How distributors decide your song is "AI", and what to do when they get it wrong

Most independent artists only find out how their distributor screens for AI when it goes wrong: a release held in review, a rejection email, a track pulled. What happens before that has been private. A court filing this month gives the first official look inside DistroKid's process, and it is mostly a list of questions nobody has answered for artists.

What we actually know about DistroKid's detector

On 14 October 2026, Universal, Capitol and Sony asked a federal court in Manhattan to force DistroKid to comply with a subpoena issued in their copyright case against Suno, which is being heard in Massachusetts. The labels' proposed order, filed the next day, is the useful document for artists. It refers to a third-party "prediction model" that DistroKid described in an October 9 response to the subpoena, and asks for two things.

First, month-by-month figures from the date DistroKid started using the model: how many recordings it delivered to streaming services, how many it tested with the model, how many the model scored as likely AI-generated in whole or in part, and the streams and royalties those flagged recordings earned. Second, the model's name and vendor, whether it can tell which AI service made a track, and any analysis DistroKid has of the model's accuracy, including the basis for statements it made about false positives. The proposed deadline is 30 October. As of Beyond Copying's report, the court had not ruled.

Read that from the artist's side and three things stand out. DistroKid runs uploads through an outside model. That model produces a score or category, not a simple yes or no, since the order covers whatever "probability levels or categories the model reports". And DistroKid has told the labels' lawyers something about false positives. We couldn't find any of it in DistroKid's help centre, which explains AI credits in detail but doesn't describe a detector.

The vendor is not public, and we couldn't confirm any name from DistroKid or from the filings, so we aren't guessing. Zinstrel notes that DistroKid won't say what tools drove its recent removals. In a statement to Billboard, DistroKid said (as reported by MBW) it is a third party and is "committed to transparency".

There is one detail worth noticing. The order asks for streams and royalties earned by recordings the model scored as likely AI, counted among recordings DistroKid delivered. That suggests a "likely AI" score has not automatically meant a blocked release. What a score does trigger, whether a label, a review or nothing, is not public.

The part you control: AI credits

DistroKid's upload form asks whether any of the track was generated by AI. If yes, you pick at least one of four boxes: the lyrics, the music, all of the audio, or part of the audio. Pick "all of the audio" and it also asks whether the artist name is a human or an AI persona. You can add or change credits after release from the album page.

DistroKid's explainer says Spotify, Apple Music and YouTube currently show these credits to listeners. It also draws a line: you don't need AI credits for pitch correction, AI-assisted mixing or mastering, or AI-assisted workflows. That line is the right one. Whether the detector draws the same line between generating and assisting is exactly what nobody has published.

How the platforms talk about their own detectors

Deezer, which built its own system, at least publishes numbers. In July it said its detector is 99.8% accurate, may miss about 2 in 1,000 AI-generated tracks, and may wrongly flag fewer than 1 in 10,000 human-made ones. It applies the AI tag per album rather than per track to cut errors, and removes tracks it detects as AI from algorithmic recommendations and editorial playlists. Since January it has offered the technology to other music companies, distributors included.

Spotify's October update says it combines third-party detection with listener, upload and metadata signals, and admits that AI detection "has limitations, and it will only get harder as generative AI evolves." It doesn't describe how an artist would appeal a detection decision. We covered what that means for discovery in our piece on Spotify's 50 million demoted tracks.

Deezer's figures are its own and we haven't seen an independent test. Even taken at face value, they show the scale problem. Deezer says nearly 90,000 fully AI-generated tracks a day made up more than half its daily uploads at the June peak. By our arithmetic, that leaves fewer than 90,000 other tracks a day, and a rate of 1 in 10,000 would mean up to about nine human releases wrongly flagged every day on one service. Each of those is someone's single.

What to do when you're flagged

• Read the email, then reply to it. DistroKid's rejection guide says the details are in the rejection email, and that if you think it was rejected in error you should reply to that email. Ask in writing which rule was applied and whether the decision came from DistroKid or from a store.
• Don't re-upload the same audio by reflex. The guide says you'll usually need to delete and re-upload to fix what was flagged. If the flag is wrong, the identical file will just be judged again.
• Keep receipts. Save project files, stems, dated bounces, sample and plug-in licences, and voice memos of the idea. A session with forty tracks of your own takes is better evidence than any statement you can write.
• Credit honestly, including partial use. If a generated element is in the final mix, the "part of the audio" box exists for it. A disclosure you made is a better position than a detector finding something you didn't declare.

This is the same distributor at the centre of the labels' data fight with Suno and of UMG's takedown campaign, and the labelling rules differ again on each store, as we set out in how platforms label AI music.

What we think

The most telling thing about this filing is who gets the accuracy data first. If the court grants the order, the first detailed evidence of how well DistroKid's detector judges independent music will go to major-label lawyers, to use against Suno. The artists whose tracks were scored aren't a party to that case and get nothing. Detection accuracy is being treated as discovery material, not as a consumer protection.

That should flip. Any distributor that runs a detector on paying customers' music should publish four things: the vendor, the false-positive rate on human-made music at the threshold it actually uses, what happens at each score, and a human review path with a deadline. A detector score should never be the only reason a release is blocked. Deezer has shown that publishing a number doesn't break anything. Until distributors do the same, treat your session folder as the only appeal you're guaranteed.

Sources

• Proposed order, UMG Recordings et al. v. Kid Distro Holdings (S.D.N.Y. 1:26-mc-00510), filed 15 Oct 2026
• How to Fill Out AI Credits, DistroKid Help Center
• What Are AI Credits?, DistroKid Help Center
• Handling a Rejected DistroKid Release, DistroKid Help Center
• AI music has surpassed 50% of new music uploads for the first time, Deezer Newsroom, 21 Jul 2026
• One Year In: Putting Human Artistry First on Spotify, Spotify Newsroom, 20 Oct 2026
• UMG and Sony ask court to force DistroKid to hand over AI music data for Suno case, Music Business Worldwide, 19 Oct 2026
• UMG and Sony ask a court to order DistroKid to hand over data from its AI detection model for the Suno case, Beyond Copying
• DistroKid's Unlimited Model Is Buckling Under the Weight of AI Music, Zinstrel, 19 Oct 2026