GPT and Copilot. The newspapers explicitly ask the court to destroy copies of their work, training datasets and AI models that contain that material. In our assessment, the most important shift is not the lawsuit itself, but that "destroy the model" is now a concrete demand in an ongoing procedure and no longer a theoretical figure.
What exactly do the Seattle Times and Newsday demand from OpenAI and Microsoft?
According to Reuters' reporting on the lawsuit by the Seattle Times and Newsday, the case turns on three key points:
- the unauthorised collection (scraping) of their journalism, including paid articles;
- the inclusion of that material in training datasets for GPT and Copilot;
- a request to the court to destroy copies, datasets and models that contain the articles.
The first two points are familiar from earlier procedures. The third point sharpens the case legally and operationally: it asks not for damages or a licence, but for the undoing of the training itself.
How does this destruction demand relate to the case of The New York Times?
The demand does not stand alone. In the original complaint by The New York Times against Microsoft and OpenAI from December 2023, the newspaper likewise asks, on the basis of 17 U.S.C. § 503(b), for the destruction of all GPT and other language models and training datasets that contain the Times' work. That statute is cited in the United States by the newspaper as the basis for a request to the court to order the destruction of infringing copies and tools.
In addition, MediaNama describes a lawsuit of 26 June 2026 in which, according to that publication, 35 local and regional publishers, together accounting for roughly 400 newspapers, speak of "industrial-scale" automated scraping, the removal of copyright information and the memorising and reproduction of articles by the models. In our assessment, a recurring picture is emerging here: in the cases mentioned, multiple US publishers link the use of their texts during training to a remedy that may extend to destruction.
Why does the destruction demand clash with the US government's fair-use position?
Opposing that demand is a contrary signal from the federal government. Reuters reported on 2 September 2026 that the US Department of Justice, in the case of The New York Times, took a position that supports OpenAI, and that it generally defends the training of AI models on copyright-protected text as fair use rather than as infringement. That runs directly counter to the newspapers' premise that training itself already amounts to infringement.
Our analysis: this creates an open legal question with three layers. Is a trained model legally an infringing copy, a derivative work or a lawful tool? Is there any precedent for "rolling back" a complex, already deployed system? And how does a court weigh destruction against alternatives such as damages, licences or retraining? The sources give no outcome here; they show that the question is undecided. We therefore express no expectation about the outcome.
What does a destruction or retraining order mean operationally for users of these models?
Even without predicting the outcome, you can think through the operational consequences conceptually. An order to destroy training datasets or AI models that contain the newspapers' works could, in our assessment, have an impact on:
- the removal or shielding of specific datasets at the provider;
- the retraining or replacement of model versions;
- the keeping of a verifiable trail of the provenance of training data;
- consequences for customers who have incorporated such a model into products and workflows.
That last point directly affects organisations in trust-sensitive domains. Anyone who has to absorb a ban or retraining at a provider benefits from arrangements agreed in advance. In an AI contract, that includes concrete audit rights and evidence obligations. The question of whether you build on open or closed models also influences how quickly you can switch, as does the fact that model versions can change within the same name.
Which questions about your AI stack must you now be able to answer?
Regardless of how courts rule on destruction, these cases make the provenance and governance of training data a practical control duty. We therefore place this case in the broader context of AI governance and concrete control duties. Make sure that, per workflow, you can answer the following questions:
- Which provider models do we use, and in which workflows?
- What is publicly known about the training sources and the copyright position of those models?
- What evidence do we have of permissions, licences or fair-use analyses?
- How quickly can we replace or restrict a model if a provider faces a ban or mandatory retraining?
A verification layer can support this. Vera is not a chatbot and not its own language model, but makes visible, per task, which independent models are deployed and which verification steps, corrections and sources accompany them. That can help to link your AI workflows to concrete models and assumptions about data provenance, so that you gain more insight into your exposure if a procedure forces changes. It gives no guarantee of correctness and does not take over your professional final judgement; that remains with you.
Sources and references
- Seattle Times, Newsday sue OpenAI, Microsoft, alleging copyright infringement
- The New York Times Co. v. Microsoft Corp. et al. – Complaint
- 35 US Newspaper Publishers Sue OpenAI, Microsoft Over Alleged Copyright Infringement
- US government backs OpenAI in New York Times copyright case
- The New York Times v. Microsoft and OpenAI – Case overview
Sources: The article draws on reporting by Reuters, the complaint of The New York Times via CourtListener, reporting by MediaNama and a case overview on Wikipedia.