# My song was flagged as AI-generated. What do I do? > An AI flag is a classifier's guess, not a finding of fact, and it is usually resolved by responding through your distributor with whatever record of the work you already have — project files, dated session backups, stems, and drafts. Nothing you do after the fact creates new evidence, so the practical goal is to present what already exists clearly, and to start keeping a better record for the next release. _By [Chronatum](https://chronatum.com/) · Published 2026-08-06, updated 2026-08-08_ If you are reading this because a release got held, rejected, or quietly taken down with a note about AI-generated content, first find out which policy and which signal triggered the action. Some notices concern an automated score; others concern impersonation, metadata, spam, or a rights-holder claim. Those are different cases with different remedies, and a short support message often collapses them into the same word: AI. ## What the flag actually means A detector does not find AI. It scores how closely a recording resembles the output of the generators it was trained on, and a platform picks a cutoff above which it acts ([the mechanics, and where they break](/learn/how-ai-music-detection-works)). Two things follow from that, and both matter when you write your appeal. **A score is not evidence of what happened in your room.** It is a statement about how your finished audio compares to a training distribution. A recording can land above the line for reasons that have nothing to do with how it was made. **The cutoff is a business decision.** The same file can pass one platform and fail another, or pass today and fail after a model update. If your track was fine last year and is flagged now, nothing about your track changed. ### Why clean, modern production gets caught A detector learns statistical regularities in its training data, not a universal signature of AI. Peer-reviewed work has found that systems can depend on low-level artifacts and can change their answer after an ordinary transformation such as resampling. That means production and delivery choices can move a file across a threshold even though the history of the session did not change. > **Note.** The reverse is also true and worth saying plainly: a track that passes a detector has not been shown to be human-made either. Detection is weak evidence in both directions. Spotify's current approach, for example, combines separate policies for impersonation and spam with optional, role-level AI credits supplied through participating distributors. ## What to do this week 1. **Find out what was actually flagged.** Ask your distributor which release, which track, and whether the objection is AI generation, an undeclared AI component, or something adjacent like a suspected sound-alike of an existing recording. These get different responses, and support tickets frequently conflate them. 2. **Answer the disclosure question honestly, before anything else.** If a generator wrote the lyrics, sang a vocal, played a part, or produced the instrumental you rapped over, that has to be declared — including when you performed over the top of it. An inaccurate declaration is a much worse problem than a flag, and it is the one thing in this process that can follow you. [What you actually have to declare](/learn/ai-disclosure-when-you-release-music) is narrower than most people assume. 3. **Gather what you already have.** Project files with their edit history, dated backups or cloud-sync versions of the session, stems and bounces at different stages, voice memos, lyric drafts, plugin licences, sample receipts, and anything a collaborator can confirm. Do not create new files to look like old ones; the metadata will not support the story and it turns a resolvable dispute into a credibility problem. 4. **Write a short factual reply.** When the session happened, what you used, who was involved, and what you are attaching. Skip the argument about whether detectors work. The person reading it has a queue and a checklist. 5. **Escalate once, in writing, if the first reply is a template.** Ask specifically what evidence would resolve it. That question is answerable and it moves the ticket to someone who can act. ## Why proving it is harder than it should be The uncomfortable part is that a finished audio file carries almost no information about how it came to exist. Two identical-sounding exports — one played and mixed over three weeks, one generated in twenty seconds — are, as files, nearly indistinguishable. Everything that made them different happened before the export and was never written down. That is why the evidence people reach for is always indirect: project files, timestamps, drafts, witnesses. It usually works, because it is genuinely hard to fake convincingly. But it is slow, it depends on habits you had months earlier, and it puts the burden on the artist to reconstruct a story from whatever survived. [What counts as evidence, and in what order](/learn/how-to-prove-you-made-your-music) is worth reading before you need it. ## Where Chronatum fits — and where it does not We make [Chronatum](/), so read this section with that in mind. It is a macOS app that watches a DAW session while you work and records the production activity it can observe — edits, arrangement changes, the shape of the session over time — then structures that into a signed record bound by hash to the exact file you exported. **It cannot help with the track that was already flagged.** There is no way to reconstruct a session that has already happened. If Chronatum was not running while you made it, it has nothing to report about it, and we would rather say so here than have you install it expecting otherwise. **It also does not certify that music is human-made.** That is a stronger claim than any tool can support, and we do not make it. A record says what was observed in one DAW session — it cannot see what happened outside that session, what was imported, or what was made somewhere else and brought in. A proof separates what was directly observed from what was inferred, what you declared, and what could not be determined, and it leaves the unknowns visible instead of rounding them into a verdict. What it is genuinely useful for is the next release: having a contemporaneous, verifiable record of the work already in hand, so that responding to a flag is attaching one file rather than assembling an archaeology project. If that is the problem you have, [the setup guides](/setup/ableton) take a couple of minutes, and anyone can [check a proof](/validate) for free without an account. ## Keeping a better record from here Independent of any product, these habits cost little and are what actually resolves disputes: - Keep versioned project files rather than overwriting one. Dated saves at natural milestones are enough. - Let cloud sync or Time Machine keep its history on your projects folder — automatic, dated, and hard to fabricate after the fact. - Bounce rough mixes as you go and keep them. A sequence of drafts is one of the more persuasive things you can show. - Keep receipts for samples, presets, and any AI tool you legitimately used, along with the licence terms. - Write down who played what, and when. A collaborator's confirmation carries real weight. - Declare AI components accurately at upload, every time. Consistency across releases is itself a signal. ## The short version A flag is a probability, not a finding. Answer the disclosure question honestly, present the record you already have plainly, and ask what specifically would resolve the ticket. Then make the next record better than the last one — because the moment to document how a track was made is while you are making it, and that moment does not come back. ## Sources and further reading - [Spotify Strengthens AI Protections for Artists, Songwriters, and Producers](https://newsroom.spotify.com/2025-09-25/spotify-strengthens-ai-protections/) — Spotify - [AI credits on Spotify](https://support.spotify.com/us/artists/article/ai-credits/) — Spotify for Artists - [Music that impersonates another artist's voice](https://support.spotify.com/us/artists/article/music-that-impersonates-another-artists-voice/) — Spotify for Artists - [The AI Music Arms Race: On the Detection of AI-Generated Music](https://transactions.ismir.net/articles/10.5334/tismir.254/) — Transactions of the International Society for Music Information Retrieval --- More: [llms.txt](https://chronatum.com/llms.txt) · [sitemap](https://chronatum.com/sitemap.md)