Believe flagged my song as AI. Now what?

The email lands at some ugly hour. It's from Believe, or DistroKid, or TuneCore, or CD Baby. Subject line: “Content Review — Action Required”. Body: your track has been flagged as potentially AI-generated, in violation of their content policy, and the release has been paused pending proof of human authorship. Streams stop. Royalties freeze. The release date you promised your fans slips.

The worst part is that they don't tell you which model flagged it, which threshold you crossed, or which specific 30-second segment triggered it. What you get is a link to an appeals form asking for “evidence of human authorship” — usually a video of you performing or producing the track. So you record the video. You send it. Sometimes it works. Sometimes it doesn't. And even when it does, the underlying audio file still trips the classifier the next time it's scanned.

If you're here, you're living this right now. Let's walk through what's actually happening and what your real options are.

Why distributors are pulling human tracks in 2026

In early 2025, Deezer publicly reported that roughly 18% of daily uploads to their platform were fully AI-generated (Suno / Udio / Riffusion). By mid-2026 that number was estimated at 28-34% depending on the aggregator. Streaming platforms responded by strong-arming distributors into pre-filtering uploads through third-party AI-audio classifiers — the two most common being Modulate's Velma and PEX AI Audio Fingerprint. Believe, TuneCore, DistroKid, and CD Baby all now run at least one of these on every upload.

The classifiers were trained on Suno v3 / Udio v1.5 outputs from 2023-24. They're now being asked to classify 2026 human productions on the same decision boundary. The false-positive rate in the wild, based on the data I'm seeing across DECODED's incoming scans, is somewhere between 11% and 17% for “radio-ready” indie tracks — particularly anything mastered with LANDR, eMastered, or iZotope's Ozone Master Assistant.

The classifier has no idea what you look like on video. It has no idea whether you toured the record last summer. It processes 30 seconds of audio and returns a probability. That's it.

Does the video-evidence workaround actually work?

Sometimes. Here's the honest breakdown from ~140 appeals I've watched artists push through this year:

• Believe: Appeals success rate ~42% when a full session-timeline screenshot + a >30-second video of you performing a recognizable section of the vocal are included. Turnaround: 5-14 business days. If you have a co-writer split sheet, attach that too — it doubles the odds.
• DistroKid: ~28%. Their internal reviewers lean hard on the classifier score. A video helps mainly on second appeals after the first is auto-rejected.
• TuneCore / CD Baby: ~35%. Faster response (3-7 days) but they'll ask for the raw multitrack session file if they're on the fence. If you don't have DAW project files — e.g. you mixed in-the-box and consolidated tracks — your odds drop to under 15%.
• Symphonic / AWAL: ~55%, mainly because they have live A&R staff who can listen and override. Almost no algorithmic appeals process.

Two things to know before you record the video:

• Distributors are looking for continuity of performance, not talent. Sing a section, play a hook on the actual instrument in the recording, or show the mix session with the timeline scrubbing. Don't just hold up a piece of paper with your name.
• Include a timestamp overlay or a visible date (newspaper, phone lock screen). Multiple appeals I've reviewed were rejected because reviewers assumed the video was pre-recorded from years ago.

Why the video isn't enough on its own

Even after a successful appeal, the appeal only clears that release, on that distributor. The audio still fails the classifier when:

• You upload the same file to a different distributor for a re-release.
• Spotify runs its own periodic AI-audit sweep (they crawl the back-catalog quarterly).
• A curator submits your track to SubmitHub or Groover for playlist consideration.
• YouTube Content ID re-scans your video's audio track.

You're fighting the same battle over and over on the same audio file. The video appeal is a bandage; the audio fingerprint is the wound.

The audio fix: neutralize the fingerprint, keep the record

The classifiers key on signal-level fingerprints, not on musicality or lyrics. Which means those fingerprints can be neutralized in a way that's inaudible to a human ear on any listening system, but moves the track out of the classifier's “AI-shaped” region.

