Your track was wrongly flagged as AI. What to do.
You wrote it, played it, mixed it, and a machine decided a machine made it. This happens, it is not rare, and the appeal paths are worse than the detection. Here is what is actually known.
Start here. Do not delete or re-upload the track. Gather dated evidence of the creation process first, because that is the only thing that has been shown to reverse these decisions.
It happens to real bands
In April 2026 the metalcore band ZAO was flagged by their distributor TuneCore for suspected generative AI use on an unreleased recording, which delayed the release. The band published a screenshot of their DAW session as proof and said they could not get a person to explain the accusation. The flag was lifted after they made it public.
Source: Vice, Metal Injection and Theprp.com, late April to early May 2026, citing the band's own posts.
One case does not measure how common this is, and nobody publishes that number. But it shows the shape of the problem: an automated flag, no human to talk to, and resolution through public pressure rather than through process.
What actually triggers a false positive
Independent research has identified concrete triggers, and they are not what you would guess.
Your export settings can look like a generator
A study from KTH Royal Institute of Technology found that detectors lean on encoding correlations rather than musical properties. Suno output is consistently 48 kHz at 192 kbps, Udio at 48 kHz and 320 kbps, while a reference corpus of human music was mostly 44.1 kHz with variable bitrates. A human musician exporting at 48 kHz and 320 kbps presents the same encoding profile as Udio.
High-pass filtering can flip a detector completely
The same team found that applying a high-pass filter at 8 kHz or above made one commercial detector label every test excerpt as AI, including human tracks from the reference corpus. That test ran on a handful of samples, so treat it as proof that the failure mode exists, not as a frequency.
Source: Cros Vila, Sturm, Casini and Dalmazzo (KTH), "The AI Music Arms Race", Transactions of the ISMIR, June 2025.
AI mastering is the one production tool with measured impact
A 2026 benchmark applied a generative mastering tool to human-composed, human-performed tracks and measured the AI flag rate rise across every detector tested, in one case from 0.1% to 52%. That preprint has not been peer reviewed and it tested a single mastering tool, so do not extrapolate to every plugin. But if you are being flagged and you cannot work out why, your mastering chain is the first place to look.
Source: Go and Kim, "HAIM: Human-AI Music Datasets", arXiv preprint, June 2026. Not peer reviewed.
What does not seem to trigger it
Pitch correction, virtual instruments and quantised electronic genres come up constantly in artist forums as suspected causes. There is no published measurement supporting any of them. That does not make them safe, it means nobody has checked.
The evidence that works
Detection decisions get reversed on process evidence, not on argument. Assemble, in this order:
- The project file itself, with its full track and edit history. This is the single strongest artefact. A session with hundreds of edits, punch-ins and automation moves is very hard to fake and immediately readable by a human reviewer.
- Dated raw recordings and stems, ideally with the original interface metadata intact.
- Timestamped process footage. Phone video of a tracking session, even thirty seconds, carries a lot of weight.
- Your export chain, written out: DAW, sample rate, bitrate, and every plugin in the mastering path. If a reviewer can see there is no generative tool in the chain, that closes the question.
- Prior catalogue released before the generator in question existed, which establishes a consistent style.
How to escalate, by platform
A flag can come from your distributor before release or from the streaming service after it. They are different problems.
Distributor flags, before release
This is where the money is blocked and where the process is weakest. Reply to the ticket with the evidence list above attached, in one message rather than a thread. Ask explicitly for human review and for the specific basis of the flag. If you get no substantive answer, the ZAO case suggests that public, factual, non-abusive escalation works when the ticket queue does not.
Streaming service labels, after release
On Deezer, the label sits at album level and removes you from recommendations rather than from the catalogue. Contact the service through your distributor, since that is the contractual relationship, and supply the same evidence pack.
Why the industry is quietly cautious about all this
Spotify said publicly in September 2025 that detection systems are imperfect and produce a lot of false positives, and that they are "not helpful for policy enforcement". That is a major platform saying in the open that it does not trust automated AI detection enough to enforce policy with it.
It is a reasonable position, and it is the same reason we do not publish a single accuracy number for our own engine. A detector's score is an indication about a rendering chain, not a verdict about a person. If a flag has cost you a release, that distinction is the argument to make, and it is well supported by the published research above.