We use AI ourselves and we actively encourage our partners to use it. We accept AI-assisted music. What we do not accept is a recording with no human author in it — because a work with no human author has no rights to distribute, and no revenue to collect.
Generative tools are now part of professional production, the same way samplers, autotune and DAWs were before them. We build with AI internally, we distribute records made with it, and we would rather our partners learn these tools early than defensively. Nothing on this page is an argument against using AI in your process.
It is an argument about authorship. A distributor's job is to move rights, and rights need an author. When there is no human in the chain, there is nothing to own, nothing to license, and nothing to defend when someone else uploads the same output. That is not our opinion — it is the structural problem underneath the entire category, and it is why stores are removing this material at scale.
The question is never did you use AI. The question is whether a person authored something we can point to.
There is a single distinction and it is not subtle. On one side, a human wrote, performed, played, programmed or produced something that materially shapes the record. On the other, a prompt was typed and a file came out.
Your composition, your melody, your lyrics — with generative tools used for arrangement, instrumentation or production. The authorship is yours; the tool executed it.
A real vocal performance carries authorship into the recording, whatever produced the backing.
Parts performed on an instrument, or programmed and sequenced by a person, sitting alongside generated elements.
Generated stems that a producer has selected, restructured, re-arranged, mixed and mastered into a record that reflects deliberate creative decisions.
Visual assets generated with AI are fine, provided the underlying recording meets the standard above and the imagery clears each store's own policy.
A text prompt into a generative model, exported and uploaded. No writing, no performance, no production decisions. This is detected and rejected.
High-volume uploads of generated material built to farm playlist placements or algorithmic streams. Rejected on sight, and grounds for termination.
Recordings using a cloned or trained voice of an identifiable performer without documented authorization. Separate from the authorship question and refused outright.
Material framed to be mistaken for a named artist's catalog, regardless of how it was produced.
A generated track with a token edit applied to clear a checkbox. Trimming an intro is not authorship, and it does not survive review.
A release needs at least one of these, and it needs to be real enough to describe and, if we ask, to evidence. Most legitimate AI-assisted records clear all three without trying.
A person wrote the music, the topline or the lyrics. Session files, lyric sheets, publishing registrations or writer splits all establish this. If a work is registered to a writer, the writer exists.
A person sang or played something that is audible in the master. A recorded vocal, a tracked instrument, a programmed part with human phrasing decisions in it.
A person made the record what it is — structure, arrangement, selection, editing, mix and master. Deliberate, documentable creative choices rather than an exported default.
Copyright attaches to works of human intellectual creation. Where there is no human author, most major jurisdictions — the United States most explicitly — hold that there is no protectable work. No protectable work means no rights: nothing to assign to us, nothing to claim on YouTube, nothing to license, nothing to enforce against a third party who releases the identical output tomorrow.
This is not a moral position, it is a mechanical one. We are a rights business. If we deliver a recording that cannot hold rights, we have delivered a liability into your catalog and ours, and the stores will eventually strip it out along with the streams and the revenue attached to it. Legal frameworks around AI-generated output are still developing and vary by territory; our policy is set conservatively so that your catalog stays deliverable as they settle.
Fully generated audio carries signatures. Our ingestion pipeline screens for them, and flagged material goes to human review rather than straight to rejection — because the tools that produce a fully generated track are the same tools that produce a legitimate AI-assisted one, and the difference lives in what you did with the output.
Deliveries are analyzed on arrival. Material with no detectable human authorship is held before it reaches a single store, so the problem never becomes a takedown across two hundred platforms.
Pre-DeliveryA flag is not a verdict. Where the signal is ambiguous, we come back to you and ask what you did — session files, stems, a vocal take, a writer split. Straightforward answers resolve it quickly.
Human ReviewRejected material is not redelivered under a different UPC. Repeat attempts and bulk generated catalog end the partnership — protecting every other catalog we deliver from association with it.
No ExceptionsSeveral stores now require AI involvement to be declared in metadata, and more are adding it. Tell us what was used and where. Accurate disclosure is not a penalty — it is what keeps an AI-assisted release accepted, credited and monetized instead of quietly suppressed. Concealing it is the fastest way to lose a delivery.
Content delivered through MusicYes is not cleared for use in training AI systems. We do not grant that permission, we do not licence catalog for it, and we do not treat delivery to us as consent to it. That is the default position for every partner, on every release, unless you tell us otherwise.
We will be direct about the limits here, because you deserve an accurate answer rather than a reassuring one. We cannot verify whether an AI company has scraped or trained on music we distribute without our permission. No distributor can. The material sits on public platforms, and enforcement against unauthorized training is an unresolved problem across the entire industry.
What we can do, and do: withhold permission explicitly rather than by silence, so that any such use is unauthorized on the record; refuse to licence catalog for training as a revenue line of our own; pass through the opt-out signals the platforms make available to us; and enable training use only where a partner has asked for it in writing. If that position changes because a store or a partner deal makes it necessary, we will tell you before it applies to your catalog — not after.