A peer-reviewed paper by Sarah H. Wu of Stanford University and Kevin J. Holmes of Reed College finds that people engage less with music they believe came from AI, and that the label alone is enough to cause it. Published in March in Cognitive Research: Principles and Implications, the study arrives while platforms and the industry argue over whether AI music should be disclosed, and how.
The paper argues that truthful attribution has real consequences for how music is heard. The researchers ran two preregistered experiments with 399 US participants, measuring what they call narrative listening: whether an instrumental piece makes the listener picture a story. In the first, people heard six human-composed works, including pieces by Beethoven, Mozart, Debussy and Ravel, with no information about their origin. The more a listener suspected a piece was computer-made, the less likely they were to imagine a story, and the flatter those stories were.
The second experiment mixed four human pieces with four made by the AI composer AIVA and tagged each one as composed by a human or by AI while it played. Pieces tagged as AI produced fewer and weaker imagined stories whatever their real origin. The effect was somewhat stronger on genuine AI pieces, which the authors read as listeners also picking up audible traces of machine composition. They link the drop to people assuming AI music carries less communicative intention, so they expect less meaning and intensity from it and let their imagination do less work. In the authors' account, knowing there is a human mind behind a piece is part of what makes it prompt a story at all.
The authors are careful about what this implies. Hiding AI origin might make a track feel richer, they write, but it does so at the expense of human creators, who lose credit and pay for the work AI systems are built on. Their paper opens with The Velvet Sundown, the AI-made act that passed a million monthly Spotify listeners in 2025 before its makers admitted what it was, as an example of AI music slipping by unnoticed. They suggest much of it is designed to smooth away the distinctive qualities of human work, producing a kind of algorithmic easy listening that audiences meet only shallowly.
A separate working paper from economists in Kiel and Hamburg, not yet peer-reviewed, points the same way: listeners, especially pop fans, would pay less for AI music once told how it was made. Its authors conclude that informed choices depend on artists and platforms being open about AI use. Earlier in May, a music supervisor also argued in MBW for a nutrition-label style disclosure system with AI-generated as one of four required origin categories.
The findings arrive as platforms settle on disclosure. Apple Music's Transparency Tags and Spotify's AI credits both rely on labels and distributors to declare AI use, Spotify has barred AI-persona profiles from its new verified badge, and Deezer detects and tags AI tracks itself, reporting about 75,000 fully AI uploads a day. MBW also points to the harms that push this debate: a North Carolina folk singer who found AI-cloned covers on her own Spotify page, Sony's requests to remove more than 135,000 AI impersonations of its artists, and ASCAP's call since 2023 to tell AI and human works apart.
Sources · 1