Suno, Udio, and the Lawsuit Over How AI Learns Music

This article explains the two lawsuits most likely to decide how artificial intelligence is allowed to learn from recorded music in the United States. It is informational only. It is not legal advice, and it creates no attorney client relationship.

In June 2024 the major record labels stopped warning about AI music and started litigating it. Coordinated by the Recording Industry Association of America, they filed two parallel suits against the two best known AI song generators. In UMG Recordings, Inc. v. Suno, Inc., No. 1:24-cv-11611 (D. Mass.), and in UMG Recordings, Inc. v. Uncharted Labs, Inc., the developer of Udio, No. 1:24-cv-04777 (S.D.N.Y.), the labels allege copyright infringement on a scale the recording industry has not tested in court before.

The core allegation

The complaints do not focus on the songs the tools produce. They focus on the songs that went in. The labels allege that the developers copied enormous libraries of copyrighted sound recordings, without a license and without permission, in order to train their models. To support the claim, the complaints point to outputs that allegedly echo specific famous recordings closely enough to suggest the originals sat in the training data. In plain terms, the labels argue you cannot build a machine that convincingly generates music in the manner of protected recordings unless you first copied a great many protected recordings to teach it.

That framing matters. It places the act of copying for training at the center of the case, which is the precise question courts across the country are now wrestling with in every corner of the AI industry.

The fair use battleground

The developers lean on fair use, the doctrine that permits some unlicensed copying. Courts weigh four factors: the purpose and character of the use, the nature of the original work, how much was taken, and the effect on the market for the original. The developers argue their use is transformative, that a model learns statistical patterns rather than storing songs, and that this sits closer to how a human musician studies records than to piracy.

The labels aim straight at the fourth factor, market harm, and they have a recent decision to build on. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025), a federal court held that copying protected material to train an AI research tool was not fair use, leaning heavily on the harm to the market for the originals. That case did not involve a system that generates new expression, so its reach is debated, but it handed rights holders a template, and they are using it.

The picture is genuinely mixed, which is why confident predictions deserve suspicion. In the parallel litigation over training large language models on books, a California federal court found that training on lawfully acquired copies could be transformative fair use, while separately holding the company liable for assembling a library of pirated copies. That company, Anthropic, later agreed to a settlement reported at roughly 1.5 billion dollars. Those cases involve text, not music, and the music suits will rise or fall on their own factual records, but they show the doctrine cutting in more than one direction at once.

Why the stakes are close to existential

Copyright law allows statutory damages of up to 150,000 dollars per work infringed when the infringement is willful. Multiply that ceiling by the number of recordings a large model may have ingested and the arithmetic stops being ordinary litigation and starts being a threat to a company's continued existence. That math is not incidental. It is leverage. It is a large part of why many observers expect these cases to pressure the developers toward licensing deals or settlements rather than a final trial verdict, in the way the Anthropic matter resolved with a payment rather than a definitive rule.

A settlement would answer the practical question, whether AI music companies must pay to train on catalogs, without answering the legal one, whether the law required them to. A ruling, by contrast, could set a precedent that reaches far beyond these two companies and reprices the raw material of an entire industry.

What to watch as the cases move

Through early 2026 both suits remained in active litigation, grinding through the slow and expensive machinery of discovery, with no merits ruling on the central fair use question. Three things are worth watching. First, whether the courts let the labels inspect the training data directly, because what that data contains may decide the case. Second, how each court treats the transformative use argument in the specific setting of a tool built to generate music, as opposed to text or research results. Third, whether either side blinks and settles, which would shape the market through private contracts rather than public law.

For an independent artist, the practical takeaway is patience paired with attention. Whether these models were built lawfully is unresolved and moving month to month. Building a business on the confident assumption that training is clearly legal, or clearly illegal, is a gamble in either direction. This remains general information, not legal advice, and the law here is being written in real time. For your own situation, consult a licensed attorney.