Browse articles

AI disclosure, explained

The three AI disclosures Run collects, what the four levels mean, the separate question on each credited person, and why disclosure is never a penalty.

4 min read Updated

On this page
  1. The three disclosures
  2. The four levels
  3. The question on each credited person
  4. What the checks do with your answers
  5. Where the answers go
  6. Changing a disclosure after delivery

Run asks three questions about AI on every release, and one more about every person you credit. All of them are required: a release does not submit until they are answered.

Disclosure is not a penalty. Every level on the scale is accepted and distributed, fully AI-generated music included. What Run refuses is not a level: it is conduct, such as cloning a real singer’s voice or passing a release off as someone else’s. See Run’s AI policy for that list.

The three disclosures

DisclosureWhere you answer itWhat it covers
ArtworkRelease section, next to the coverThe cover image only
RecordingEvery track, under AI involvementThe sound recording: performance, vocals, instruments, production
SongwritingEvery track, under AI involvementThe composition: melody, lyrics and structure

Artwork is one answer for the whole release. The other two are answered on each track separately, because the two are genuinely different questions: a song you wrote yourself can be performed by an AI voice, and a fully generated instrumental can carry lyrics you wrote by hand.

The four levels

All three disclosures use the same four-step scale.

LevelWhat it meansExample
NoneNo AI anywhere in that part of the workYou played and sang the track, and mixed it yourself
AI-assistedAI tools helped, but did not generate any part of the workAn AI mastering or noise-removal tool, AI stem separation, a synth preset engine
Partly AIAI generated material portions of the workThe second verse vocal is generated; the rest is you. Or AI wrote the hook lyric
Fully AIThe whole of that part is AI-generatedYou generated the recording from a prompt and released it as it came out

The line between AI-assisted and Partly AI is whether anything a listener hears was generated. A tool that cleans up, masters or reshapes something you performed is assistance. A vocal, melody, beat or lyric section that came out of a model is a material portion, even if it is eight bars long.

Answer for the version you are releasing. If you generated ten drafts and then re-sang the whole thing, the recording you are delivering is what the disclosure describes.

The question on each credited person

Song credits carry their own, separate AI question, one answer per person, describing that person’s own contribution rather than the track:

  • Human: they did their part themselves.
  • Human + AI: they used AI in their part.
  • Fully AI: their part was generated.

One rule follows from it: at least one credited person on a track must not be marked fully AI. A track where every credit is fully AI does not submit. Someone chose the prompt, picked the take and decided it was finished, and that person belongs in the credits. Credits, contributors and publishing splits covers how credits work.

Keep the two consistent. If a credit says AI was involved while both of that track’s disclosures say None, the checks raise it for a reviewer.

What the checks do with your answers

  • Missing disclosures block the submission: the artwork answer at release level, both answers on every track.
  • A credit marked as AI-involved against two “None” disclosures is flagged for a person to look at.
  • A track that discloses AI-generated audio (Partly AI or Fully AI) with no AI voice or likeness consent document attached goes to a reviewer. There is no field anywhere that says “this is a real person’s voice”, so a person checks rather than software guessing. If your track uses an identifiable voice, open a request from Support pinned to the release and send the consent there: the release form has upload rows for cover, public domain and sample clearances, but not for this one.
  • Titles in the “AI cover” or “AI voice” format fail the checks outright. That phrasing is the marketing format of voice-clone releases. The vaguer “AI version” and “AI remix” wording is flagged for a reviewer instead.
  • “In the style of” in a title fails the checks, with or without AI.

Where the answers go

Your disclosures travel with the release when we deliver it. What each store does with them after that is that store’s own policy, and stores are the ones setting the AI rules they enforce, which is why a wrong answer is worth correcting rather than leaving.

One consequence inside Run: a track whose recording is disclosed as Fully AI cannot be opted into Content ID. See choosing stores and Content ID.

Changing a disclosure after delivery

Settle these before you submit. Changing an AI disclosure on a release that is already out is not a quiet metadata edit: it sends the release back through a fresh review in our distribution pipeline, and availability on the stores can be affected while that review runs. It is also not a self-serve edit, because a delivered release is locked. Open a support request.

Tell us anyway if an answer was wrong. An honest correction is always better than a wrong disclosure sitting on a live release.