How to tell if a song is AI generated

Where an answer is actually recorded

• Deezer tags albums that contain fully generated tracks. The tag is applied at album level rather than track level to cut down on mistakes, and it goes on whether or not the uploader declared anything.
• Spotify is adding AI disclosures to song credits, using a standard developed through DDEX, so a credit can say that AI appeared in the vocal, in the instrumentation or in post-production. It arrives release by release as distributors start sending the field, which began in 2026 with a small number of them.
• YouTube asks creators to disclose realistic synthetic content, and the disclosure appears in the expanded description.
• The release paperwork. A label, a distributor, a producer and a registered songwriter are all things a person has to supply, and a large share of generated uploads carry none of them.

Two checks you can run yourself, both free

Neither belongs to us. We would rather send you to the tools that read audio than print a verdict on a format we do not analyse.

• SynthID Detector, Google's portal. It reads the inaudible watermark Google embeds in audio from its Lyria music model and from NotebookLM's audio overviews, and it marks the segments where it found the watermark instead of returning one number for the file. Google says the mark survives MP3 compression, added noise and speed changes.
• Deezer's AI music detector, opened to everyone in June 2026. You connect a playlist from any of about twenty streaming services, competitors included, and it reports how much of that playlist its model reads as fully generated. Deezer states 99.8% accuracy for it. That is Deezer's figure, measured by Deezer, on material Deezer chose.

How much generated music is being uploaded

Deezer publishes the number, and it has moved quickly. In January 2026 the company put fully generated tracks at 39% of daily uploads. By April it was 44%, roughly 75,000 tracks a day. By the middle of the year it reported close to 90,000 a day, which carried it past half of everything delivered on the busiest days.

Listening did not follow. Deezer puts streams of those tracks at a low single-digit share of total listening, and says about 85% of that listening is fraudulent, which is why it stopped paying royalties on it. The flood is aimed at payout systems rather than at audiences. That is worth holding on to before you assume the song you liked last week is part of it.

What listening gets you, honestly

• Voices in no particular room. Breaths in odd places, or no breaths, and a reverb tail that does not match any space the singer could be standing in.
• Lyrics that scan and do not land. Exact rhymes every time, lines of even length, and almost no proper nouns, because a model with no life behind it writes about feelings rather than about a street.
• Detail collapsing in the dense passages. Cymbals turning to hiss, a guitar losing its pick attack, consonants smearing when several parts arrive together.
• Arrangements that never take a risk. Sections turn on the bar you expect, nothing is held a beat too long, and the tempo does not breathe.
• Endings. A fade that was cut rather than played, or a stop that lands as though the file simply ran out.

Questions people ask before they argue about it

Why we do not offer an audio check

Deepfake-Eval-2024 gathered media that genuinely circulated during 2024, including 56.5 hours of audio, from 88 websites in 52 languages. Open-source audio detectors lost roughly 48% of their area under the curve on that material compared with the academic datasets those same models score well on. The free audio detectors offered to the public are largely those models behind a web form.

Vendors selling into the music industry publish better numbers. Ircam Amplify, spun out of the French research institute of the same name, states 99% accuracy with under 1% false positives for its music detector. Like every figure in this market it is the vendor's own, produced on material the vendor selected, and there is no published test set anyone can reproduce it against. We publish no accuracy number for any detector until our own bench has run.

When the audio gives you nothing

• Read the artist page instead of the track. Three hundred songs, no photographs, no live dates, and a catalogue that starts in March is a stronger signal than anything in the waveform.
• Search the songwriter registrations. ASCAP, BMI and PRS all run public repertory search, and a released song by a working writer usually has a registration with human names on it.
• Read the credits. A producer, a mixing engineer and a studio are people, and people can be looked up.
• Ask for the stems. Whoever recorded the song has them, and whoever generated it does not.