A group of authors who opted out of Anthropic's $1.5 billion Bartz class-action settlement has filed a separate $75 million copyright lawsuit, targeting specifically how the company sourced their books to train its models. It lands the same week that two prominent legal scholars published an argument that AI model weights may be "probabilistic copies" - a category copyright law was never written to handle - putting the technical physics of memorization at the center of the AI copyright wars.
Key facts
- Authors who opted out of the $1.5B Bartz settlement filed a new, separate suit seeking about $75 million, focused on how Anthropic acquired their works.
- The suit coincides with a new SSRN paper, "Probabilistic 'Copies' in Generative AI Models," by Stanford's Mark A. Lemley and A. Feder Cooper.
- The core legal question: does storing weights that might or might not generate a copyrighted work when prompted constitute a "copy"?
- The $1.5B Bartz settlement was one of the largest copyright resolutions in the AI era; opting out let these authors pursue their own claim.
The lawsuit is the loud part; the paper is the important part. For two years, the fight over training data has been argued with borrowed metaphors - "the model memorized my book," "no, it just learned patterns" - and courts have struggled because the technology doesn't fit the categories copyright law was built on. Lemley and Cooper's contribution is to name the actual question precisely. When a model is trained on a book, it doesn't store the text like a photocopy. It adjusts billions of numerical weights so that, given the right prompt, it might reproduce a passage - or might produce something that merely resembles it, or nothing like it at all. Is that stored bundle of probabilities a "copy" of the book?
Their answer is that the legal outcome depends on a question upstream of traditional copyright analysis: whether a stored representation counts as a valid "system description" of the protected work - a compressed but faithful specification from which the work can be regenerated. If the weights functionally encode the work, that looks like a copy even if no single file contains the text. If they only encode diffuse statistical tendencies that happen to sometimes reconstruct it, the analysis is different. The framing matters because it moves the debate off vibes and onto a testable technical property: how reliably, and under what conditions, can the work be pulled back out?
This is where the theory and the courtroom collide. The $75 million suit turns on how Anthropic obtained the books in the first place - the sourcing, not just the outputs - which is exactly the upstream question the scholars are formalizing. The earlier Bartz settlement, at $1.5 billion, was one of the largest copyright resolutions of the AI era, and its size sent a message that this exposure is not theoretical. By opting out, these authors bet they can do better pressing their own claim, and the new academic framing gives them sharper language to argue that storing patterns extracted from their work is itself the infringing act.
Why should anyone outside the litigation care? Because whatever standard emerges will apply to every company that trains on scraped text - which is all of them. If courts adopt a "probabilistic copy" logic that treats sufficiently faithful weights as copies of their training data, the compliance burden on model builders changes overnight: they would need to prove their systems can't reliably regenerate protected works, or license the data outright. If courts reject it and demand a specific infringing output, the burden shifts back onto rights-holders to catch the model red-handed. The whole economics of open and closed model training sits on which way that goes.
The honest caveat: an SSRN working paper is an argument, not a ruling, and the "probabilistic copy" framing is contested precisely because it is powerful. Critics will note that plenty of learning - human and machine - involves compressing works into patterns without infringing, and that a standard keyed to "could this be regenerated" risks sweeping in ordinary generalization. The $75 million figure, too, is a demand, not an award. What is not in doubt is that the legal system has stopped hand-waving about how these models work and started demanding a rigorous account of the difference between memorizing a book and merely being shaped by it - and that account will decide who pays whom.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)