How to Tell If a Song Is AI-Generated

There is a short version of this answer and a long one. The short version is that by ear, in 2026, you probably cannot, and the research agrees with you: in a survey of 9,000 listeners, 97% could not reliably distinguish AI-generated music from human music.

What actually carries signal

These are the things a detector looks at. Every one of them also occurs innocently, which is exactly why no single item is a verdict.

Room tone and the noise floor

A microphone in a room picks up the room. Air conditioning, traffic, the chair, the performer breathing between phrases. Generated audio often has an unnaturally clean floor, because there was never a room. The catch: a track made entirely inside a DAW with virtual instruments also has no room, and that is most electronic music ever released.

Breath and handling noise

Singers inhale. Guitarists' fingers squeak on wound strings. Drummers hit slightly off-centre. Generated vocals frequently have breath that is either absent or placed too regularly, because it was modelled rather than taken.

Micro-timing

Human performance drifts against the grid in ways that are not random: players rush into fills and lay back on the two and four. Generated output can be too even, or unevenly even in a way that does not correspond to how anybody plays. The catch, again: heavily quantised electronic music is deliberately grid-locked, and gets flagged for it constantly.

Spectral residue from the vocoder

This is the one that is genuinely technical rather than aesthetic. The neural vocoders inside generative music models use upsampling layers, and the mathematics of those layers requires them to leave regularly spaced peaks in the frequency spectrum. It is a property of the architecture, not the training data. It is published, peer-reviewed and reproducible with an ordinary spectrogram.

It is also removable. Pitch shifting moves the peaks. Re-encoding at low bitrate can smear them. Future models can use different upsampling and erase them entirely. Which is why it can be a strong signal and never the only one.

What does not work

• "It sounds soulless." Not measurable, not reproducible, and demonstrably wrong at the rate listeners actually perform.
• Lyrical banality. Plenty of human songs are lyrically banal, and generators have improved fast on exactly this axis.
• Checking the metadata. Stripped by ordinary pipelines and trivially omitted.
• One detector's score, taken alone. Detectors disagree with each other on the same file routinely, which means at least one is wrong every time it happens.

The case that beats almost everything

A real singer over a generated backing track. The mix averages to something inconclusive, the detector shrugs, and the track passes. This is the single most common way detection fails in practice, and the only fix is to separate the stems and score them independently.

Why hybrids break detectors goes through that case properly.