On July 20, a federal judge in California granted final approval of the $1.5 billion settlement Anthropic negotiated with a class of authors. It’s the largest copyright settlement in US history: roughly 500,000 works, about $3,000 per work. Every headline says the same thing: the biggest number ever recorded.

The number isn’t the story. The story is the line the court drew to get there, and that line still isn’t on most legal teams’ vendor checklists.

The question your procurement team isn’t asking yet

Here’s the detail most of the coverage glossed over. The same court had already ruled, in an earlier decision from Judge William Alsup, that training an AI model on copyrighted books is fair use. That question is settled. It is not why Anthropic paid $1.5 billion.

What cost the money was how the copies were obtained. Anthropic bought and scanned millions of physical books, and the court found that part entirely lawful. But it also downloaded hundreds of thousands of titles from piracy sites like Library Genesis and Pirate Library Mirror, and that’s where liability attached. Per TechCrunch’s coverage of the ruling, the court described the roughly $3,000-per-work payout as four times the statutory minimum for ordinary infringement, specifically because the material was pirated, not because it was used for training.

Translate that into a procurement question: the question that matters for any company evaluating an AI vendor isn’t “did you train on copyrighted material?” The answer is almost certainly yes, in some form, for any large model on the market today. The question that now carries an actual price tag is “how did you acquire the data you trained on?” That price, as of this month, is roughly $3,000 per work when the answer turns out to be “we downloaded it from a piracy site.”

What the settlement doesn’t resolve

The scope matters here, because most of the press coverage flattened it. The settlement covers specific claims against Anthropic through August 2025. The Authors Guild was explicit that future claims and output-based claims are excluded from the release, per its statement following final approval. Authors keep the right to sue over future conduct.

The settlement also requires Anthropic to destroy every original pirated file it downloaded from Library Genesis and Pirate Library Mirror, per the Authors Guild. That’s not just a payment, it’s an order to eliminate the physical evidence of how the data was acquired, a detail none of the “biggest settlement in history” headlines mentioned. Participation was high too: roughly 91% of eligible authors filed a claim before the March 30 deadline, according to case reporting, suggesting most affected rightsholders knew exactly what had been done with their work.

There’s a bigger point here for any company that licenses or embeds AI models. Because Anthropic settled rather than litigating to a final verdict, the case never reached an appeals court. No binding precedent was set on the underlying question. Google, Meta, OpenAI, and Midjourney all still face similar unresolved suits, and every judge retains the discretion to reach a different conclusion on different facts. This settlement closed one company’s exposure for one acquisition method in one specific time window. It did not close the industry’s exposure.

Why this matters even if your vendor isn’t Anthropic

Here’s the connection most people aren’t making yet: Anthropic isn’t an isolated case, it’s just the first one to resolve. If your company depends on a language model for anything that touches intellectual property, customer-facing content generation, or any workflow where data provenance carries legal weight, your vendor’s exposure functions, in practice, as your contractual exposure. Most AI model licensing agreements still don’t have indemnification clauses written for this specific scenario, because until this month there was no real reference number to write them against.

Now there is. $3,000 per work is a concrete anchor for modeling risk, something any legal team signing a significant AI vendor contract should be running the numbers on this week, not after a demand letter shows up.

What we do differently in vendor evaluation

In the AI Maestro discovery process, we evaluate AI vendors as part of the Process Reality Map, not as a separate legal formality bolted on at the end. That means asking about training data provenance before a contract is signed, not after a headline like this one forces the question. We’ve already written about how to structure that trust and privacy evaluation, and the pattern keeps repeating: the Claude Mythos leak tested vendor trust the same way, before the companies using it had time to react.

The lesson here isn’t “watch out for Anthropic specifically.” It’s that the data-provenance question now has a market price, and it still isn’t in most of the contracts your company is signing this quarter.

Add data provenance to your vendor evaluation