AI music has moved from novelty to legal test case. Two startups that can generate complete songs from a prompt in seconds are now facing major record labels in court, while YouTube is reportedly exploring licenses that could reshape how music models are trained.
The dispute is not only about one set of tools. It may decide whether AI music companies can build powerful models without paying for recorded music, or whether access to catalogues becomes the central cost of competing.
The lawsuits put training data at the center
On June 24, Sony Music, Warner Music Group, and Universal Music Group sued Suno and Udio. The labels claim the companies used copyrighted music in training data “at an almost unimaginable scale” and that the models can generate songs that “imitate the qualities of genuine human sound recordings.”
Suno and Udio have both responded by pointing to efforts to keep their systems from copying protected works. Udio said its model “has ‘listened’ to and learned from a large collection of recorded music.” Two weeks before the lawsuits, Suno CEO Mikey Shulman said the company’s training set is “both industry standard and legal,” while also saying the exact recipe is proprietary.
Neither company has specified whether copyrighted works are in its training data. That silence matters because the record labels are arguing not only about what the models produce, but also about what they may have learned from.
Why music is a harder AI battleground
Training-data disputes have become familiar across generative AI. Companies have faced legal pressure over books, news articles, images, and other copyrighted material, and some have responded with licensing deals while cases continue.
Music is different in several practical ways. The public domain is much narrower for the kind of music most listeners expect, so an AI company has fewer useful alternatives if it cannot train on modern catalogues. The source article also notes that rights in music are more concentrated than in film, images, or text.
The three biggest record labels, whose publishing arms collectively own more than 10 million songs, control much of the music that shaped the last century. The filing names artists the labels allege were wrongly included in training data, ranging from ABBA to those on the Hamilton soundtrack.
There is also the creative challenge. A readable poem or passable illustration is one kind of technical problem. Producing music people want to hear asks a model to absorb taste, structure, sound, performance, and genre cues in a way that can be difficult to separate from the recordings used to train it.
The output claims may be just as important
The record labels allege infringement on both the training side and the output side. In other words, they argue that Suno and Udio learned from copyrighted recordings and can generate music too close to protected works.
James Grimmelmann, a professor of digital and information law at Cornell Law School, said the plaintiffs have unusually strong odds against an AI company. He compared the case with the New York Times case against OpenAI, but said the claims against Suno and Udio are “worse for a bunch of reasons.”
One reason is plausibility. In a text case, a company might argue that copyrighted articles appeared inside a broad web scrape without specific intent. Grimmelmann said that defense is harder to make for commercial recordings, because “It’s pretty clear that they had to have been pulling in large databases of commercial recordings.”
The labels also claim simple prompts produced songs that resembled existing works. They cite a Udio prompt, “my tempting 1964 girl smokey sing hitsville soul pop,” which they say yielded a song that “any listener familiar with the Temptations would instantly recognize as resembling the copyrighted sound recording ‘My Girl.’” They also cite a Suno example called “Prancing Queen,” generated with the prompt “70s pop” and the lyrics for “Dancing Queen.”
Suno and Udio say they use safeguards. Shulman wrote that Suno is “designed to generate completely new outputs, not to memorize and regurgitate preexisting content” and said it does not allow prompts that reference specific artists. Udio said it uses “state-of-the-art filters to ensure our model does not reproduce copyrighted works or artists’ voices.”
The labels argue those protections have loopholes. The source describes examples in which artist-name filters blocked direct requests but could be tested with spaced-out names, though similar workarounds were blocked on Udio.
What the court could decide
Grimmelmann outlined three possible paths. The first would favor the AI startups: the lawsuits fail, and the court finds that training and outputs do not violate fair use or copy protected works too closely. If that happened, songwriters and rights holders would need another legal route to seek compensation.
The second outcome would be mixed. The court could find that training is allowed, but that the companies must better control outputs so models do not imitate protected works improperly. Grimmelmann compared this with an early Napster ruling that required the company to ban searches for copyrighted works in its libraries, even though users found workarounds.
The third outcome would be the most severe for AI music companies. The court could find problems with both training and outputs. That could mean the companies cannot train on copyrighted works without licenses, cannot allow close imitation of protected works, and may have to pay damages that could run into the hundreds of millions for each company.
Licensing could decide who survives
Two days after the lawsuits, the Financial Times reported that YouTube is taking a different route. Instead of relying on secret training sets, YouTube is reportedly offering unspecified lump sums to top record labels for licenses to use their catalogues in training.
If courts make free training on label catalogues legally risky, licensing may become the only realistic path for strong AI music models. That would favor companies with deep pockets, because the source article notes that music data is expensive, concentrated, and difficult to replace with public-domain material.
The licensing question is complicated by music copyright itself. There are two different copyrights at play: one for the song, covering composition such as music and lyrics, and one for the master, covering the recording. Some artists, including Taylor Swift and Frank Ocean, have come to own the masters of their catalogues after drawn-out legal battles. Many others retain only the song copyright while labels retain the masters.
Musician groups are divided on whether licensing AI is a path to compensation or a threat to future work. SAG-AFTRA adopted contract rules in April allowing AI clones of member voices, with minimum rates for compensation. In December, the Indie Musicians Caucus criticized the American Federation of Musicians, saying it would oppose any agreement “obligating AFM members to dig [their] own graves by participating—without a right to consent, compensation, or credit—in the training of our permanent Generative AI replacements.”
Kenneth Shirk, international secretary-treasurer at AFM, framed the tension directly: “We want musicians to get paid. But we also want to ensure that there’s a career in music to be had for those that are going to come after us.”
That is the core issue now facing AI music. The technology can already create songs quickly, but the legal and business foundation underneath it is unsettled. If licensing becomes mandatory, the future of AI music may be shaped less by model design alone and more by who can afford access to the catalogues that listeners recognize.