Most listeners in a Deezer survey missed AI-made music entirely

The tool moment

On 12 November 2025, Deezer published a newsroom report on a survey it had commissioned from Ipsos, described as “the world's first survey focused on perceptions and attitudes towards AI-generated Music.” Ipsos surveyed 9,000 people across the United States, Canada, Brazil, the UK, France, the Netherlands, Germany and Japan. As part of the exercise, participants were asked to listen to three tracks, two generated by AI and one recorded by a human, and to identify which was which. According to Deezer's report, 97% of respondents got it wrong.

What the documents show

Deezer's report states that 71% of respondents “were surprised by these results” and 52% “felt uncomfortable by not being able to tell the difference.” The same report says a majority of respondents wanted AI-generated music clearly labeled and were concerned about fair payment to artists whose work trains AI models. This is a company-commissioned survey, run by Ipsos on Deezer's behalf and published on Deezer's own newsroom, not an independently designed academic study; Deezer has a direct commercial interest in the transparency argument the survey supports. Deezer's newsroom separately covers the company's own AI-detection and tagging system as an ongoing product feature, distinct from this survey's findings about listener perception.

What stays with the musician

A 97% miss rate in one blind test with three tracks says something about that specific test, not about every listening context a musician's own audience might use, background listening on a phone speaker versus close listening on studio monitors, for instance. A musician using AI-assisted tools still has to decide, independent of this survey, what to disclose to their own audience and how, since the survey itself finds an appetite for labeling regardless of detection ability. The finding is about identification, not about whether AI-generated or AI-assisted output is acceptable to any given audience.

Judge it by listening

The report does not publish the three tracks, the exact instructions given to participants, or details of how “fully AI-generated” was defined and verified for the test, so a reader cannot rerun this specific experiment. What can be checked independently is the general pattern: short, casual listening comparisons of AI-generated and human-made tracks in a reader's own genre, noting which cues, if any, actually distinguish them, rather than assuming either that detection is easy or that it is impossible based on one company's survey alone.

• What exactly counted as “fully AI-generated” in this test, and how was that verified?
• Would the same miss rate hold for genres or production styles this survey did not test?
• Does this outlet's or artist's own audience want labeling regardless of whether they could detect AI involvement themselves?

A large, company-commissioned survey is real evidence about self-reported attitudes and one blind test's outcome. It is not a peer-reviewed study, and its sponsor has a stated interest in the transparency conclusion it reaches.