Home/ The Signal/ Industry/ AI Music Generation Deals: Opt-In Is Only as Strong as Your Contract
Rights

AI Music Generation Deals: Opt-In Is Only as Strong as Your Contract

Opt-in does not mean you get to decide. For most artists and songwriters, the consent question in AI music generation will be answered by whoever holds the rights on paper — a label, a publisher, an…

A close-up photograph of a signed recording contract lying on a dark walnut conference…

Opt-in does not mean you get to decide.

For most artists and songwriters, the consent question in AI music generation will be answered by whoever holds the rights on paper — a label, a publisher, an estate, a catalogue fund — and the answer may reach you as a press release, at the same moment it reaches everyone else. That is not a forecast. It is how the paperwork already works, and the deals now being signed between rights holders and AI companies are being written on top of that paperwork, not instead of it.

A couple of years ago the industry story was litigation: infringement claims against model builders, discovery fights over what went into the training data. The story now is alliances. BMG and Suno have announced a global partnership covering a catalogue in the millions of songs, with artist choice and compensation stated as the organizing principles. As of this writing, specific products have not been announced. What has been announced is a shape — consent, payment, attribution, negotiated in a term sheet rather than argued in a courtroom.

That shape deserves to be taken seriously. It also deserves to be read the way a rights manager reads any term sheet, because the word doing the heaviest lifting is participation, and participation is a contractual position before it is an ethical one.

What a licensing deal actually changes

A license converts an unsettled legal question into a commercial one. Before: was ingesting this recording lawful? After: what is the rate, what is the term, and how is usage counted? That conversion is genuinely useful. Unlicensed training gives an artist a lawsuit they probably cannot afford to run; a license gives them a line item. Line items can be audited.

But a license negotiated at the catalogue level does not decide what any individual creator receives. That is decided by the chain underneath it — the recording agreement, the publishing agreement, the split sheet, the producer letter. A rights holder can sign a deal that is admirable in public and still route the money through a clause written in 2011 by someone who had never heard of a diffusion model. Both things can be true at once, and usually are.

Can an artist stop their music being used to train an AI model?

Usually not on their own, and not from the recording side. In a standard label deal the master is controlled by the label; in a standard publishing deal the composition is controlled by the publisher. If both have granted broad rights, the practical decision sits with them, and the artist's leverage is whatever their contract reserved — a consent right, an approval right, a carve-out, or nothing. Where artists tend to have independent standing is around name, voice, and likeness, which in many jurisdictions are treated separately from the recordings and are frequently the strongest lever an artist actually holds. None of this is legal advice, and it varies by territory and by contract; the point is where to look first.

Rights layer Who typically holds the switch What to check
Master recording Label, or the artist if unsigned/reverted Grant language, term, whether new-use consent survives
Composition Publisher, co-writers Whether administration includes novel-use licensing
Name, voice, likeness Usually the artist Explicit carve-out from any catalogue-level deal
Stems, session files, sample packs Whoever delivered them Work-for-hire language, buyout terms
Catalogue sold to a fund The fund Whether artist consent rights transferred with the assets

Six questions to ask before a deal gets announced for you

A short list to run against your own paperwork, and against any framework you are asked to endorse.

A photorealistic wide shot of a lone musician standing in a dim recording studio…
  • Does the grant cover training? Old contracts lean on "all media now known or hereafter devised." Whether that language reaches model training is contested, and it is the sentence your position rests on.
  • Is the income treated as a license or a sale? Many modern agreements split license income near evenly and pay sale income at the artist royalty rate after recoupment. Same money in, very different money out.
  • Per-use accounting or pooled pot? If revenue is a lump advance divided by catalogue market share, the largest rights holders absorb most of it regardless of what the model actually leaned on.
  • Is opt-in per work or per catalogue? Catalogue-level opt-in with a per-artist opt-out is a different instrument from work-level consent, and it defaults in the opposite direction.
  • What happens at the end of the term? Models are not meaningfully untrained. Ask what the deal says about outputs that already exist when the license lapses.
  • Do consent rights survive a sale? Catalogues change hands. A protection that does not travel with the asset is a protection with an expiry date nobody told you about.

What career impact actually looks like

Not replacement. Displacement, unevenly distributed, and it starts where the brief is functional rather than authored.

I have scored enough indie games and short films to know which of those briefs is fragile. The ninety-second loop that has to survive being heard four hundred times in one level. The two-minute tension bed under a documentary interview. The corporate explainer that needs something in D minor at 92 BPM that does not distract. Library and production-music budgets were already thin, and generative music tools land directly on that work, because the buyer's requirement was never a specific artist's voice — it was clearance, tempo, and a 48kHz WAV by Thursday.

Be equally honest about where the tools fall down. Dense arrangements still mush in the low mids. Vocals remain the hardest thing to get right, and a convincing performance — breath, timing that pushes and drags, a phrase that means something — is still where most renders give themselves away. Stem separation from a generated master rarely gives you clean, independently mixable parts. Prompt roulette is real; you get the take you wanted on the eleventh try, or not at all.

So the pressure is asymmetric. Work paid for by the minute is exposed. Work paid for because someone wanted your decisions — the wrong chord that turned out to be right, the revision after the director changed the cut — is exposed later and less. That is not consolation. It is a map of which contracts to renegotiate first.

What a framework that actually protects creators would contain

Judge any of these deals on five things. Consent recorded at the work level, not asserted at the catalogue level. Output-level accounting an artist can audit, closer to a mechanical statement than a revenue share. A published list of who is in, so participation is verifiable from outside. Name and voice handled as a separate grant with separate consent. And rights that survive the sale of the catalogue they were attached to.

None of the announced frameworks has demonstrated all five in production yet, because none of them has been running long enough to produce a statement anyone can check. That is the honest position: the direction is better than litigation, the machinery is unproven, and the first real audit will tell us more than any press release. We will see whether the model works when someone shows their arithmetic.

Until then, the artist protection that exists is the protection written into your agreements, and the negotiating window is open right now, while rights holders want catalogue participation to look voluntary.

Someone is going to answer the consent question for your catalogue. The only thing still up to you is whether you have read the clause that lets them.

Not sure which tool to use?

Compare the top AI music and sound tools side by side — honest reviews, real pricing, no sponsorships.

Compare the Tools
T

Theo Brandt

Tutorials Writer

Theo Brandt writes step-by-step tutorials for AI music tools — prompting, stem workflows, and release prep — from a bedroom studio that started with a cracked DAW and a $60 mic. More by Theo Brandt →