Do people prefer AI-generated music? A useful answer needs evidence about what listeners choose. A test of whether they can identify how a recording was made measures a different response.
That distinction matters when reading the Deezer/Ipsos survey alongside the blind-voting results published by music discovery platform Kallitone. Both involve listening, but the questions and methods differ. Reading them carefully opens a more interesting discussion about what makes an unfamiliar song worth hearing again.
What the Deezer/Ipsos listening test measured
In November 2025, Deezer published an Ipsos survey involving 9,000 adults across eight countries. Participants heard three recordings, comprising two AI-generated tracks and one human-made track, and were asked to identify their origins. Deezer reported a 97% failure rate on that classification task. [1]
The finding concerns recognition within that particular test. It does not establish that 97% preferred the AI tracks, that every answer they gave was wrong, or that the result would be identical with other recordings.
Imagine hearing an unfamiliar song and being unsure whether software generated it. You might still dislike the chorus. You could also enjoy it while remaining uncertain about its origin. Identifying a production method and responding to a piece of music are separate tasks.
What Kallitone reported from its beta
Kallitone uses paired listening comparisons to help organise music for discovery. Artist identities are hidden during voting. Listeners hear passages selected randomly from within two songs, with a standard minimum of 15 seconds per track, then choose their preference. Votes feed an Elo-style rating system. [2]
On 26 October 2026, Kallitone reported 25,000 completed comparisons across 500 tracks in its first-year beta. The catalogue included 460 tracks by independent musicians and 40 identified as AI-generated. Each track participated in 100 comparisons. All 40 identified AI tracks finished below the reported catalogue average Elo rating of 1,250; the highest scored 1,244. [3]
The comparison count covers the whole catalogue. It is neither a count of individual listeners nor a claim that 25,000 dedicated human-versus-AI trials took place.
These are platform-reported beta observations. They describe the participating recordings and listeners, and should not be treated as a representative verdict on all AI-generated music.
Why the two findings can coexist
A listener could struggle to identify AI-generated recordings while consistently preferring other songs in a particular selection. There is no contradiction: the first response concerns a guess about origin; the second concerns a choice between recordings.
The distinction becomes clearer when we separate three questions:
| Question | What the answer can tell us |
|---|---|
| How do you think this recording was made? | The listener's ability to identify its origin under the test conditions. |
| Which of these two recordings do you prefer? | A relative preference between the options presented. |
| Which recording do you return to, save or seek out? | Whether that initial encounter develops into further interest. |
Someone designing a listening study must decide which response matters. Keeping that choice visible helps readers understand what a finding actually establishes.
The Deezer/Ipsos survey and Kallitone's beta also used different recordings, audiences and procedures. Placing the findings alongside each other helps explain the distinction between recognition and preference. It does not turn them into a single controlled experiment.
What a broader preference study could investigate
A follow-up study could ask whether preferences vary across musical styles, generation tools and listening conditions. It could recruit a defined listener sample, document how recordings were selected, and distinguish between reactions to short excerpts and complete songs.
It could also compare responses before and after production information is disclosed. If a listener changes their assessment after learning how a recording was made, that would raise a further question about the role of context in musical judgement.
For a platform comparison, useful reporting would include the number of unique voters, the pairing process, repeated participation, track provenance and uncertainty around the results. A large vote count alone cannot answer all of those questions.
Why this matters for independent music discovery
For an independent musician, the practical challenge is earning attention from someone who has never heard their name. A convincing recording still needs a listener who wants to stay, return or share it.
Kallitone's contribution to that problem is a specific discovery format: an unfamiliar song can enter a listening comparison before the voter sees the artist's identity. The resulting ratings organise choices made by participants. They do not remove taste or turn a short encounter into a complete assessment of artistic value. [2]
The beta offers a concrete case for examining what listeners choose when some familiar presentation cues are withheld. Its significance will depend on how those encounters develop as the catalogue and community grow.
You can explore Kallitone's current rankings and form your own response to the music. Positions can change as voting continues.
A useful next question for the AI-music debate is what happens after a recording passes as plausible: does anyone want to hear it again?
Sources and methodology notes
- Deezer/Ipsos survey on AI-generated music, published 12 November 2025. Primary source for the survey and its reported classification result.
- How Blind Music Voting Works: Can Songs Be Judged Without Popularity? Kallitone's explanation of its voting format and limitations.
- Can Human Judgment Scale? Kallitone Reports 25,000 Blind Music Comparisons, published 26 October 2026. A press release issued by Orkaan Studio S.R.L. (Kallitone), rather than independent reporting.
Sources checked on 31 October 2026. Published by Kallitone; beta figures originate in the platform's own report.