Buying a music model on training provenance

For two years the way to pick an AI music model was to generate the same brief on each one and listen. That still works, and it is still the first thing to do. It is no longer the thing that decides the purchase for anyone shipping client work, because the models converged on "good enough for a bed" and diverged on something you cannot hear at all: where the training data came from.

Two positions define the current field. ElevenLabs built its music model on licensing agreements with Merlin and Kobalt, reported by Billboard, which means the corpus question was answered before the product launched. Suno's licensing is arriving through a settlement track instead: Warner in November 2025, then BMG in August 2026. As of mid-August 2026, the label-partnered model that track points toward has not shipped, and V5.5 — released in March 2026, on the same V5 line the litigation concerns — is still what answers a generation request.

Those are two genuinely different postures, not two points on a spectrum, and they produce different answers when someone in procurement asks where the audio came from.

Three provenance postures

Posture	What it looks like	What you can say when asked
Documented licensed corpus	Named rightsholder agreements in place before the model shipped; vendor will state what it trained on	"The model was trained under agreements with named rightsholders, here they are"
Settlement track	Litigation being resolved deal by deal; licensed model announced but not yet the one serving requests	"Disputes are being settled, and the model that produced this file predates those settlements"
Unstated	No published training-data disclosure at all	"We don't know"

The third bucket is larger than most buyers expect, and it is worth checking for explicitly rather than assuming a vendor's silence means the answer is fine. A vendor that has done licensing deals says so, loudly, because it is the most expensive thing it has bought.

Two clarifications that keep this honest. First, a settlement track is not a bad posture, it is an unfinished one, and vendors on it are moving in the right direction with real money behind it. Second, "documented" refers to what the vendor has disclosed, not to an audit you have performed. Nobody outside these companies has inspected a training corpus. You are buying a disclosure and the vendor's willingness to stand behind it.

For the output-quality half of the same decision, the head-to-head is in Suno V5 vs ElevenMusic, and the wider field including instrumental-only options is in best AI music generators.

What provenance buys, and what it does not

It is easy to oversell this. Provenance is a probability instrument, not a shield.

What it buys:

• A lower chance of a claim. A model trained on licensed catalogue is less likely to reproduce a recognisable fragment of something a rightsholder is actively policing.
• An answer to the question. Half of what a legal review wants is not a guarantee, it is a defensible account of how the asset came to exist. "We don't know" is what escalates.
• Procurement survivability. Enterprise vendor forms have started asking about training data directly, in the same section as data residency and retention. A vendor that cannot answer fails the form regardless of output quality. The general version of that questionnaire is in security and data questions for AI content tools.

What it does not buy:

• Indemnity. Unless the contract says the vendor will defend and cover you, licensed training data is a fact about the model, not a promise to you. Read for the indemnity clause separately; it is the clause with actual money attached.
• Freedom from output-side infringement. A model with a clean corpus can still generate something that resembles an existing work, especially when you prompt it toward one. Provenance is about the input, not the output.
• A grant you do not have. The vendor's licence to train is upstream of your licence to use. You still need the right plan terms covering your actual distribution, which is a separate check, mapped in the AI music licensing guide.
• Coverage for personality rights. Voice-likeness is its own regime and clean training data says nothing about whether a vocal resembles a specific identifiable person.

The question to put to the client's legal team

Do not ask "is AI music allowed?" You will get "it depends," which costs a week. Ask a question that forces a decision they can actually make:

For this campaign, what is the minimum standard we must meet on generated audio: (a) a vendor with published, named licensing agreements covering its training data; (b) a vendor with a documented licensing programme underway; or (c) no requirement, provided the plan terms grant commercial use for our media plan? And do you require contractual indemnity from the vendor as a separate condition?

That is answerable in one reply. It gives you a tier to buy against and it separates the two things that get conflated: training provenance and contractual indemnity.

Four follow-ups worth sending in the same message, because each is a yes or no:

1. Does the media plan include broadcast, cinema, in-store or DSP distribution? These are often outside a standard commercial grant and need checking line by line.
2. Do we need to keep the generation record, and for how long? Model, version, date, account, prompt. Assume yes and agree a retention period.
3. Does the disclosure obligation sit with us or the platform? Clearance and disclosure are separate duties; satisfying one does nothing for the other.
4. Is a cloned or artist-adjacent vocal ever acceptable on this account? Usually a flat no, and it is much better to have that in writing before someone generates one.

Running the comparison in an afternoon

Provenance narrows the shortlist. It does not pick the track. Once you know which tier you are buying in, the practical comparison is short:

1. Write one brief and freeze it. Same genre, same tempo, same length, same instrumentation, same structural instruction. Every word you vary between candidates invalidates the comparison.
2. Generate three variations per model, not one. A single generation tells you about a roll of the dice. Three tells you about the model's centre of gravity.
3. Score against picture, not in a player. Music that sounds good alone and wrong under footage is the standard failure mode, and it is invisible on a waveform.
4. Score the paperwork on the same sheet. Provenance posture, whether the plan grant covers your distribution, whether indemnity exists. Give it the same weight as the audio, because it is the criterion that can veto the choice.
5. Record everything as you go. The record you take during evaluation is the template for the record you keep in production.

Inside Versely, the music side of the catalog runs on the Suno Sounds line, and each model publishes what it is and what it costs on its own page: Suno Sounds V5.5 is listed at 2 credits a generation with looping, tempo and key control plus lyrics capture. That is the right shape for beds, loops and sound design, where volume is high and exposure is low. The AI music generator is the entry point, and every generation you run through the workspace lands in the same library as your video and voice assets, which is what makes the generation record a by-product rather than a chore.

Where ElevenLabs sits in a wider audio stack, including its dubbing line, is covered in ElevenLabs Music v2 and Dubbing v2. If you want to see how any two catalog models line up side by side on specs and cost, the comparison hub does that pairwise.

FAQ

Is a model with licensed training data automatically safe for advertising?

No. It improves the input-side position and answers the procurement question, but you still need the plan-level grant that covers your specific distribution, and you still need to check that the output does not resemble an existing work. Provenance is one of three or four checks, not a replacement for the rest.

How do I verify a vendor's training-data claim?

You cannot, independently. What you can do is capture the claim: the vendor's published statement, dated, saved as a file, alongside the terms that were in force when you generated. That converts a marketing claim into a documented representation you relied on, which is a materially different position from having read it once on a website.

Does this matter for a fifteen-second bed under a social post?

Much less. Exposure drives the standard. Organic social with a generated bed carries different risk from a national campaign with a media budget behind it, and applying the same procurement bar to both wastes time you should be spending on the campaign. Pick the tier per use, not per company.

Will the settlement-track vendors catch up?

Probably, and the deals point that way. The thing to watch is not the next announcement, it is the product cutover: the day the generation endpoint starts calling weights trained under those licences, and the day the terms are updated to reflect it. Until both happen, an announcement changes the trajectory rather than the file you just generated.