How to check if a song is AI-generated

Step 1 — Get the best copy you can

Every step after this depends on audio quality. A screen recording, a phone capture of a speaker, or a 96 kbps stream strips out most of what analysis relies on. Ask for the original file. If you can only get a stream rip, note that in your conclusions and expect an inconclusive result.

Aim for 45 seconds or more of continuous music. Skip intros and fade-outs; take a chorus and a verse if you can.

Step 2 — Run a signal analysis

Run the file through the detector and record three things: the classification, the plausible range and the confidence level. Then open the technical details and note the per-segment scores. Segments that disagree wildly are informative in themselves — that pattern is common in hybrid tracks.

Run it twice with different excerpts of the same track. Consistent results across excerpts are worth far more than a single run.

Step 3 — Read the metadata

Metadata is often more revealing than acoustics. Check ID3 or Vorbis tags for encoder strings, creation dates, tool names and comment fields. Some generation and distribution pipelines leave recognisable traces. Note that metadata is trivially editable, so its presence is suggestive while its absence proves nothing.

Where present, Content Credentials or C2PA-style provenance manifests are much stronger evidence than any detector, because they are cryptographic rather than statistical.

Step 4 — Check provenance and release history

• Does the artist have a plausible history, or did twelve albums appear in one month?
• Are there live performances, session photos, stems or dated project files?
• Do release-date, distributor and ISRC records line up?
• Is cover art and artist copy stylistically consistent across releases?
• Does the same “artist” span unrelated genres with identical production?

This evidence is usually decisive in a way that acoustic analysis is not, and it is what journalists should lead with.

Step 5 — Listen critically

Trained ears still catch things detectors miss: lyrics that scan oddly, phrasing that never breathes, transitions that arrive exactly on the bar without any performance friction, instruments that never change articulation, reverb tails that behave identically across sections. Note these as observations, not proof — human production can produce all of them too, especially in programmed electronic genres.

Step 6 — Weigh the evidence together

Rough ordering, strongest first:

1. Cryptographic provenance, or a direct admission
2. Verifiable production artefacts: stems, dated drafts, session files, live recordings
3. Platform and release-history anomalies
4. Metadata traces
5. Detector output
6. Subjective listening impressions

If your only evidence is item five, you do not have a conclusion — you have a hypothesis that needs work.

When the evidence conflicts

A high detector score with a solid provenance trail usually means the detector is reacting to production style, not origin. A low score with no provenance at all means very little, because post-production removes detectable artefacts. In both cases the correct output is “undetermined”, and saying so publicly is far safer than guessing.

Before you publish or accuse

Contact the artist and give them a genuine chance to show their working. Read the results disclaimer first: a detector estimate should never be the sole basis for an accusation, takedown or enforcement action.