Human music beat AI at eliciting a target feeling

The system

The system is again a listening study, not a music tool: 'Understanding Listener Perceptions of AI and Human-Composed Music in Emotional Applications,' by Kimaya Lecamwasam and Tishya Ray Chaudhuri, an arXiv preprint first submitted 3 June 2025 and revised through a September 2026 version. It studies functional emotional music, tracks built to elicit a specific target feeling such as calm or upbeat energy, framed for affective computing and wellness-technology design.

What the documents establish

According to the full preprint text, 152 participants heard four one-minute instrumental clips, two human-composed and two AI-generated, across Calm and Upbeat cases, split into Unlabeled, Correctly labeled and Incorrectly labeled groups. Labeling significantly predicted stated preference (a chi-square test gave 16.51 on 4 degrees of freedom, p=0.002), with participants told an AI piece was human showing higher AI preference than those told the truth. But on efficacy, how well a clip elicited its target feeling, labeling had no significant effect (p=0.259), and all three groups rated human-composed clips more effective overall, by margins of 61 to 63 percent, with no significant differences between groups. A separate GEMIAC emotion-profile measure did not differ significantly by origin in category terms, though its intensity scores ran higher for human music in specific cases, so the paper's own efficacy ratings and its GEMIAC scores are reported as related but distinct findings.

Craft and rights

This separates what a label does to preference from what a track actually does functionally: belief about origin swayed which clip people said they liked, but did not change their judgment of which clip worked at its stated emotional job. For anyone selling music as functional, calming, focus, wellness, that is a meaningful split between marketing effect and measured outcome. The preprint's qualitative theme ties perceived 'humanness' to imperfection and flow rather than to any acoustic measurement, a subjective attribution rather than a property of the audio. Whether wellness and mental-health apps built on AI-generated soundtracks should disclose that origin, given this paper's own finding that mislabeling shifts preference, is an editorial question the study raises but does not answer.

Outcomes and open questions

The authors' own limitations section calls for future work with clinical populations, real-time physiological markers, or longer exposure, meaning this result rests on self-report from a single sitting, not biometric evidence of emotional response. As a preprint still being revised as of the version read here, it has not gone through the field's normal peer-review cycle on the record retrieved. Whether the efficacy gap and the preference-labeling effect both hold outside a lab listening task remains open.

• Am I choosing this track because it works for its stated purpose, or because of what I was told made it?
• Would I want a wellness app to disclose whether its soundtrack is AI-generated before I use it?
• Has this preprint's efficacy finding been replicated with physiological rather than self-reported measures?

The clearest fact this preprint establishes is a gap between belief-driven preference and self-rated functional effectiveness, with human-composed clips holding an edge on the latter that labeling alone did not erase.