They Read Everything.
Here is the plot twist nobody put on a banner: the AI labs mostly won the copyright fight. The reading was legal. The shoplifting was not.
Training a model on the entire internet? Judges keep calling it fair use. Transformative, even. Meta won that argument in June 2025. So did Anthropic. The labs aren’t paying for what they learned. They’re paying for how they got the books — and for the days the model forgets it isn’t supposed to recite them.
Across town, The New York Times isn’t arguing the reading was illegal — it’s arguing ChatGPT handed its articles back, word for word. In Munich, a court already agreed on song lyrics and ruled OpenAI directly liable. In London, the judges went the other way: a model’s weights, the High Court held, are not a copy of anything.
So the rule, as of this fortnight: learn all you want. Don’t keep the receipts. Don’t quote.
Footnotes
- Meta prevailed on the training-as-fair-use question in Kadrey v. Meta (J. Chhabria, 25 June 2025). A court likewise found Anthropic’s training fair use in Bartz v. Anthropic; the liability that produced the $1.5 billion settlement attached to acquisition — books obtained from pirate libraries — not to the training itself. The settlement covers roughly 500,000 works at about $3,000 each. Authors urged final approval at a fairness hearing on 14 May 2026, with claims by then at roughly 93%.
- Munich: GEMA v. OpenAI, 11 Nov 2025 — a German court held OpenAI directly liable for memorized song lyrics. London: Getty Images v. Stability AI, 4 Nov 2025 — the High Court held that a model’s weights are not an infringing copy, though Getty retained limited trademark findings. Neither is a clean sweep for anyone.