The row in a legal review that says "training data provenance" does not get easier to fill in this week. Liability in these cases has been landing on how a corpus was obtained rather than on the act of training. A 2025 decision found that Anthropic could lawfully train on books it had purchased while remaining exposed on piracy [6], and the piracy exposure is what produced a $1.5 billion settlement with authors [7]. Nothing in the government's filing disturbs that split, and the filing effectively says so when it puts the inquiry on the specific facts and uses at issue in each case [5].
Teams often assume a content licence buys cover for whatever went into the model. What it actually buys is narrower: a dated document naming a scope and a counterparty who cannot later claim surprise. Those are two different products at two different prices. Only the second one survives contact with a discovery request.
Watch the plaintiff as an operator rather than as a victim. The Times is suing OpenAI over training [2] while licensing its articles and recipes to Amazon for use in Amazon's generative AI tools [11]. For that publisher, litigation and licensing are running at the same time rather than as a choice between them [14]. Dozens of outlets have signed with OpenAI, including the Associated Press and Vox Media [10]. The licence market did not wait for the doctrine, and it will not close because of one brief.
On how much weight the brief carries: the administration has used statements of interest heavily in private litigation, and one official described them as "incredibly" successful at advancing policy aims [8]. It had already argued that AI training should count as fair use in its National AI Legislative Framework [9], and Trump is separately pursuing a defamation suit against the Times [12]. Taken together, the filing looks like a policy position that happened to find a convenient docket, not a finding about these particular facts.
So draw the two axes the rulings actually turn on. One: can you produce a transaction for every copy in your training or retrieval corpus? Two: does the output substitute for the source it learned from? The argument filed this week helps in exactly one quadrant, lawful acquisition paired with non-substituting output. The Anthropic number prices the quadrant directly below it. Everything on the substitution side is still an evidence fight about your specific product, which is what the filing's own carve-out says [5].
The forcing question for each model, dataset and scraper in your stack is which document proves acquisition. Where there is no document, the fair-use argument the government made is not one you can borrow, and a licence bought now is buying a paper trail rather than a legal opinion. The deck usually claims more than that. What holds up on Friday is smaller.
