AI copyright news, graded by source

Anthropic describes a ledger of what trained each Claude model, and says its book-scanning project cost tens of millions of dollars

Opposing a motion to compel, Anthropic and its employees describe two things in sworn declarations and a brief. First, a "Tokens Formula Log": a ledger with more than 40 formulations of the training data mix for its models, listing datasets, their origin and their weights. Anthropic says copies were produced by mistake in March 2025 and taken back, and that the ledger can be inspected only in a secure room; its chief executive and chief science officer are among those with access. Second, the dataset it built by buying and scanning print books, which it says cost tens of millions of dollars and whose existence was a closely guarded secret until this case. Anthropic also says it offered to admit that the training mix for a non-commercial, abandoned model called Finch included datasets associated with Books3, LibGen and the Pirate Library Mirror. It describes the downloaded book datasets it gave the plaintiffs as publicly available on the internet. These are Anthropic's own statements.