Sony is pressing a second lawsuit against AI music generator Udio, this time over more than 30,000 copyrighted recordings identified through audio fingerprinting after Sony gained access to Udio’s training data in discovery. A judge blocked Sony’s attempt to fold these tracks into the original 2024 suit, which remains capped at 333 works, so Sony filed fresh. Udio’s own admissions that its models were built by ingesting “a vast amount of different kinds of sound recordings” from sources including YouTube are now central to Sony’s case, with damages of up to $150,000 per work at stake. Read the full Sony v. Udio filing breakdown<\/a> at The Verge.<\/p>\n Any enterprise that has ingested third-party content to train or fine-tune an internal AI model should feel the temperature of this case. The liability pattern here isn’t about generating infringing output. It’s about what went into the model in the first place. Sony’s audio fingerprinting approach demonstrates that plaintiffs can now audit training corpora with technical precision, and Udio’s own documentation became its worst exhibit. If your data governance program doesn’t know exactly what your models were trained on, that gap is a legal exposure, not just a compliance gap.<\/p>\n The settlement dynamic is worth noting because it cuts both ways. Universal and Warner both settled with Udio and moved to licensing partnerships, which suggests the litigation isn’t purely punitive but is functioning as a forced negotiation. Sony’s decision to keep fighting and expand its claim signals that not every major rights holder will accept that trade. For any enterprise that licensed, scraped, or tolerated ambiguity in training data sourcing, the practical question is whether your vendor has a clean settlement or an open war, because the two situations carry very different indemnification postures.<\/p>\n The falsification condition here is clear: if courts consistently rule that audio fingerprinting evidence from post-hoc training audits is inadmissible or too speculative to support per-work damages, this litigation strategy collapses. But if Sony wins even a fraction of the 30,000-work claim at the statutory maximum, it sets a precedent that makes unsettled training data liability an existential balance-sheet risk for any AI vendor without documented licensing. The budget item your procurement team hasn’t priced yet is retroactive indemnification from the AI vendors whose models touch your workflows.<\/p>\n Audio fingerprinting converts a sound recording into a compact digital signature, the acoustic equivalent of a hash, that can be matched against other audio even when the output sounds different on the surface. Platforms like Shazam use it for song identification; Sony is using it forensically to show that specific recordings appear in Udio’s training data. For enterprise AI buyers, the business relevance is that the same technique can, in principle, be applied to text, code, or images, meaning training data provenance is now technically auditable by opposing counsel.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Audio fingerprinting<\/h2>\n