Licensing Your Music to AI Companies: What to Ask Before You Sign

For two years, AI training deals were a story about giants: labels and publishers negotiating catalog-wide agreements while everyone else watched. That phase is ending. Aggregators, data brokers, and AI companies themselves are now reaching independent artists directly, with offers that range from genuinely fair to quietly extractive, often wearing the same friendly email.

Saying yes is not a betrayal and saying no is not a strategy. Licensing your work for machine learning is a legitimate income stream, one that did not exist five years ago, and it deserves the same scrutiny you would give a publishing deal. Here is what to ask before anything gets signed.

First, understand what kind of deal it is

AI music deals are not one thing. The offer in your inbox is probably one of three shapes, and everything else depends on which:

• Training licenses: they pay to learn from your recordings. Your work shapes a model's understanding; it is not supposed to come out the other side recognizably.
• Output or generation licenses: they pay to produce material based on you, your voice, your style, your stems. This is closer to a work-for-hire crossed with an impersonation permit, and it should cost them far more.
• Reference or catalog access: they pay to index, analyze, or match against your catalog, for search, recommendation, or detection. Least invasive, and priced accordingly.

If the contract does not make clear which of these it is, that is not an oversight. Ambiguity in scope always favors the party that wrote the document.

The questions that separate good deals from grabs

Scope: which works, which models, which uses? A license for named recordings to train specified model families is a deal. A license to all works, present and future, for any machine learning purpose is a donation. Watch especially for future works clauses; you are being asked to price music you have not written yet.

Term and revocation: can you leave, and what happens when you do? Trained weights cannot practically unlearn you, which makes exit rights tricky, but good deals handle it honestly: fixed terms, renewal payments, and commitments about future training runs after termination. A perpetual, irrevocable license for a one-time fee means the relationship ends the day the check clears, and only for you.

Exclusivity: almost never worth it. Exclusive training rights lock you out of every future deal in a market where prices are still being discovered and mostly rising. If a company insists on exclusivity, the number should be dramatic, and even then, think hard.

Outputs and attribution: what happens when the model produces something close to you? Ask directly: does the license cover outputs that resemble your work or voice, is there attribution, is there a share of output revenue, is there a process when something crosses the line? Style mimicry is the live wire in all of this; a contract that is silent about outputs has answered your question in the company's favor.

Payment structure: flat fee or per-use? Flat fees are simple and final; per-use structures pay as the model earns. The honest answer is that per-use is better when you have leverage and auditability, and flat is better when you do not trust the meter. Which raises the next question.

Audit and reporting: can you verify anything? A per-use royalty without reporting obligations and audit rights is a promise, not a term. Ask what gets reported, how often, and what happens when the numbers look wrong.

Warranties: what are you promising them? Every deal will ask you to warrant that you own what you are licensing. This is where unregistered works and handshake splits become dangerous: if a collaborator surfaces later, the indemnification clause makes their problem yours. Do not warrant a catalog whose paperwork you have not actually done.

Negotiate like a machine can read you

The artists getting the best terms right now share one trait: their catalogs are legible. Works registered with timestamps and fingerprints, contributors and splits recorded, permissions published in machine-readable form, including an explicit position on AI training, priced rather than merely forbidden.

Legibility changes the negotiation twice. It makes you safer to license, clean provenance is precisely what AI companies' lawyers are hunting for, and it gives you a public reference price. When your published terms already say what training access costs, the conversation starts at your number instead of theirs. Silence, by contrast, is being treated as consent by the worst actors and as risk by the best ones. It is the only position with no upside.

A short list of red flags

• Rights to future works, or to your name, image, and voice, folded quietly into a music deal.
• Perpetual and irrevocable paired with a modest one-time payment.
• No definition of what happens to derived data and model weights at termination.
• Confidentiality terms that prevent you from ever discussing what you were paid.
• Pressure to sign inside a deadline measured in days. Real buyers of decade-long rights can wait two weeks for a lawyer to read the document.

None of this requires you to love or hate the technology. It requires you to treat a new class of buyer with the same professional skepticism the last century of music history earned. The companies training on music are not going away, and neither is the leverage of the artists who kept their records straight. Get the paperwork done, set your terms, and make them meet you at your number.