Listeners liked the same piano recording less when told it was AI

The audio task

A label deciding whether to disclose that a keyboard part was played by software, or a platform weighing whether an “AI performed” tag will hurt a track's reception, is really asking a question about perception, not audio fidelity: does a listener's belief about who or what performed a piece change how they rate what they hear, even when the sound itself never changes?

What the documents show

A study published in Empirical Studies of the Arts, with the accepted manuscript archived in full on OSF, tested this directly. Alessandro Ansani and six co-authors recruited 120 online participants for a cross-over design: each watched three videos of classic piano performances in two versions with identical audio, one captioned as a professional pianist playing and one captioned as the piano playing “automatically, thanks to an AI.” Because the soundtrack never changed, any rating difference could only come from the caption. The manuscript reports the human-labelled version scored higher for likeability, engagement, emotional valence, and perceived quality, an effect the authors say did not depend on participants' own musical expertise but was moderated by their general attitude toward AI. When asked afterward what differences they had noticed, participants described rhythm, dynamics, and dissonance differences that, by the study's own design, were not actually present, which the authors read as evidence people improvise a technical story to justify a belief-driven reaction.

Rights status

The paper makes no rights or licensing claim; it is a study of listener psychology, not a legal or regulatory document, and neither the manuscript nor its abstract states anything about copyright, consent, or disclosure obligations. What it establishes is empirical: labelling a performance as AI-driven measurably lowers audience ratings of identical audio. That is a finding about listener bias, not a statement that any specific disclosure rule is required, recommended, or prohibited.

What to check before you use it

This is an editorial checklist. Before citing this study to justify a disclosure decision, confirm the manuscript's own limitation: the authors did not compare a real AI-driven performance against a real human one, only belief about a single unchanged recording, so the finding concerns perception under a stated label, not whether listeners can actually detect AI performance in the wild. Confirm the sample was 120 online participants rating three classical piano pieces, not a cross-genre or professional-listener population. Check whether the OSF manuscript text matches the version of record before quoting a specific figure, since an accepted manuscript and a published article can differ in minor ways.

• Does the disclosure decision rest on whether AI performed a part, or on whether a listener merely believes it did?
• Would the same bias appear across genres and performer demographics the study did not test?
• Is a rating drop being treated as proof of detectability, when the study measured belief rather than discrimination?

The result is narrow but useful: a caption alone can move how identical audio is judged, and any policy built on a “listeners can tell” claim should specify whether it means genuine detection or simply the label doing the work.