A mid-career novelist checks whether her book appears on the leaked "Books3" list that surfaced during discovery in the Anthropic litigation. It does. In 2025 Anthropic agreed to pay a class of authors roughly $1.5 billion, close to $3,000 per work, to settle claims it trained Claude on pirated copies. She qualifies, because her book was registered and the copy was exact enough for a court to recognize. Most working writers checking that same list, without a registered copyright or a class-action law firm behind them, do not have an equivalent path.
The underlying tension is that a distinctive style, the thing that makes a creator's work recognizable and valuable, is exactly the kind of pattern a model is good at learning and reproducing at a fraction of the cost. A creator's competitive advantage and a model's training signal are, for practical purposes, the same object viewed from two different sides.
The industry is already splitting into visibly different responses to that tension. Adobe built its Firefly image generator's entire enterprise pitch around training only on licensed Adobe Stock content and public domain material, specifically so it could offer commercial customers indemnification that models trained on scraped web data cannot promise. Getty Images, by contrast, is still pursuing its UK lawsuit against Stability AI over the same underlying claim, a case that has not resolved the core question either way.
Established creators with recognizable brands and legal resources have started negotiating licensing deals, and Universal Music Group, Sony, and Warner sued the AI music startups Suno and Udio in 2024 rather than negotiate first. Emerging and mid-career creators, who often have the most distinctive and therefore most trainable styles precisely because they have not yet been flattened by market pressure toward convention, tend to have neither the leverage to negotiate nor the resources to sue.
What breaks is the old assumption that a style, as opposed to a specific work, is not really property, an assumption copyright law has generally supported and that made sense when imitating a style required years of a human learning to paint or write like someone else. Model training compresses that acquisition cost close to zero.
What is opening up, unevenly, is a market for consent-based training data and creator-side tools that watermark or flag work that has not consented to that use. Neither is close to universal, and neither fully solves the enforcement problem once a model has already been trained on something.
The Anthropic settlement did not decide whether training on someone's style, rather than a pirated copy of one specific book, is theft. It proved that when the copy is exact enough and the class action large enough, there is a number attached to it. Everyone whose case is blurrier than that is negotiating without one.