What we do in the DECODED pipeline, specifically for tracks that were pulled by distributors:

• Stem-isolated re-editing. We split the track into drums, bass, other, and vocals. Then we scan each stem independently against the live classifier. In practice, 9 out of 10 distributor flags are anchored on one stem — usually the mastered mix bus or the AI-assisted vocal tuning — and the other three stems are clean.
• Targeted counter-fingerprints on the flagged stem only. Micro-timing offsets (2-8 ms), envelope perturbation, ±30-cent wandering formant shift, sub-audible pink noise floor, and an all-pass phase rotation stack that breaks the “coherent mastering” centroid without touching perceived loudness or EQ.
• Stochastic best-of-N re-roll. The pipeline renders the track 4-8 times with different random seeds against the same counter-fingerprint stack. Only the render that scores lowest on the live classifier ships. The others are discarded.
• Certificate of Authorship. We generate a signed PDF report showing the before/after classifier score, timestamped, tied to your account. Believe, TuneCore, and CD Baby all currently accept this document as supplementary evidence on appeal. (DistroKid does not accept it as primary evidence but will reference it.)

Turnaround: 4-6 minutes end-to-end. You upload the exact WAV or FLAC that got pulled, and you get back a bit-for-bit new master + PDF certificate + before/after score. You only pay if we bring the score under the platform's threshold. Failed renders do not charge your card.

The exact playbook if Believe just pulled your release

1. Don't retract the release yourself. Retracting resets your Believe royalty accrual and makes the appeal harder to file — the release ID gets archived. Wait for the appeal to resolve.
2. File the first appeal within 48 hours with only the evidence you already have. Session screenshot, one video clip, a short written statement. Don't burn your ammo. If it gets accepted (~42% odds), you're done for this release.
3. In parallel, run the master through a scan. Free at DECODED. You'll see the actual classifier probability the platforms assigned your track — the number Believe won't tell you. If it's over 49%, the odds of other platforms flagging it later are high, and you should preemptively defend the audio.
4. If the appeal is rejected, escalate with a new master. Re-submit the same release with the defended WAV as the audio asset and the Certificate of Authorship as supplementary evidence. The appeal reviewer sees a lower classifier score on re-scan and the appeal moves. In my sample, second-appeal acceptance rate is near 78% when both the defended audio and the certificate are attached.
5. Re-upload the defended master to every other distributor and platform you use. This is the step most artists skip. The old audio file is still sitting in Spotify Content ID, SubmitHub's hash cache, and YouTube's ID system. Ship the new master to your DSP delivery so future scans see the clean version.

Common questions

Does defending the audio violate Believe's content policy?

No. Believe's policy prohibits “audio generated in whole or in substantial part by generative AI models” without disclosure. Your track is a human recording. We are not adding AI-generated material; we're applying analog-style audio-engineering perturbations to counter a misfired classifier. It is functionally the same as re-mastering.

Will the certificate hold up if a label / publisher audits me later?

The certificate documents the pipeline that was applied and the classifier score improvement. It doesn't claim your track is AI-free (obviously we can't verify a stranger's track was human-written); it documents that the audio was defended against false-positive classifier flags. Every upload requires an attestation of ownership that you sign. If a label asks, that attestation + certificate + your session files are the audit trail.

What if I actually used Suno for one section?

Then Believe's policy applies and you should disclose it on upload — that's a separate legal question, not something we help with. If the AI section is under ~10% of the track and the rest is human, our pipeline sometimes still fails the scan (the AI stem's fingerprint is that strong). In that case we refund and you go back to the drawing board. We built this for humans getting mistreated, not for machine outputs being laundered.

How much does this cost vs. losing the release?

DECODED is $19 flat for a passed defense. Believe royalties on a typical indie release run $80-400 in the first quarter. If your appeal is trending toward rejection, the math is a single flat fee to keep the release live — there's no subscription and no upsell.

Closing note

If you're staring at a Believe / TuneCore / CD Baby email right now telling you your track is flagged as AI: it's not just you, it's not personal, and it's fixable. The video appeal is worth filing today. The audio fix is worth running before the appeal is rejected. And the certificate is worth attaching with the appeal — not after.

Upload the file. Free scan. Honest verdict. If we can't save it, we'll tell you what to change and how — no charge.

— Kenny, founder, DECODED