The case for letting labels license AI voice deals is real. The voice still needs the artist's own signature

A label can license a recording. Only the artist should be able to license their voice singing something new. The UMG and ElevenLabs deal is a good moment to draw that line, because its fan platform hasn't launched and its rules aren't written yet.

The news hook

On 10 Sep 2026 Universal Music Group and ElevenLabs announced a multi-year licensing agreement. The first product is a fan platform for remixes, mashups and "personalized vocal experiences", "built on licensed music and artist participation". ElevenLabs says artists will be "fairly compensated" and should "share in the value created". The post doesn't describe a consent process, and it says the platform "will be distinct" from ElevenMusic and the Music API. We covered the deal and ElevenMusic's terms in UMG signs ElevenLabs, and ElevenMusic adds downloads, and argued there that voice needs a per-artist opt-in. This column takes on the best argument against that.

The best case for letting the label decide

Put it as a label's digital team would. Licensed AI only beats unlicensed AI if it launches with a catalogue fans want. Ask thousands of artists to sign individually and the platform opens with a few willing names and a lot of gaps, while unlicensed tools offer everyone. Many record contracts already give the label rights to use an artist's name and likeness to promote releases, so the label isn't claiming something new. A single licence also gives the label the leverage to win a decent rate, which a mid-tier artist negotiating alone won't. And "participating" artists in the announcement may well mean artists who opted in; nothing published says otherwise.

That is a serious argument. On the recordings, we mostly agree with it.

Where it breaks: a voice isn't a catalogue

A remix rearranges something the artist already performed. A "personalized vocal experience" can make the artist perform something they never did. Those are different acts, and only the second one puts words in a person's mouth.

ElevenLabs' own terms show what happens when voice is folded into a blanket grant. On ElevenMusic, an uploader licenses their name, image, likeness and voice ("NILV") royalty-free, and the terms give as an example of an Adaptation "using an artist's NILV as depicted in Uploaded Music as the vocals for newly generated music." If the artist removes their music, "previous Adaptations using that Uploaded Music or NILV" may stay available and "will be unaffected." The licence "does not by itself entitle you to, any payout". The terms do bar ElevenLabs from commercialising a voice "on a standalone basis" without permission, which is a real limit. But it doesn't stop the voice being used inside other users' songs.

The UMG platform is formally separate, so none of that automatically applies. It is still the clearest evidence of the default this company writes when nobody negotiates voice separately.

The template already exists

The efficiency argument assumes per-artist consent has to be slow. The US Senate's NO FAKES Act of 2025 bill suggests otherwise. For a living person, it says a digital-replica licence is valid only if it "is in writing and signed by the individual or an authorized representative", includes "a reasonably specific description of the intended uses", and lasts no more than 10 years. A licence covered by a collective bargaining agreement that deals with digital replicas is exempt. The bill is still a bill, not law. As a drafting checklist, though, it's short: a signature, a scope, an end date. A label that already has a relationship with each artist can collect that alongside the catalogue deal. The bill doesn't say that a label licence counts as the artist's own signature, and we don't think it should.

ElevenLabs also already runs a consent-led model for famous voices. Its Iconic Marketplace, announced in November 2025 with Michael Caine and more than 25 other voices, has companies "request access" to a voice for projects and frames each collaboration as "rooted in respect, consent, and creative authenticity". It doesn't publish how approvals work, but the shape is per-voice and per-project, not blanket.

What we think

Split the deal in two. UMG can license recordings at catalogue level, and the efficiency argument wins there. Voice should work like the NO FAKES draft: the artist's own signature, a stated use, a time limit, with a union agreement as the alternative. Add the one thing ElevenMusic's terms get backwards: when an artist withdraws their voice, the fan-made outputs using it come down too.

That would cost the platform some names on day one. We think that's the right trade. A platform that launches with fewer voices, all of them willing, is a better argument for licensed AI than one where a fan can make any UMG artist sing anything. We also suspect that the artists most likely to say no are the ones whose voices are worth the most, which is exactly why the label would rather not ask. That last point is our reading of the incentives, not something either company has said.

What would change our mind: if UMG publishes consent rules showing each participating artist signs for voice separately and can pull outputs, the label-level deal is fine as it stands. Until then, "artist participation" is a phrase, not a policy.