You hear a track on a playlist, in a short video or in a shop, and something about it feels a little too smooth. The voice never quite breathes, the lyrics say nothing in particular, and the artist has released forty albums since spring. Is this song SI (AI)? The honest answer is that your ears alone will rarely settle it. This guide explains why the question has become so common, what listening can and cannot tell you, which outside clues are worth checking, how SI music detectors actually work, and where they fail.
Why everyone is suddenly asking
Music generators such as Suno and Udio turn a text prompt into a finished song with vocals, instruments and lyrics in under a minute. The volume that followed is easiest to see on the services that measure it. In a press release dated 20 April 2026, the streaming service Deezer said it receives nearly 75,000 fully SI-generated tracks a day, about 44% of all daily uploads and more than two million a month, up from around 10,000 a day a little over a year earlier. It says it detected and tagged more than 13.4 million such tracks in 2025.
Uploads are not listening, though. The same release puts SI-generated music at roughly 1% to 3% of total streams on Deezer, and says that in 2025 up to 85% of the streams on fully generated tracks were fraudulent, meaning bots playing songs to collect royalties. Those streams are demonetised. In other words, most of the flood is not aimed at your ears at all. It is aimed at the payout system.
Some generated music does reach real listeners. The best known case is The Velvet Sundown. According to The Guardian (14 July 2025), the "band" collected more than a million streams on Spotify within weeks and released two albums in June 2025 in a country folk style. It first denied being an SI creation, then a person describing himself as an "adjunct" member told reporters the songs were made with Suno, and finally the project confirmed it was SI-generated, calling itself "a synthetic music project guided by human creative direction". Listeners who liked those songs had no label to warn them.
Can you hear the difference?
Mostly not, and this is the most important thing to accept before you start looking for clues. In a survey run by Ipsos for Deezer in November 2025, with 9,000 people in eight countries, 97% of respondents could not tell fully SI-generated tracks from human-made music in a blind test. Eighty per cent said fully generated music should be clearly labelled.
That does not mean listening is useless. It means a single impression is weak evidence. Things people often notice, such as lyrics that stay vague, a voice with a strangely even tone, or a mix that sounds polished but flat, also describe plenty of human pop, library music and amateur home recordings. Treat what you hear as a reason to look further, never as a verdict. The same rule applies to images and text, as we explain in our guide to how accurate SI detectors really are.
Check the context before the sound
Because the audio itself is hard to judge, the fastest signals usually sit around the song rather than inside it. None of these proves anything alone. Several together start to tell a story.
- The release pace. A new artist with dozens of albums or hundreds of tracks in a few months is unusual for a human act. Generators make that pace trivial.
- The credits. Look for named writers, producers, musicians and a label or distributor you can find elsewhere. A song with no human name attached anywhere deserves a second look.
- A life outside the streaming page. Real bands usually leave traces: concert dates, interviews, local press, videos of people actually playing. A project that exists only as album covers and a bio is a warning sign, as The Velvet Sundown showed.
- The artwork and photos. Generated band photos often show the usual image tells, which we cover in how to spot SI images. A generated cover does not make the music generated, but it raises the odds.
- Platform labels. Some services now say something, with limits we describe next.
What the streaming platforms tell you
Platform policies differ a lot, and none of them gives a complete answer.
Deezer began tagging fully SI-generated music in June 2025, using a detection tool it launched in January 2025. Its April 2026 release says detected tracks are also removed from algorithmic recommendations and editorial playlists, that it no longer stores high-resolution versions of them, and that its tool can recognise output from prominent generators such as Suno and Udio. If you use Deezer, the tag is the most direct signal any service offers today.
Spotify took a different route. In an announcement dated 25 September 2025, it said it had removed more than 75 million "spammy" tracks in the previous twelve months, announced a stricter policy against unauthorised voice clones, a music spam filter, and support for an industry standard for SI disclosures in music credits developed through DDEX. A later update on the same page says a beta launched on 16 April lets artists show, through their label or distributor, where SI played a role, for example in vocals, lyrics or production. Spotify adds a caveat worth quoting: because it depends on what artists declare, "the absence of a credit doesn't mean AI wasn't used."
Bandcamp went furthest. In a policy post dated 13 January 2026, it said music and audio "generated wholly or in substantial part by AI is not permitted", asked fans to report suspected tracks, and reserved the right to remove music on suspicion of being SI-generated. That makes Bandcamp a place where generated music should be rare, but the rule relies on reports and review, not on a guarantee.
