AI Music Got a License. Read What It Actually Covers.

For two years the story around AI music was lawsuits. Labels sued the generators, artists sued the labels, and everyone waited for a court to decide whether training on songs was allowed. That phase is ending, and what replaces it is less dramatic and more consequential: a set of license terms that describe what you are actually allowed to do with a generated track.
The clearest marker is Suno. In September it launched v6, and by Music Business Worldwide's account the model was trained from scratch on licensed music from Warner Music Group, BMG and Believe. That is a structural change. An earlier generation of these models learned from whatever was online. A model trained on licensed catalogs has a defensible supply chain, and it also inherits the obligations that come with one.
The terms changed in September
Suno's current terms took effect on September 3 and redrew the line between free and paid. Free users get a handful of lifetime trial downloads with no commercial rights. Pro and Premier subscribers get monthly download caps, and the company assigns its rights in what they generate to them. There is one sentence worth reading twice: Suno says the rights it grants continue after you cancel, but it does not promise that copyright will exist in the output.
That last point is the one people miss. A license to use a song is a contract between you and the tool. Copyright is a property right the government grants. Suno can license you the track. It cannot manufacture a copyright in it, and the US Copyright Office has said plainly that copyright does not extend to purely AI-generated material, or to material where there is insufficient human control over the expressive elements. Prompts alone, in its reading, do not make you the author.

Practically, that means the melody Suno generates is yours to use because Suno said so, not because you own it. If someone else reproduces it, your remedy is thin. What you can claim is the part you wrote. If you wrote the lyrics, they are your expression, and a competitor who copies the words is exposed in a way that a competitor who copies the tune is not. The obvious advice follows from that: write your own lyrics.
Everyone else is moving in different directions
Udio, the other name in this space, took a different route. It signed licensing deals with Universal Music Group and Warner Music Group, and while it rebuilds as a licensed platform its terms restrict output to personal, non-commercial use and downloads are switched off entirely. Strong output, no practical way to publish it. If your plan involves a client or a monetized channel, Udio is not a tool you can use right now.
Google's Lyria 3.5 comes from the other end of the industry. It generates songs with vocals and timed lyrics at 44.1 kHz stereo, takes your own lyrics with section tags, and stamps every output with a SynthID audio watermark. The Clip variant always returns 30 seconds, which happens to be the standard length of a broadcast ad. What Google is selling is closer to a jingle than a song, with a provenance marker baked in.
ElevenLabs Music has the most legible terms of the group. Self-serve plans allow commercial use online and offline, with one carve-out: film, TV, radio and studio games are excluded and require an enterprise agreement. Read that against the tools' typical buyers, and the product is aimed at social and online advertising, where a licensed track from a subscription is exactly what a small team needs.
MiniMax Music 3.0 closed its paid music API to new users from August 20 and published open weights instead, which shifts the question from licensing to hosting. If you run it yourself, the terms you answer to are the model's license, not a subscription.
Why this matters more for ads than for albums
Music is where AI rights get tangled fastest, because the output sits close to the input. A language model that summarizes a book is a step removed from the text. A song in the style of a specific artist invites an immediate comparison, and courts have been less willing to accept a transformation argument for audio than for text. Suno lost a ruling in Germany brought by GEMA, the performing rights society, over the use of protected recordings and compositions. That is the case to watch, because an injunction stops a service rather than taxing it, and a stopped service has no terms at all.
For anyone producing advertising, the licensing question has become the first question, ahead of which model sounds best. A jingle used to need a writer, a singer and a studio session. Now a model sings the lyrics you typed, and the hard decisions move from the recording to the rights. Some of that pressure is commercial rather than artistic. Research from Ipsos on 2,015 US video ads found that ads carrying a sonic brand cue were far more likely to land in the top third for branded attention than in the bottom third. That is a correlation measured on attention rather than a promise about sales, and it explains why brands keep paying for a distinctive sound even when a generator can produce something passable in seconds.
The label deals are the real signal in all of this. When the largest rights holders license their catalogs to a generator, they are setting the terms on which it can operate rather than blessing the technology, and those terms then become the baseline for every smaller catalogue owner who follows. That is how the sampling and streaming fights ended two decades ago, and the shape looks familiar: a period of open infringement, a wave of litigation, then a licensing market where the incumbent rights holders collect and the tool vendors absorb the cost.
For a buyer, the consequence is that music is becoming a line item with a source attached. Cheap tracks still exist. What is disappearing is the assumption that a track generated from a prompt is free of claims, which was never true and is now demonstrably false.
What to do before you publish
Three habits cover most of the risk. Check the terms on the day you generate, because nearly every tool in this space rewrote them during 2026. Keep a record of the plan you were on, the date, and how you downloaded the file, since a dispute about rights usually turns on which terms applied when. And treat free tiers as experiments, because none of the major tools grant commercial rights for free, and a track generated on a free plan does not become licensed retroactively when you upgrade.
The bigger picture is that AI music has moved from a legal question to a procurement one. The lawsuits are not over, and the models are still trained in ways artists dispute. But the tools now publish terms, the terms differ sharply by vendor, and the difference between them is the difference between a song you can put in a client's ad and a song you can only play for yourself. Reading the fine print has become the job.
Related articles
The Memory Chip Is Now the Bottleneck in AI Compute
The logic roadmap is public and predictable. The memory roadmap is gated by packaging yields, a years-long talent base and three companies' capital plans.
Hengdian's Studios Are Empty While Asia's AI Film Pipeline Runs
Hengdian's empty stages are the evidence that the industry has already placed its bet, whether or not the audience agrees.
Google Bought 890 Megawatts of Nuclear to Feed Its Data Centers
The compute is available. The electricity is the thing that has to be arranged, years in advance, the way a supply chain is arranged.
Meta Settled the Book Lawsuit. Permission Now Has a Price.
Meta settling removes one landmark from the map. It does not draw the rest of it, and the next few settlements will decide whether permission becomes a market or stays a series of exceptions.