The dispute between AI music generators and the record industry has been fought mostly in courtrooms, over whether training a model on copyrighted songs is fair use or theft. On September 9, Suno answered that question with a change to the way its newest model was built.
Suno released version 6 of its music-generation model and said explicitly that this version no longer trains on the catalog data used for its earlier models. Instead, the company said, v6 is built on licensed music, a reversal tied directly to its legal battles.
The company is facing multiple copyright lawsuits from major record labels, which argue that Suno and its rivals trained their models on vast collections of copyrighted recordings without permission. Those cases are among the highest-stakes disputes in the AI industry, and their outcome could reshape how every generative model is built.
The shift to licensed data is a practical acknowledgment of the pressure those suits create. If Suno can point to a model built only on music it paid for, it strips the labels of their central complaint about v6, even if the earlier models remain in dispute.
The move also reshapes the economics of the business. Licensing music costs money, and the terms of those licenses are not public. Suno will have to price its subscriptions to cover the cost, or accept thinner margins than a model trained on freely scraped data.
In August, Suno said it would begin watermarking the songs its models generate, another concession to an industry worried about a flood of machine-made music diluting the market. The watermarking and the licensing move together suggest a company trying to turn itself into a lawful partner rather than a defendant.
The broader AI music sector is watching closely. Udio, another AI music startup, faces similar lawsuits, and the outcome of the label cases could determine whether the entire category survives in its current form or must be rebuilt on licensed data.
The record labels have not said whether v6 changes their view. Their lawsuits target the earlier models and the practices behind them, and a new model built on licensed data does not erase the claims about what came before.
For musicians, the shift offers a measure of what they have demanded for years: compensation and control over how their work is used. If AI models must license the music they learn from, songwriters and performers get a seat at the table instead of watching their catalogs absorbed without consent.
The question is whether licensing will scale. Training a capable music model on a sufficiently broad licensed catalog is expensive and requires the cooperation of rights holders who have been reluctant to engage. Suno’s v6 will test whether a licensed model can match the quality of one trained on everything available.
The company has not disclosed which catalogs it licensed or on what terms, leaving open the question of how much music v6 actually learned from. A model trained on a small licensed library may sound thinner than its predecessors, and users will notice.
The stakes extend beyond music. The same legal theories at play in the Suno cases are being argued over text, image and video models across the AI industry, and a settlement or ruling in one sector could shape all the others.
Suno’s v6 is best read as a hedge: if the courts rule against AI companies, Suno wants a product already built on the terms the courts are likely to impose. If the courts rule the other way, the company will have paid for data it could have scraped for free.
The legal fight has been running for more than a year. The major labels, led by the Recording Industry Association of America, have accused AI music companies of building their products on the industry’s catalog without consent or payment, and they have sought damages that could run into the billions. Suno has denied wrongdoing, arguing that training on publicly available music is lawful.
Some corners of the industry have already chosen negotiation over litigation. A number of labels and publishers have struck licensing deals with AI companies, selling access to their catalogs under terms that give rights holders a cut. Suno’s v6 appears to put it in that camp, at least for its newest model.
What happens next depends on the courts as much as on Suno. A ruling that training on copyrighted music requires a license would force every AI music company to follow Suno’s lead; a ruling the other way would leave the company having paid for data its rivals still use free. Both outcomes are on the table, and the labels are pressing for the first.
Either way, the days of training music models on whatever the internet offers, without asking, appear to be ending. Suno just became the first major player to say so out loud.
