Can AI companies launder copyrighted music?

Suno allows users to generate music by entering text descriptions. The record labels, who have not licensed their music to the company, claim that Suno originally trained its AI on vast amounts of their copyrighted recordings without permission. Suno recently released a newer version of its music-making tool, which it says was trained on a new set of data. However, the record labels allege that this new version was trained partly on the outputs of previous, infringing versions. Because those older versions were fed stolen songs to learn the structures of music, the labels argue that the new version is simply inheriting the same underlying issues.

AI systems are built through a process called training, where they study millions of examples to learn patterns. Think of this like teaching a student by showing them every painting in a museum. If the student studies stolen art, they learn the techniques of that art. The record labels argue that Suno is using a technique called distillation. In this process, a newer, smaller AI is trained to mimic the specific behaviors and patterns of a larger, older AI. If the older AI was originally trained on copyrighted songs, the new AI essentially copies the knowledge—and the legal flaws—of its predecessor. The labels contend that even if the new version did not ingest original copyrighted audio files directly, it still carries the influence of those files because it was taught to act just like the version that did. To the labels, this is like washing dirty money through a bank; the origin of the value remains tainted even after it has been moved around.

This case challenges the idea that AI companies can simply wipe their hands clean by training new versions of their tools. It suggests that if an AI foundation is built on unauthorized data, that data might be permanently baked into the system's identity. If the court agrees with the labels, it could force companies to rethink how they update their software, potentially requiring them to prove that their new versions were built entirely from scratch using only licensed or clean materials. This could set a major precedent for any AI company that creates tools using massive, unregulated datasets, forcing them to address the legal debt inherited from their earlier versions.