A signed label deal is not a cleared model
Suno added BMG in August, but the licensed model has not shipped and V5 still powers the product. What must be true before an AI track runs in a paid ad.
Digital Music News reported a Suno deal with BMG on 12 August 2026, adding another rightsholder to the track that started with the Warner settlement on 25 November 2025. Inside a week the framing in creative-team Slack channels had already collapsed into "Suno is licensed now, we can use it in the campaign."
That is not what happened, and the gap is worth being precise about, because it is the kind of thing that gets found in a legal review three days before a media buy goes live. As of mid-August 2026, the label-partnered model those deals point toward has not shipped. The product still runs on V5.5, released in March 2026 and built on the same V5 line the litigation is about. A signature on a settlement is a change in the relationship between two companies. It is not a change in the weights that produced your file.
Four things "a deal" can mean, and only one of them helps you
The word does a lot of undifferentiated work in headlines. Pulled apart, there are four distinct events, and they land in a specific order:
| Event | What it changes | Does your output become cleared? |
|---|---|---|
| Litigation settlement | The dispute between the label and the AI company ends | No |
| Training licence | The AI company may lawfully train on that catalogue going forward | No, not until a model trained under it ships |
| Distribution or product partnership | Where and how the output can be commercially exploited by the partners | Only for the partners, not for you |
| Licence granted to you, in your plan terms | What you may do with your generated file | Yes, and only this one |
Almost every announcement you read is one of the first three. The fourth is not usually announced at all, because it lives in the vendor's terms of service and changes without a press release.
The practical reading: a rightsholder announcement tells you the legal weather is improving. It tells you nothing about the file sitting in your downloads folder.
The gap between a signature and a shipped model
Even in the best case, where a deal genuinely does lead to a licensed model, four things have to happen in sequence:
- The agreement is signed. This is the part that gets reported.
- A model is trained under it. New corpus, new run, new evaluation cycle. Months, not days.
- The product cuts over. The generation endpoint stops calling the old weights and starts calling the new ones. Until this happens, every generation is from the old model regardless of what was signed.
- The terms are updated to reflect it. The grant to you changes only when the contract you accepted changes.
Steps 3 and 4 are the ones that affect your risk, and they are the two that are least visible from outside. There is no reliable way to tell from a track's audio which model produced it, which is why the record you keep at generation time carries the whole burden.
Meanwhile, the streaming layer moved in a different direction entirely. Platforms have largely landed on labelling AI involvement rather than removing it, which changes the discovery consequences without touching the clearance question at all. That distinction is worked through in tagged, not removed.
Five things that must be true before an AI track goes into a paid ad
This is the checklist. It is deliberately boring, and every item is a thing you can actually verify rather than a thing you can feel confident about.
1. You can name the model that produced this specific file. Not the vendor. The model and version, recorded at the time of generation, along with the date. A vendor's licensing announcement in August does nothing for a track you generated in June, and if you cannot say when a file was made, you cannot say which model made it. In Versely, every model in the catalog has a page publishing what it is and what it costs, so Suno Sounds V5.5 is a named, dated thing in your record rather than "the music tool."
2. Your plan granted commercial rights at the moment of generation. Rights are almost never applied retroactively across an upgrade. A track made under one grant does not silently acquire a better one because you later paid for a better plan. If the campaign matters, generate under the terms you intend to ship under, from the start.
3. The grant actually covers this media buy. This is where most teams get caught. "Commercial use" in a vendor's terms is frequently narrower than a media plan assumes, and paid social, broadcast, in-store, cinema, and distribution to streaming DSPs are often separate grants with separate conditions. Read the specific clause against the specific line items on the plan. The general shape of these grants is mapped in the AI music licensing guide.
4. You hold the generation record. Prompt, model, version, date, account, and the output file's original hash if you can get it. Clearance disputes are resolved by paperwork, and the paperwork has to exist before the dispute does. This is ordinary asset governance, not a music-specific problem, and it belongs in the same system as everything else you ship, as in AI content governance for brands.
5. Disclosure is handled for every destination the ad runs on. Clearance and disclosure are different obligations and satisfying one does nothing for the other. Ad platform rules, and now regulatory ones, are their own checklist, laid out in AI ad disclosure compliance.
If any of the five is unknown rather than false, treat it as false. "We think it is probably fine" is the state that produces the awkward call.
What to do in the meantime
You have three defensible options while the licensed models are still in transit, and picking one deliberately is better than defaulting into whichever tool is open.
Use the model whose training provenance is already documented. ElevenLabs' music model is the one built on licensed catalogue deals with Merlin and Kobalt rather than on a settlement track, which is a different starting posture from "the dispute is being resolved." The full head-to-head on output quality and terms is in Suno V5 vs ElevenMusic.
Keep AI music on the organic side of the line. A generated bed under an organic social post carries a different risk profile from the same bed under a paid placement with a media budget behind it, because the exposure, the scrutiny and the cost of a takedown are all different. Plenty of teams run AI music everywhere except paid, and that is a coherent policy rather than a fudge.
Licence production music for the paid work. Unfashionable, still correct for a hero campaign. The cost of a production library track is small relative to a media buy, and the clearance question is answered by an invoice.
For the day-to-day work that sits below the paid line, the AI music generator covers beds, loops and sound design where the exposure is low and the volume is high. The judgement is about which side of the line a given track is on, not about whether AI music is usable at all. The broader version of that argument, across images and video as well as audio, is in the AI copyright and safety guide.
FAQ
If Suno has settled with Warner and BMG, what is still unresolved?
The settlements resolve those rightsholders' disputes and set up a licensing path. They do not retroactively change what the currently shipping model was trained on, and they do not cover rightsholders who have not signed. Until a model trained under those licences is the one actually answering your generation request, the output you get is from the same weights as before.
How would I even know which model made a track?
Only from your own record. Audio does not carry a reliable, tamper-proof statement of which model produced it, so the answer has to be captured at generation time: model name, version, date, account, prompt. Build that into the workflow and the question is trivial. Skip it and there is no way to reconstruct it later.
Does a paid plan make output safe to use in advertising?
It makes the grant from the vendor to you commercial, which is necessary and not sufficient. The training-data question sits underneath the grant, and a vendor cannot grant more than it holds. That is precisely why provenance and the plan-level grant are two separate items on the checklist above.
Is this an argument against using AI music at all?
No. It is an argument for matching the tool to the exposure. Beds, loops, sound design, internal work, organic social and pitch material are all low-exposure contexts where generated audio is straightforwardly useful. A track carrying a seven-figure media buy is a different decision, and it should be made with the five items above answered rather than assumed.