Seattle Times and Newsday sue OpenAI and Microsoft, demand the models themselves be destroyed

What happened

The Seattle Times and Newsday filed a lawsuit against OpenAI and Microsoft, alleging the companies ingested their journalism into AI training data without permission and that the resulting models reproduce passages from their reporting when prompted. The plaintiffs are seeking not just damages but the destruction of the training datasets and the AI models themselves.

Context

The New York Times brought the first of these actions in 2023, and the line of cases has since grown to include Ziff Davis, Merriam-Webster, Encyclopedia Britannica, and a group of nearly 400 local newspapers alleging that chatbots divert readership and cost subscription revenue. The Seattle Times and Newsday enter that sequence with a particular complication: Microsoft and OpenAI have funded some of the Times' journalism projects and fellowships, meaning the defendants have both a financial stake in the plaintiff's output and a training interest in its text.

How it works

The technical allegation is that text from the two publications was incorporated into training corpora used to fine-tune large language models, and that those models can reproduce passages from the source articles when a user prompts them. Microsoft is named as a defendant because Copilot is built on OpenAI's model technology, making it a downstream product in the plaintiffs' framing. The remedy goes well beyond the injunctions typical of software-copyright cases: the plaintiffs want the defendants to identify and eliminate the specific weights or checkpoint layers that encode their text, destroy the training datasets, and discard stored copies of the articles. Neither source describes the model architecture, parameter count, or the technical method the plaintiffs intend to use to demonstrate that specific text is recoverable from specific parameters.

Our read

The remedy is the part that should make people pause. Ordering the "destruction" of models that "incorporate" a plaintiff's works asks a court to require identification and elimination of specific weights in a model whose internal structure neither source specifies. Proving which subset of a model's parameters encodes a paragraph of Seattle metro reporting is an unsolved problem; neither source names the testing methodology the plaintiffs intend.

The more durable shift is the legal framing. The complaint does not stop at verbatim copying. It characterises the output as "derivative imitations" of the source. If a court accepts that a model's output is a derivative work whenever it was trained on the underlying text, the standard reaches any fine-tuning pass, any LoRA, any checkpoint whose corpus touched the plaintiff's material. That is a wider net than "you scraped our website."

The funding relationship complicates the narrative the complaint leans on. Microsoft and OpenAI have directly funded the Times' reporting. That is not a legal contradiction, but it gives the defendants a settlement lever the New York Times, which received no such funding, does not have.

What this changes

For a small studio running ComfyUI and local inference, nothing in the day-to-day pipeline changes. What shifts is upstream. If the "derivative imitation" standard gains traction, the pool of legally available fine-tuning datasets narrows. If the studio performs any LoRA training on scraped text-to-image or text-to-video data, this case reinforces that unlicensed training data is an escalating risk. Practical step: audit the provenance of any custom checkpoints, confirm the training data was licensed or public-domain, and keep that documentation in one place. The separate 400-newspaper suit targets revenue loss from chatbot traffic diversion, a theory that does not touch a local inference stack.

License

This is a lawsuit, not a model release or software launch. No licence applies.

Key takeaways

• The Seattle Times and Newsday seek destruction of the models and training data, not just damages, an aggressive remedy that requires identifying specific weights encoding specific text.
• The "derivative imitation" framing in the complaint, if adopted by a court, would extend copyright exposure to any fine-tuning or LoRA trained on the plaintiff's material, not just verbatim scraping.
• Microsoft and OpenAI's prior funding of Seattle Times journalism gives the defendants a settlement dynamic the 2023 New York Times suit did not have.
• For a local-model studio, the actionable step is auditing training-data provenance for any custom checkpoints before the legal standard tightens further.

Sources

1. Seattle Times and Newsday are the latest publications to sue OpenAI and Microsoft — tier 2
2. Seattle Times and Newsday sue OpenAI and Microsoft for infringement — tier 2