Trump administration told a court OpenAI's training data is fair use. The judge still decides.

The question of where your model gets its training data used to sit with legal and engineering. It just moved up to your desk. The Trump administration put its name behind OpenAI in court. New publishers are suing. Google is calling Hollywood with a checkbook. Let me catch you up on what changed and what you need to say in your next review.

The deep cut

• Data provenance is a product decision now. OpenAI faces the New York Times, the Seattle Times, Apple, and a federal court all at once.
• A government brief is not a court ruling. The Trump administration backed OpenAI, but Judge Stein still decides the fair use question.
• Licensed data buys goodwill, not just legal cover. Google offered Disney around $40 million per character to make AI feel safe.

Washington put a thumb on the scale

The government stepped into a private copyright fight. In a 20-page brief, US attorneys argued that training LLMs on copyrighted text is fair use, and that ruling otherwise would "thwart such creative and scientific progress while hindering American prosperity."

Read the fine print before you relax. The brief itself concedes that "the fair-use inquiry hinges on the specific facts and uses at issue in each case." Hayden Field notes the administration has leaned on these statements of interest across many cases, and one official calls them "incredibly" successful. But the authors have no jurisdiction here. Judge Sidney Stein in Manhattan still rules.

The word-count defense

Microsoft is not waiting on Washington. It handed publishers' experts 8.2 million Copilot chat logs, picked because they were the ones most likely to spit back news content. Even in that stacked sample, Microsoft says only 59,545 chats shared even 16 words with news articles, and the authors' suit found just 24 responses with 30 matching words across all those logs.

The argument is that reproduction is rare, so training is transformative. The counterexamples are ugly, though. The Seattle Times complaint shows ChatGPT reproducing an 88-word verbatim stretch of its Pulitzer-winning Boeing 737 MAX coverage from just a headline and URL. When a model can regurgitate your best work on demand, "rare" stops being reassuring.

The lawsuits are getting personal

This is not one case anymore. The Seattle Times is suing companies that fund it. Microsoft Philanthropies underwrites some of its journalism, and both firms backed a $10 million Lenfest fellowship that included the paper. CEO Alan Fisco called it "not an easy decision" but said they must defend content that costs "millions of dollars a year to produce." Then the paper's own union pointed out the company won't promise not to replace newsroom jobs with AI.

Apple is fighting on a different front. It accuses OpenAI of destroying evidence after a former Apple engineer allegedly kept a company MacBook, downloaded a confidential circuit schematic, and used it at OpenAI. OpenAI calls the whole thing "a mess of Apple's own making." Either way, provenance disputes now reach into your hiring and your hardware.

Paying for the data cleans up the story

Google is taking the other road. It is calling Disney, Warner Bros., and Universal, offering around $40 million for the right to generate outputs featuring a single copyrighted character, with totals that could reach billions. Licensed data is cleaner than scraped data, and Charles Pulliam-Moore makes the sharper point: an official Disney-approved Darth Vader from Gemini makes AI feel legitimate to people who now view it negatively.

There is a cost signal here for you. OpenAI has struck deals with the AP, News Corp, and Axel Springer, three of which top $300 million combined. Clean data has a price, and it is climbing. The question is whether you pay it up front or fight it in discovery.

Three questions for your team

• Can we name, right now, where the training data behind every model we ship comes from, and can we defend each source in front of a judge?
• If a court sides with publishers next quarter, which of our features breaks, and what is our fallback plan for the data underneath them?
• Where would a licensing deal buy us real trust with users, and what is that worth against the cost of a lawsuit like the Seattle Times filed?