Plenty of teams are running a Claude key in staging right now, with a legal reviewer asking where the training data came from and a ship date already on the board. For them the piracy allegation is the least interesting part of this filing, because Anthropic has answered a complaint about Library Genesis before and that answer cost $1.5B [5].
Scraping happened at a particular time, on particular machines, and a court can order it stopped. What the Regional Court of Munich addressed was different: whether lyrics retained inside a model amount to reproduction, and whether a model reciting them counts as communication to the public. Ruling against OpenAI rather than Anthropic, the court said yes to both [8]. It also held that the text and data mining exception did not cover this, because permanent memorisation goes beyond transient analysis and the rightsholder had opted out [9]. Europe's exception carries a second condition that no lab can retrofit: it only applies to works the miner had lawful access to, and a pirate library is never that [11]. Read that way, the copy lives in the weights themselves, not in a crawler log a lab can promise to purge, and the weights are what gets served, tuned and hosted.
Two things keep this from being a repricing event today. The Munich judgment is not final, and a German regional court does not bind a jury in Northern California [10]. The $150,000 per composition is also the statutory ceiling for wilful infringement rather than a sum any court has awarded [6].
The settled authors' case paid about $3,000 a title, split with the publisher, leaving roughly $1,500 for each side [7]. Divide the demand by that settlement and you get 50 times the per-work figure, or 100 times the author's half [13]. The smaller number is what two sets of lawyers agreed to. The larger one has not been argued in front of anybody.
Plenty of teams treat training data as the vendor's problem on the theory that they do not train models themselves. In practice they pick one frontier model, tune prompts and evals to its habits, slide a fine-tune into the loop, and find out a quarter later that the abstraction layer was never built.
So use two axes, both answerable this week without counsel: the number of weeks a model swap would really take, and whether the product puts long verbatim runs of someone else's text in front of users or reads that text and emits something short. Quick swap with short outputs means you are watching this, not exposed to it. Quick swap with verbatim output makes it a filtering and logging job you own either way. Slow swap with short outputs means the indemnity wording is the whole of your position. Slow swap plus verbatim runs is the quadrant where the Munich reasoning, if it survives appeal, lands on the buyer rather than the lab, and that is the one worth pricing before anyone asks.
