How can you tell if a song was made by AI?

In a few seconds, Suno or Udio produce a complete song: verses, chorus, vocals, mix. The result is often stunning, and generated tracks now circulate on every streaming platform, sometimes under the name of “artists” who do not exist. How can you tell the difference? Here is a three-step method: listen, investigate, analyse.

1. Listen: clues you can hear

None of these clues is proof on its own, but when they add up, you should be alert.

The voice

• Lyrics that “slide”: swallowed syllables, invented or badly articulated words, especially at the end of phrases and in languages other than English.
• Breathing that is missing or oddly placed: a singer breathes at logical moments; a generated voice sometimes strings together phrases impossible to hold in one breath.
• A timbre that changes from one verse to the next, as if several singers were taking turns.

The instrumental

• A “veiled” sound in the highs: cymbals and hi-hats lack definition, with a slight metallic or “hissing” grain.
• Instruments that merge: it is hard to follow a guitar or bass line from start to finish; it seems to dissolve into the mix.
• A structure that is too perfect… or inconsistent: a chorus that comes back identical, or on the contrary an ending that stops dead, a bridge that leads nowhere.

2. Investigate: the artist and the file

• Does the artist exist? Gigs, interviews, consistent photos, musician credits, long-standing social media… A “band” with no trace outside the platforms, releasing an album a month, should raise suspicion. In 2025, the case of The Velvet Sundown, which had built a large audience on the platforms before its artificial nature was acknowledged, showed that it doesn't only happen to others.
• The release pace: dozens of titles in a few weeks is rarely human.
• The file metadata: if you have the file, its information (encoding software, comments) sometimes mentions the generator. Careful: it can be erased in one click, so its absence proves nothing.
• Platform labels: some streaming platforms have started labelling AI-generated content. But labelling is far from systematic.

3. Analyse: what a detector sees

Music generators do not “play” real instruments: they directly produce the audio signal. This process leaves traces in the sound spectrum, inaudible but measurable, which a model can learn to recognise by studying tens of thousands of real and generated examples.

The trap is hybrid tracks: a real singer on a generated instrumental, or a generated voice on real music. Analysed as a whole, this kind of track can get a low score. That is why our detector separates the vocals and the instrumental before analysing each one on its own.

Key takeaways

• Your ear gives leads, but recent generators have become very convincing.
• Investigating the artist is often as telling as listening.
• A detector gives a probability, to be cross-checked with the rest. No tool is 100% reliable (here is why).