AI music detection is the analysis of a recording’s audio to estimate whether an artificial intelligence (AI) system generated it. Knowing how AI music detection works is what separates a useful result from a misleading one.
This guide covers where the signal comes from, what a likelihood means, why Unsure is a valid answer and why a person should decide. SEAM, short for Sonic Evidence of AI in Music, is the worked example: an audio-recognition service from The InterSpace Group, introduced in our meet SEAM launch article.
How AI music detection works: the signal is in the audio
Detectors analyse the audio itself for traces that generation systems leave behind, then return a likelihood rather than a certainty. An academic review of AI-generated music detection describes the traces as statistical patterns too subtle to hear, such as characteristic frequency patterns, left during synthesis.
One approach turns the audio into a spectrogram, a picture of sound in which time runs across, frequency runs up and brightness shows energy. In one of the first published music deepfake detectors, amplitude spectrograms gave the strongest results among the inputs tested when a model learned to tell real recordings from machine-rebuilt copies.
Watermarks, markers a generator embeds on purpose, are a second route, but they help only when the generator adds one and the file keeps it. The review notes they can be deliberately removed or degraded.
Metadata is weaker still. A title, credit line or “no AI used” declaration records what someone typed, and Spotify, which depends on artist disclosure, says “the absence of a credit doesn’t mean AI wasn’t used.”
That is why SEAM reads the audio itself, not the metadata and not the uploader’s word. A credit line can be changed in seconds without touching the sound.
What a likelihood actually means
A likelihood is a strength-of-evidence reading: how strongly the audio points one way. It is not the share of a track that is machine-made, and it says nothing about intent or who is responsible.
A precise-looking number can still mislead. The authors of that early detector wrote that “a good test score is not the end of the story,” then listed what can go wrong once deployed: scores that may not mean what they seem, edited audio, unseen generators, and results nobody can explain or contest.
In their tests, performance dropped sharply after pitch shifts, added noise and low-bitrate re-encoding such as MP3. So scan the best file you have, ideally the original master.
How to read each verdict, including Unsure
SEAM returns one of three verdicts per track, with a likelihood:
- Likely human: the audio points toward a human-made recording. It is not a clearance for rights or originality, so your normal checks still apply.
- Likely AI: the audio points strongly toward AI generation. Listen closely, ask for disclosure and apply your own policy.
- Unsure: the audio does not settle it. Listen, then ask for stems or session files and a short note on how the track was made.
Unsure is information, not a malfunction. The review notes that real production “rarely yields a single global label”: AI-generated accompaniment can sit under a human vocal, and remixing or mastering can make a single yes-or-no answer insufficient.
Spotify describes AI use as “increasingly a spectrum, not a binary.” A detector that never says Unsure has not removed that uncertainty. It has hidden it.
Why a person should make the final call
A score cannot know your policy, your risk or the artist’s side of the story. The costs of being wrong differ: a wrong “AI” call lands on a real artist, while a wrong “human” call lets a breach of your rules through.
Detection and disclosure answer different questions: disclosure records what the artist says, and detection reads what the audio shows. Where the two disagree, ask.
InterSpace Distribution’s mandatory AI disclosure policy applies to every new release and asks artists to say whether, and how, AI was used in the artwork or the audio. Spotify has said it will support an industry standard for AI credits developed through DDEX (Digital Data Exchange).
Every SEAM verdict is advisory: it never rejects, holds or approves a track by itself, and a person decides. Those authors also flagged contesting a result as a weak point, so keep a record of the file, the verdict, the date and your reasoning, and let the artist respond.
“Likely” is a reason to look closer and ask for disclosure. It is never proof you can publish about someone.
SEAM runs in a web console and through an API (application programming interface, the way software talks to software), and the first 1,000 tracks are free. Run a few of your own tracks through SEAM and see how each verdict reads.
