What is still litigating in AI music
Universal and Warner settled with Udio, Warner with Suno, Sony with neither. UMG v. Suno continues and no merits ruling on AI music training exists.
The mental map most creative teams carry on AI music is one settlement behind and at least one plaintiff short. The shorthand has collapsed into "the labels settled, it's fine now," and every part of that sentence needs qualifying.
Here is the state of play as of August 2026, with the parts that are actually resolved separated from the parts that are not.
Who has settled with whom
| Rightsholder | Udio | Suno |
|---|---|---|
| Universal Music Group | Settled, October 2025 | Still litigating |
| Warner Music Group | Settled, November 2025 | Settled, November 2025 |
| Sony Music | Not settled | Not settled |
Three readings follow directly from that grid.
Sony has settled with neither. This is the single most commonly missed fact in the whole area, and it is the one most likely to matter, because a rightsholder without a settlement is a rightsholder whose claims are intact.
UMG v. Suno continues, with Suno defending on fair use. That is a live dispute between the largest rightsholder and one of the two main defendants.
The settlements were paired with forward licences for opt-in artist catalogues. That is genuinely significant for where the market goes. It is not a ruling, and it is not retroactive cover for anyone who is not a party.
The lyrics side is separate and less watched: Concord, UMPG and ABKCO v. Anthropic comprises two live suits over song lyrics, with no settlement. Music publishing rights are a distinct body of claims from recording rights, and they have not moved with the label deals.
Why a settlement is not a ruling
A settlement ends a dispute between two companies. It does not decide the legal question, it does not bind anyone who was not a party, and it creates no precedent for the next case. Courts do not cite settlements.
There is a second gap that matters more operationally. A signature does not change the weights that produced your file. Even where a deal genuinely leads to a licensed model, the sequence runs: agreement signed, model trained under it, product cuts over to the new endpoint, terms updated to reflect the new grant. The reported event is step one. The steps that change your position are three and four, and they are the least visible from outside. That progression is worked through in detail in a signed label deal is not a cleared model.
So the correct statement is uncomfortable but short: no merits ruling on AI music training exists. Nobody has one to point at, in either direction.
The nearest thing to precedent is not a music case
Because the music matters have settled or are pending, the doctrine is being made elsewhere, on text and images. Three matters carry most of the weight, and none of them is binding appellate law yet.
Bartz v. Anthropic (N.D. Cal.) held in June 2025 that training on lawfully acquired books was fair use, and that retaining a pirated library was not. The class settlement received final approval on 20 July 2026. The distinction is the useful part: the largest sums so far in this field were paid over how material was acquired, not over what was done with it in training. Settlements still create no precedent.
Kadrey v. Meta (N.D. Cal.) granted summary judgment for Meta on training in June 2025, expressly on a thin record, with the court signalling that market-dilution theories could succeed if properly evidenced. An interlocutory appeal was denied in July 2026, and the torrenting and distribution phase remains unresolved with a hearing set for 25 February 2027. A defendant win that reads as a comment on the evidence rather than a rule.
Thomson Reuters v. Ross (D. Del.) went the other way in February 2025, holding that a non-generative legal research tool acting as a direct market substitute was not fair use. It is on appeal, the Third Circuit heard argument on 11 June 2026, and the decision is pending. This will be the first US appellate ruling on AI training and fair use. Until it lands there is no controlling answer, which is why the confident takes in both directions are worth discounting.
In the UK, Getty v. Stability AI ended in November 2025 with Getty dropping the training and output claims mid-trial for want of territorial evidence, the secondary-infringement claim failing, and an extremely limited trade mark win on watermarks. UK training legality was never decided. It is not a precedent for either side, however it was headlined.
The dispute that opened downstream
One development is easy to miss because it does not involve an AI company at all. In June 2026 the American Federation of Musicians sued Universal and Warner over the allocation of settlement proceeds to session musicians.
That is a fight about who gets paid out of money already agreed, and it points at the structural question sitting under the licensing deals: catalogue owners are settling and licensing, and the performers on those recordings are asking where their share sits. Anyone planning to rely on "licensed AI music" as a category over the next few years should watch how that allocation question resolves, because it shapes what an opt-in catalogue licence is actually worth and how willing artists are to opt in at all. The case for treating training provenance as a purchasing criterion rather than a marketing line is in licensed training data as a buying criterion.
What this means for a track you generate this week
The legal weather is improving and the legal position is unchanged. Both are true, and the operational answer is to sort work by exposure rather than to wait for a ruling that may be a year away.
Low exposure: generate freely. Beds, loops, sound design, internal video, pitch material, organic social. The volume is high, the scrutiny is low, and the downside of a problem is re-scoring a clip. This is most of what most teams need, and the AI music generator covers it.
High exposure: answer three questions first. Paid media, broadcast, DSP distribution, anything with a budget behind it.
- Which model made this file, and when? Named model, version, date, recorded at generation time. A licensing announcement in August does nothing for a track generated in June, and audio carries no reliable statement of what produced it. Every model in the catalog has a page naming what it is, so Suno Sounds V5.5 is a dated entry in your record rather than "the music tool."
- Does the grant cover this specific placement? Commercial use in vendor terms is routinely narrower than a media plan assumes, and paid social, broadcast, in-store, cinema and DSP delivery are frequently separate grants. The AI music licensing guide maps the usual shapes.
- Is there a recognisable voice on it? If a listener would name the singer, that is a publicity-rights question that no settlement between labels and AI companies touches, and it is the fastest-moving risk in the field.
And know that clearance is not the only cost. Platforms have largely converged on labelling AI music rather than removing it, and at least one streaming service excludes tagged tracks from editorial and algorithmic playlists. That is an economic penalty applied to lawful content, and it belongs in the plan even when the legal question is clean. The three different platform responses are separated out in tagged, not removed.
For comparing the tools themselves rather than their legal posture, Suno vs Udio vs Stable Audio is the head-to-head. None of this is legal advice.
FAQ
If Warner settled with Suno, can I use Suno commercially?
Your right to use the output comes from your plan's terms, not from Warner's settlement. The settlement resolved a dispute between two companies. Check the grant in your own agreement, confirm it covers the specific placement, and keep the record of which model and version produced the file. Those are the things that answer the question.
Which rightsholder is the outstanding risk?
Sony Music has settled with neither Suno nor Udio, and UMG's case against Suno continues. Those are the intact claims. Publishing is separate again, with the lyrics suits against Anthropic still live and unsettled, which is a reminder that recording rights and composition rights move independently.
When will there be an actual ruling?
The nearest date worth watching is the Third Circuit's decision in Thomson Reuters v. Ross, argued 11 June 2026 and pending. It is not a music case and it concerns a non-generative tool, so it will not settle AI music directly, but it will be the first appellate treatment of training and fair use in the US. Beyond that, the Kadrey distribution phase has a hearing set for 25 February 2027.
Should we just avoid AI music until this resolves?
That is a defensible policy for paid campaigns and an overcorrection everywhere else. A generated bed under an organic post carries a different risk profile from the same bed under a media buy, because exposure, scrutiny and the cost of a takedown all differ. Plenty of teams run generated music everywhere except paid, and that is a coherent position rather than a fudge.