How SI music detectors work
Detection research on music is young, and most of it is public. Two lines of work are useful to understand.
Classifiers trained on real and generated audio
In a 2024 paper with a deliberately double-edged title, "Detecting music deepfakes is easy but actually hard", researchers at Deezer Research showed that a fairly simple model trained on real songs and copies of them reconstructed by generators' decoders reached 99.8% accuracy on their test set. They then spent the second half of the paper on the "hard" part. When they applied common changes the model had never seen, such as a small pitch shift, added white noise or re-encoding at a low bitrate, its performance dropped drastically, with some scores falling to almost zero, meaning generated tracks were classed as real. They also warned about generators the model was not trained on, about scores that are poorly calibrated, and about the need for a way to contest a verdict.
Other teams have since built larger test sets. The SONICS dataset, presented at ICLR 2025, gathers more than 97,000 songs, over 49,000 of them synthetic songs from Suno and Udio, precisely because earlier work focused on fake singing voices over real instrumentals rather than songs generated from end to end.
A fingerprint left by the decoder
A second paper from the same Deezer team, accepted at ISMIR 2025 under the title "A Fourier Explanation of AI-music Artifacts", explains what many of these models are picking up. Generators build a compact internal version of a song, then a decoder turns it back into a waveform by repeatedly upsampling a short signal. The authors prove mathematically that this step leaves small, regularly spaced peaks in the frequency spectrum, related to the "checkerboard" artifacts long known in image generation, and that the pattern comes from the model's architecture rather than from its training data. They tested commercial generators including Suno and Udio and built a very small detector on this cue that matched much larger models in several scenarios.
The same paper is candid about its blind spots. Detection dropped on a Udio version that was not in the training data, which the authors suspect means Udio changed its architecture, and they expect anything that moves frequencies, such as resampling or a pitch shift, to hurt the method. A fingerprint tied to one architecture can disappear the day a generator changes its decoder.
Key point: an SI music detector answers a narrow question, "does this file carry a trace I know?", not the broad one, "did a human make this song?". A negative result does not clear a song. A positive one is strongest when it matches the context.
What our SI music detector can and cannot do
Our SI music detector, part of the SI or Not suite for SI detection, works on the same principle: it measures the regular spectral trace that decoders leave, and it also reads the file's tags, since some generators name themselves in the metadata of their downloads. It is a beta, and we describe its scope plainly on the tool page:
- It recognises Suno well and Udio less reliably. Songs from other generators usually come back with no fingerprint found.
- It needs the audio file itself and at least 30 seconds of it. It cannot read a song from a Spotify, YouTube or SoundCloud page, and a short clip does not carry enough of the pattern.
- It does not specifically detect a cloned or generated voice over a real backing track, or a human singer over a generated instrumental.
- Heavy mastering, effects, pitch or speed changes and repeated compression can weaken or erase the trace, exactly as the research above predicts.
That is why a negative result reads "No SI fingerprint found", never "human-made". We do not publish an accuracy figure for the beta, because any number would depend entirely on the songs it was measured on. Voice clones are a separate problem, and a growing one in scams, which we cover in SI scams: voice cloning, fake photos and fake profiles.
Which signal tells you what
| Signal | What it can tell you | Main limit |
|---|---|---|
| Your own ears | A reason to look closer | 97% of people failed a blind test (Ipsos for Deezer) |
| Release pace and credits | Whether a human act plausibly stands behind it | Prolific humans and anonymous producers exist |
| Deezer tag | A platform's own detection of a fully generated album | Deezer only, fully generated tracks only |
| Spotify SI credits | What the artist chose to declare | No credit does not mean no SI |
| File tags | A generator naming itself | Tags are easy to delete or edit |
| Spectral fingerprint | A known generator's trace in the audio | Known generators only, fragile to edits |
A five-minute routine
- Look at the artist page. Count releases, check the dates, look for named people and a label.
- Search outside the platform. Concerts, interviews, videos of real performances, a social account with a history.
- Check the labels. A Deezer tag, SI credits in Spotify's Song Credits, or a removal from Bandcamp all count.
- If you have the file, run it through a detector and read the clues, not only the score. Use a full song in its original download format when you can.
- Weigh it all together. Agreement between context and a detector is meaningful. A lone score is not.
Above all, do not accuse a musician on the strength of one tool. Independent artists are already competing with a stream of generated uploads, and a false accusation can hurt them far more than a missed generated track hurts you. If the subject interests you more broadly, our piece on SI slop looks at the same flood across images, videos and text.



