97% of People Fail to Recognize AI Music, Deezer Study Finds

There are many concerns surrounding artificial intelligence and its increasingly elaborate capabilities. A key concern revolves around its growing ability to convincingly mimic reality.

Take video content, for example. In September 2025, OpenAI released Sora 2, a model capable of producing AI-generated videos that appear more realistic than ever. As OpenAI explains, “Sora 2 can do things that are exceptionally difficult, and in some instances outright impossible, for prior video generation models: Olympic gymnastics routines, backflips on a paddleboard that accurately model the dynamics of buoyancy and rigidity, and triple axels while a cat holds on for dear life.” It also lets users place real-world elements into Sora 2, including humans, animals, and objects.

The potential for drastic consequences is clear. The more convincing AI-generated content becomes, the easier it is for bad actors to spread false yet believable information. This can include simple hoaxes created by people who find it entertaining to mislead others, as well as political propaganda designed to polarize and deceive.

Another concern is that people often overestimate their own ability to recognize misinformation. Researchers at the Max Planck Institute for Human Development have confirmed this through an extensive study. Their findings show “no significant impact of education on people’s ability to distinguish between true and false information. This contradicts the widespread belief that more educated individuals are likely to be less susceptible to misinformation, especially as higher education teaches us critical thinking.” The same research also reveals that younger people are, in fact, more likely to fall for false information, despite being considered “digital natives.”

While the study does not focus on AI specifically, which means that we should not directly apply its results to AI-generated content, it offers an important reminder: Everyone can fall for information that appears real.

When it comes to AI, a recent survey by Deezer and Ipsos reinforces the concerns around distinguishing between real and AI-generated content. The study was conducted across eight countries with 9.000 individuals to better understand their attitudes towards and perceptions of AI in music. Participants were asked to listen to three tracks, only one of which was human-made, and decide whether they were fully AI-generated or not. The results are troubling: 97% of respondents failed to provide accurate answers.

As Deezer explains, “A majority (71%) of the respondents were surprised by these results and more than half (52%) felt uncomfortable by not being able to tell the difference.” The findings not only highlight how capable AI has become in such a short period of time, but also underline how important it is not to overestimate one’s ability to detect AI in music (or any other medium and format).

The difficulty of recognizing AI-generated content is already evident in the real world. A striking example is The Velvet Sundown, a band entirely created by AI, from their music and backstory to their images. According to The Guardian, the band initially denied the claims multiple times before ultimately admitting it was an AI creation. One Reddit user wrote: “I’m ashamed to say last week I listened to ‘Floating on Echoes.’ It was in my Discover Weekly on Spotify and it sounded OK. Spotify can’t be promoting this over actual up-and-coming artists (...).”The story illustrates just how much AI can accomplish, and how easy it is for content to appear convincingly real.

After examining people’s ability to distinguish real music from AI-generated content, Deezer and Ipsos investigated participant's attitudes toward AI in music. According to the data, 51% believe AI will play a significant role in music production over the next ten years, and many expect that its use will lead to more low-quality, generic-sounding music on streaming platforms.

In addition, 45% of music streaming users say they would like to filter out fully AI-generated music on their platform, while 40% say they would skip such tracks without listening if they came across them. 80% believe AI-generated music should be clearly labeled for listeners, and 52% think it should not be included in charts.

Copyright remains a major concern for artists and labels, and respondents agree that fairness toward creators matters. For instance, 65% say companies should not be allowed to use copyrighted music to train their AI models, and 73% believe it is unethical for AI companies to use copyrighted material to generate new music without clear approval from the original artist.

From a financial point of view, 70% believe that fully AI-generated music threatens the livelihood of current and future musicians, while 69% believe payout rates for AI-generated tracks should be lower.

While The Velvet Sundown shows how difficult it can be to identify AI music, especially when there’s a full backstory and images, there are practical steps anyone can take to improve their detection skills.

1. Practice mindfulness. Pay close attention to what you’re listening to. Notice if anything feels generic, overly polished, or just doesn’t “click” with the emotion you’d expect from human performance.

2. Check the track’s origin. Look up the artist, release history, and context. Fully AI-generated bands often have minimal or inconsistent backstory.

3. Look for labeling. Platforms like Deezer are starting to label AI-generated music. Always check for official indicators.

4. Notice patterns in popularity. If a track or band gains massive streams very quickly with little exposure, it could be AI-generated.

5. Examine visuals. Album covers, promotional images, or band photos can give clues. AI-generated visuals are often slightly off, overly polished, or inconsistent.

6. Use community feedback. Forums, social media, and reviews often discuss suspicious or AI-generated tracks. If multiple listeners point it out, it’s worth investigating.