On July 21, Judge Araceli Martínez-Olguín gave final approval to Anthropic’s $1.5 billion settlement with the authors whose books were pirated to train Claude: roughly $3,000 per book across more than 482,000 titles, with over 91% of them already claimed and $101 million in attorneys’ fees riding along. It is the largest copyright settlement in US history and the first of the big AI training cases to reach the end, while suits against OpenAI, Meta and others grind on.
The legal line worth memorizing came earlier, from Judge William Alsup, before he retired: training a model on books you lawfully bought is fair use. Downloading them from piracy sites is infringement, and it stays infringement no matter how transformative the model that ate them. Anthropic paid $1.5 billion for how it acquired the data. The training itself was never the exposure.
The coverage this week is authors and checks, and fair enough; $3,000 a book is real money for a working novelist. The settlement also handed a number to your buyer’s procurement team, and that is where it starts changing deals.
I’ve watched a footnote become a deal gate before
At Zeta I sold data-driven marketing through the years when GDPR and CCPA gave the question “where did this data come from” legal teeth. Before that, provenance lived in a footnote of the MSA. After, deals passed through privacy reviews that hadn’t existed a year earlier, and vendors who couldn’t produce a clean sourcing answer stopped making shortlists. Nobody announced the change. The checklists just grew a row.
Training data is next, and this time the row arrives with a dollar figure attached. Statutory damages for willful copyright infringement run up to $150,000 per work, and Alsup’s framework put Anthropic’s theoretical exposure into numbers no board could sit with, which is why the case settled. Risk teams can’t price “maybe AI has a copyright problem.” They can price a settled class action with a per-book rate. The security questionnaire that already asks about agent governance is about to grow a provenance section.
Indemnity splits the field
The frontier labs saw this coming. OpenAI has offered its Copyright Shield to enterprise and API customers since 2023; Microsoft has the Copilot Copyright Commitment; Google indemnifies across much of its generative stack; Anthropic’s commercial terms commit it to defend customers over authorized use. Read the fine print, though, and the protection is narrower than the press releases: most of these cover what the model outputs, not the vendor’s own training-data liability, and the conditions and caps vary wildly from contract to contract.
An indemnity is also only as good as the balance sheet behind it. Anthropic could write a $1.5 billion check and keep operating. The 40-person AI app in the same bake-off cannot, and its indemnity clause, if one exists, is a promise from a company that would not survive the claim. If you sell an AI product built on somebody else’s model, you inherit the stack’s provenance whether you like it or not, and your buyer’s counsel will eventually map the chain: whose model, trained on what, indemnified by whom, up to how much. Map it first.
Get the answer in the packet
If you sell AI: write the provenance answer down before somebody makes you. Where the training data came from, what was licensed, what your model vendor indemnifies and to what cap, and what you cover on top. Put it in the standard security-review packet next to the SOC 2 report. The first vendor in a category to answer cleanly sets the bar the rest get graded against, and this question is cheap to answer today and expensive to answer under deadline in a stalled deal.
If you buy AI: two additions to the evaluation. Ask whether the indemnity covers the training data or only the outputs. Then read the cap out loud in the meeting, because a $500,000 cap against a $150,000-per-work exposure is a coupon, not coverage. And ask the provenance question even when the demo is beautiful. Anthropic had the best model in the world and still wrote the largest copyright check in American history for the way it filled its bookshelf.