Transmission // 06
One and a half billion for the library
3 min
A court called AI training fair use and Anthropic still paid a record settlement. The distinction it turned on is the one every lab should have been watching.
The copyright question is not being settled by one landmark case. It is being settled by a dozen partial ones, and the shape is finally visible.
Start with the number everyone quotes and almost nobody reads correctly. Anthropic settled Bartz v. Anthropic for $1.5 billion — the largest copyright settlement on record, approved by the Northern District of California, working out to roughly $3,000 per book across more than 400,000 titles.
The split that actually happened
Judge William Alsup did not rule that training was infringement. He ruled the opposite: training on books that were lawfully acquired was fair use, and “quintessentially transformative” at that.
What was not fair use was how the corpus was obtained and kept. Anthropic had downloaded millions of copyrighted books from pirate sites to assemble a permanent general-purpose library. In Alsup’s words, it “had no entitlement to use pirated copies for a central library. Creating a permanent, general-purpose library was not itself a fair use.”
So the billion and a half was not the price of training. It was the price of the shelf. Acquisition and retention are the exposed surface; the gradient update was fine.
Three separate questions, three different answers
The discourse collapses these constantly. The courts do not.
Is training on protected work fair use? Increasingly yes, where the copies were legitimately obtained — and unresolved where the use substitutes for the original. In Thomson Reuters v. Ross Intelligence the court granted partial summary judgment for Thomson Reuters in February 2025 on the question of copying; the appeal, still pending, covers both whether the headnotes are copyrightable at all and whether the use was fair.
Can the output infringe? Yes, and this is the dangerous one for anybody deploying. Courts have held that generated outputs may infringe where they are substantially similar to protected works — putting liability on generation, not just training, and therefore potentially on the user.
Can machine output be owned? No. The Supreme Court denied certiorari in Thaler v. Perlmutter on 2 March 2026, leaving human authorship intact as a requirement of US copyright. Your prompt does not make you an author.
Europe has no fair use to weigh
On 31 July 2026 the Munich Regional Court I decided GEMA v. Suno (42 O 763/25), the first European ruling on a generative music tool. It prohibited four separate acts across six compositions: reproduction for training in the US, reproduction through memorisation in the model on German servers, communication to the public by offering the model, and reproduction through the outputs.
Read that second one twice. The court held that a work memorised in weights sitting on a German server is an infringing reproduction — the model itself as a copy. It took jurisdiction over the US training acts through a venue rule for collecting societies, applied US law to them, and rejected fair use.
What this means if you ship things
- Provenance of the corpus is the liability, not the training run. Where your data came from is a discoverable fact.
- Output you generate is not yours in the copyright sense. If ownership matters, a human has to make it.
- Substantial similarity is your problem, not your vendor’s.
- Indemnities are worth what the cap says. Read the carve-outs first.
The training question is closer to settled than the headlines suggest. The provenance question just got very expensive.