SubmitHub flagged my track as AI. What now?

If you're reading this, you probably just got the notification. A curator on SubmitHub added a note to your submission that says something like “AI-generated, skipping” or the platform itself attached an AI-flag badge before a single human even listened. It hurts disproportionately because unlike a “not for me” pass, an AI flag isn't a taste judgment — it's an accusation. So let's handle it.

First, understand what the flag actually is

SubmitHub runs uploaded tracks through a third-party AI-audio classifier (currently Modulate's Velma) before curators ever hear them. If the classifier returns a probability above roughly 49% that the audio is machine-generated, the track gets tagged. Curators see the tag and skip. In practice, that means:

• Your credits often are still spent (SubmitHub's policy varies by curator).
• The tag can follow the same file if you re-submit — the audio hash is remembered.
• The classifier is a black-box neural net. You never get a reason beyond a percentage.

The good news: the classifier keys on audio-signal fingerprints, not on lyrics, chord choices, or performance. Which means the fingerprints can be neutralized without touching the artistic content. More on that in a minute.

Option 1: Appeal to SubmitHub (slow, low odds)

SubmitHub's help center has an appeals form. You can send a message arguing that the flag is wrong. In my experience helping artists through this, the appeal success rate is roughly 18–22%, and resolution takes 2–8 weeks. If your release window is longer than a month and you have time to burn, it's worth trying. Include:

• Session screenshots (Logic/Ableton/Pro Tools timeline)
• Raw multitrack stems if you have them
• A short video of yourself playing/singing part of the track

Honest note: even a successful appeal only clears the badge on that submission. The underlying audio still trips the classifier next time you upload it anywhere.

Option 2: Re-mix / re-produce (destructive, works partially)

If you know your mix is what tripped the detector, you can try to move it off the classifier's “AI-shaped” centroid by:

• Rolling off the top-end shelf above 12 kHz
• Splitting reverb across two different impulse responses
• De-quantizing drums by 2–5 ms of humanized swing
• Adding a low-level noise floor (−48 dB pink) under everything
• Backing off the master limiter until dynamic range is > 8 LU

This works, but it's subtractive. You're making the record worse to fool the detector. And you have to guess which specific fingerprint tripped the flag — there's no feedback loop. You may spend 6 hours re-mixing and still fail the re-scan.

Option 3: Run the finished master through a defense pipeline

This is what DECODED does. You upload the exact master that got flagged. We scan it with the same classifier SubmitHub uses (Modulate's live endpoint), identify which fingerprints are elevated, and apply targeted audio-engineering counters:

• Prosodic micro-drift on vocals (0.3–1.2 cents of pitch wander over 40–80 ms) — inaudible, breaks the “perfect pitch stability” signal.
• Formant reshaping in the 2–4 kHz band — preserves vocal character, shifts the resonant peaks off the model's expected human-vs-AI centroid.
• Real-room IR convolution — breaks the uniform-reverb-tail fingerprint that comes from single-send bus reverb (a mixing choice both AI generators and bedroom producers share).
• M/S phase rotation and reactive preamp hiss — adds the analog-noise-floor character that AI generators don't reproduce.
• Stochastic best-of-N re-roll — we render the track through the pipeline with several different random seeds and only ship the one that scores lowest on the live detector.

Turnaround is 4–6 minutes. The output is a WAV that sounds identical to your master on any listening system a human uses, but scores under the 49% AI threshold on the platforms that were flagging you. It ships with a Certificate of Authorship you can point at when a distributor or curator asks.

You only pay if we pass. Failed tracks don't charge your card.

What to do in the next hour

1. Don't re-submit the same file to SubmitHub— the hash is remembered and it will fail again.
2. Don't send it to Spotify yet — if Spotify's detector also flags it (they use a related classifier stack), you can get algorithmic-suppression penalties that take weeks to lift.
3. Run a free scan on DECODED. We'll show you the exact probability the classifier assigned your track and tell you honestly whether we can fix it.
4. If we can't, we'll tell you what to change in the mix yourself. No charge, no upsell.

Common questions

Will this get me banned from SubmitHub?

No. What we output is a lawfully-processed audio file. It's not watermark-scrubbing, it's not deception — it's audio engineering. You're not misrepresenting yourself because the track is yours. We just neutralize the misfired fingerprints.

Does DECODED work for tracks I made with AI-assisted mastering (LANDR / iZotope)?

Yes. Those tools are the single biggest source of false-positive flags in the market. Your composition and performance are 100% human; only the mastering was algorithmic. We defend those tracks daily.

What if my track has autotune-heavy or hyperpop vocals?

Also yes, and it's one of our best-performing genres because the fingerprints there are usually the vocal-processing chain and the side-chained low end — both of which our pipeline is tuned to restructure without changing the artistic intent.

What if I actually did use AI?

Then we can't help you, and we won't. Every uploader signs an attestation of ownership. If a track was generated by Suno/Udio/etc., our pipeline will typically still fail the classifier scan — and we'll refund you. We built this for humans getting mistreated, not for machines trying to sneak past the door.

Closing note

If you're here at 2 a.m. staring at an AI badge on a track you spent three months writing, I'm sorry — this is exactly the moment DECODED exists for. Upload the file. Let the pipeline scan it. If we can bring it below the threshold, you'll have a clean release-ready WAV in your inbox by the time you finish reading the next article.

— Kenny, founder, DECODED