Music

Udio admits in court it scraped audio to train its AI music generator

The startup denied Sony Music's copyright claims but conceded that its training data came from "publicly available sources" — language the labels read as YouTube. UMG and Warner already settled. Sony is the last major holdout, and the fair-use ruling that follows could set the rules for every AI music tool on the market.
Alice Lange

In an answer filed in the Southern District of New York, the AI music startup Udio formally conceded the basic fact at the heart of one of the music industry’s biggest lawsuits: its models were trained on audio it scraped, not audio it licensed. The filing denies that this constitutes copyright infringement and asks the court to dismiss Sony Music’s claims with prejudice. But it concedes the activity itself. Sony, joined by Arista Music and Arista Records, is now the only major label still in court against an AI music generator, after Universal Music Group and Warner Music Group both quietly settled and signed licensing deals.

Specifically, Udio’s filing acknowledges that its system was built by feeding it “a vast amount of different kinds of sound recordings” gathered from “publicly available sources.” From those recordings, the company’s lawyers argue, the model derived “a complex collection of statistical insights about the auditory characteristics” — language designed to frame the use as transformative rather than derivative. The labels’ position, including in the amended complaint they filed last autumn, is that Udio specifically scraped copyrighted music from YouTube using tools like yt-dlp, and that doing so violated both the Copyright Act and the Digital Millennium Copyright Act’s anti-circumvention provisions. Udio’s filing does not name YouTube. Sony’s lawyers will argue that the gap between “publicly available sources” and YouTube is rhetorical, not factual.

The Sony fair-use ruling, expected this summer

The settlements that already happened are part of why the Sony case matters. Universal traded its lawsuit last autumn for an equity-and-licensing arrangement in a forthcoming joint AI music platform with Udio, with opt-in artist compensation. Warner did the same with Suno, Udio’s main rival, a month later. Both deals make the participating labels co-owners of licensed AI walled gardens rather than litigants against unlicensed ones. Sony chose differently. By staying in court, Sony is betting that a federal ruling in its favor will become the precedent that governs whether unlicensed AI music tools can legally exist at all. That precedent matters less to Universal and Warner now than it would have a year ago, since their commercial future is in the licensed alternative they helped build.

The legal question turns on the Anthropic precedent. A federal court ruled in a parallel case last year that the AI company’s unauthorized training on copyrighted books counted as fair use under US copyright law, but that its downloading of those books from online pirate libraries did not. That distinction, between training and acquiring, is exactly what Sony’s lawyers will press. If YouTube counts as “publicly available” in the same way a public library is, Udio’s defense gets stronger. If pulling audio from YouTube counts as circumvention of the platform’s technical protections, the labels’ DMCA argument carries. The judge has to decide which one the situation actually is.

The skepticism layer

This is not quite the smoking gun the headlines suggested. Udio’s admission was already implicit in the existence of its product. You cannot train a music model that produces convincing imitations of the Temptations and Mariah Carey, as the labels have repeatedly demonstrated, without having heard them in some form. What the filing changes is the legal posture. Previously, Udio could keep the question of how it acquired training data ambiguous. Now it cannot. But the ruling that follows will not undo what already happened with UMG and Warner. Two of the three major labels have already chosen to monetize the AI rather than try to kill it. The Sony case will determine what unlicensed AI music generators are allowed to do going forward, but the industry’s commercial answer is mostly already written. The licensed walled garden is the future. The lawsuit decides what survives outside of it.

Damages, if Sony wins, could be punitive. The labels are seeking up to $150,000 per work plus $2,500 per circumvention. Multiplied by however many tracks are alleged to be in Udio’s training data, the number gets very large very quickly. In practice, a Sony win likely produces a settlement before the damages phase, on terms that look more like UMG’s and Warner’s than like a death sentence for Udio. The most likely outcome, regardless of how the judge rules, is that Udio ends up with a license. The question is who writes the terms.

Udio filed its answer to Sony’s amended complaint on April 29, 2026. The fair-use ruling that the parallel Sony cases against Udio and Suno are expected to produce is widely projected for the summer of 2026 — th

Discussion

There are 0 comments.