How to Detect AI Generated Music

Why this is harder than it sounds

Two years ago, most generated music announced itself. Vocals had a smeared, underwater quality; instruments blurred at transients; and the high end fell off a cliff somewhere around 14 kHz because the models were generating into compressed representations. You could often tell within five seconds.

That era is over. Current systems produce output that survives casual listening, and increasingly survives careful listening too. Meanwhile the boundary itself has dissolved: a track might be generated end to end, generated and then re-sung by a human, human-performed and then processed through a generative stem separator, or written by a person who used a model only for a bridge they were stuck on. 'AI or not' is a binary question about a continuous reality.

This matters for detection because every method below is really answering a narrower question: does this recording carry the statistical fingerprints of a generative pipeline? A track can be substantially AI-authored and carry none of them, and a fully human track can carry several.

Step one: listen deliberately

Human listening remains genuinely useful, provided you listen for structure rather than for a vague sense of wrongness. The 'it just feels off' instinct is unreliable and heavily biased by what you already suspect about the artist.

Work through a checklist instead. Listen once for the vocal, once for the arrangement, once for the mix. Each pass catches different things, and writing down what you noticed before you run any tool stops the tool from anchoring your judgement.

• Lyrics that scan perfectly, rhyme cleanly and say nothing specific — no place names, no proper nouns, no detail a person would actually choose
• Vocal takes with no breath noise between phrases, no consonant grit, and identical phrasing across repeated lines
• Instruments that never change articulation: a guitar that strums with the same attack for three minutes, drums with no ghost notes or velocity drift
• Sections that repeat with machine exactness, or transitions that arrive without any performative lead-in
• A mix that is loud and full but strangely flat — everything present, nothing foregrounded
• Endings that fade or stop rather than resolve, and song lengths that cluster suspiciously around a default

Step two: measure the signal

Acoustic detection converts the recording into numbers and asks whether those numbers resemble patterns common in generated audio. The detector on this site samples several sections of the track, runs a Hann-windowed FFT over each, and derives six measurements: the spectral ceiling, cross-segment tonal agreement, crest factor, spectral-centroid variability, high-band energy ratio and stereo correlation.

Each one is a proxy for something. The spectral ceiling asks whether the top end stops abruptly at a suspiciously round frequency. Cross-segment agreement asks whether the verse and the chorus have improbably identical tonal balance — real performances drift. Crest factor asks how much dynamic range survives. Centroid variability asks whether brightness moves the way a performance moves. Stereo correlation asks whether the image was constructed or captured.

None of these is decisive alone, and all of them are contaminated by production. That is the whole difficulty: mastering does to a human recording much of what generation does to a synthetic one. A heavily limited EDM master and a generated track can be nearly indistinguishable on crest factor, because both have had their dynamics deliberately destroyed.

This is why the result you get is a range rather than a number, why it carries a separate confidence level, and why an explicit Inconclusive outcome exists. A detector that never says 'I don't know' is not more capable — it is less honest.

Step three: demand provenance

Provenance beats everything else combined, and it is astonishing how rarely people ask for it before reaching for a detector. If a track is human-made, there is almost always a trail: a project file with a hundred takes in it, stems, a rough mix from three weeks earlier, a phone recording of the writing session, a bounce with a different arrangement.

Ask for it. Ask what DAW, what plugins, who played what, when it was written, what got cut. A person who made a song can answer those questions in detail without preparation. Someone who typed a prompt usually cannot — and someone who typed a prompt and then genuinely produced the track around it will tell you exactly that, which is a perfectly reasonable answer.

Context helps too. An account that posted forty polished tracks across three genres in a fortnight is telling you something no spectrum analyser can. So is a back catalogue with a coherent voice, a live history, or collaborators who can vouch for the sessions.

Step four: check metadata and watermarks

Some generators embed provenance metadata, and some participate in content-credential schemes such as C2PA. When present, that signal is far stronger than anything acoustic — it is a declaration rather than an inference.

The catch is that metadata is trivially removed. Re-encoding a file, running it through a DAW, or uploading it to almost any platform strips or rewrites the tags. So metadata is close to conclusive when it is there and tells you essentially nothing when it is absent. Never treat missing credentials as evidence of anything.

Audio watermarking is more robust in principle, but it only works if the generator applied one, the scheme is public, and you have a detector for it. Coverage across the ecosystem is patchy, and re-recording or heavy processing degrades most schemes.

The five mistakes people make

Almost every bad AI-music accusation comes from the same small set of errors. Recognising them is more valuable than any additional detection technique.

• Treating a percentage as a verdict. An estimate of 78% is not a finding of guilt; it is one weak input.
• Running one tool and stopping. Different detectors disagree constantly, and agreement between two is weak evidence when both measure similar things.
• Ignoring confidence. A high probability with low confidence is much weaker than a moderate probability with high confidence.
• Testing a degraded file. A 96 kbps rip from a video has already lost the high-frequency detail the analysis depends on.
• Assuming AI use is misconduct. It is legal, widely disclosed, and frequently part of an otherwise entirely human process.

A workflow you can actually follow

Start with the best copy of the file you can get and listen through it once, deliberately, noting anything from the list above before you run anything. Then run an acoustic analysis and read the whole report rather than the headline number — the confidence level and the per-measurement reasoning tell you far more than the percentage.

If the result is inconclusive, that is usually the correct answer and you should stop rather than shop for a tool that will give you a firmer one. If it leans generated, go looking for provenance: ask for stems, ask about the session, look at the release history. If provenance is offered and coherent, believe it. If it is refused with no explanation, you have learned something a detector could never tell you.

And decide in advance what you would do with the answer. If the honest response to 'this is probably AI' is 'then I would not add it to my playlist', a probabilistic tool is entirely adequate. If the honest response is 'then I would terminate a contract', no acoustic tool is adequate and you need documentary evidence.