Models are told to name their sources

A federal court in Manhattan ruled on 8 April that generative model developers must, upon request from a rights holder, disclose the copyrighted works that went into their training data, and the ruling lands in the middle of the industry’s great copyright fight with the force of a procedural bomb. The decision does not say whether the training was legal, which is the question the industry has been litigating for three years; it says that the answer cannot be kept secret, and that the parties asking for it are entitled to know what went in. The procedural move may do more to settle the fight than any of the substantive rulings that preceded it.

The reasoning of the court is worth reading for what it does not say. The judge accepted that the technical task is difficult, that training corpora are assembled from web crawls, synthetic data and licensed collections in proportions that the companies themselves describe loosely, and that reconstructing an exact list for any given model may be impossible in practice. He then ruled that the difficulty is not a reason to withhold the information, and that the companies must produce what they can, in good faith, with the burden of explaining what cannot be produced. The standard is not perfect disclosure; it is honest disclosure, and the difference is what the industry has been resisting.

The consequences will be felt before the next substantive ruling. The request is a discovery order, which means the first wave of lawsuits can now go to the evidence stage, and the evidence stage is where the cases will either settle or collapse. The companies have maintained that their training was transformative, that the works were used to learn patterns rather than to be reproduced, and that the public benefit outweighs the artists’ claims; the disclosure will let the courts test that narrative against the actual composition of the training data, and the composition has already proved to be more revealing than the companies’ summaries suggested in the earlier rounds of this fight.

The artists’ side of the argument has been treated, fairly, with some scepticism in the coverage, because the copyright claims are broad and the remedies demanded have ranged from the plausible to the punitive. The ruling shifts the ground beneath both positions: the artists no longer have to prove what the training data contained, which was always their weakest point, and the companies can no longer treat their training corpora as trade secrets, which was always their strongest one. The fight will now be conducted on the actual record rather than on the companies’ own summaries of it, and the record, as it emerges, will tell a story that neither side has wanted to tell in public.

The ruling is closer to a reckoning than a battle. The generative industry was built on a scale of copying that no legal regime had anticipated, and the companies’ defence has always leaned on the convenience of opacity; the Manhattan decision removes the convenience while leaving the argument intact. The disclosure will be slow, partial and contested, and the fight over the substance, over what copying is fair and what is compensated, will run for years. But the era in which the industry could describe its own training data with impunity has ended, and for a field that makes its money from the arrangement of other people’s words, that is a change in the weather that everyone can feel.