Does this sound like AI?

Step one: what are you actually listening to?

• A complete song — that's the job of the free AI music detector on this site. It forwards the audio to a specialist classification service and discards it once analysis is done.
• A voice, a voice note, spoken words — different problem entirely. See what actually gives an AI voice away.
• A short clip ripped from a social video — track down the original file first. Re-capturing audio from a video mostly tells you about the encoder, not the music.

Cues that generally hold up

None of these proves anything by itself. A cluster of them in one recording is a reason to dig further.

• Lyrics that scan flawlessly but say nothing. Clean meter, generic imagery, not one specific place, person, date or detail across three minutes.
• Transitions that are too clean. Sections change exactly on the bar, with no fill, no push, nothing that reads as a player's decision.
• Instruments frozen in one gesture. The same guitar attack, the same hi-hat, the same piano touch from the first bar to the last.
• Delivery with no variation. Repeated lines phrased identically, breath that never changes, energy that neither builds nor fades across a verse.
• A tone that never shifts. Verse, chorus and bridge with the exact same spectral shape, as if the whole song were rendered in a single pass.

Cues that trick almost everyone

These are the reasons careful listeners end up accusing human artists. Every one of them is a standard production decision used on most commercial releases.

• Loud and flat. A squashed dynamic range is aggressive mastering, not proof of synthesis.
• Silence above 16 kHz. That's a lossy encoder's ceiling. It describes how the file was compressed, not how the music was made.
• Rigid timing. Quantised programming has been the norm for three decades.
• Flawless pitch. Pitch correction sits on nearly every commercial vocal released today.
• Repetitive song structure. Standard practice in template-driven pop, house, drill and library music.

A retired internal research build returned elevated historical scores for three of twelve procedural synthesiser references. That does not measure the current provider — see the electronic false-positive study. If you're listening to techno, trance, hyperpop or a game soundtrack, raise your bar before you conclude anything.

What actually carries weight after that

1. Provenance. Stems, a project file with a believable edit history, an unmastered rough mix, dated drafts, session footage. Any single one outweighs every measurement this site can produce.
2. Release context. Twelve releases in a month across unrelated genres, all mastered identically, with no live footage anywhere — that tells you more than a spectrum analysis ever could.
3. Metadata. Tags frequently retain encoder strings and tool names nobody bothered to strip. Content Credentials or a C2PA manifest is cryptographic, not statistical, and it outranks everything else on this list.
4. Acoustic analysis. Run it against two separate excerpts and compare — that's where most of the variation in a single reading comes from.
5. Your ears. The thing you noticed first and should weigh last, because a number anchors perception harder than people expect.

The complete process, mistakes and all, is written out as a ten-step walkthrough.

How to phrase it without overreaching

“This audio shows characteristics associated with generated music, and no one has provided provenance when asked” is a defensible sentence. “This song is AI” is not, and it's the one that gets quoted back at you later. When your signals disagree, undetermined is a legitimate outcome — most single-track investigations honestly land there.

Questions people ask

Want an independent classification to place beside the available evidence? Run a free check — no account required, and we never write your audio to storage.