Listeners heard AI-labeled music as less expressive in a test

The system

The system here is a two-experiment working paper, 'Humans Perceive AI-Generated Music as Less Expressive than Comparable Human-Made Content,' by Christopher William White, Kavi Kapoor, Nicole Cosme-Clifford, James Symons and Leonardo von Mutius, posted to SSRN on 8 January 2025 and retrieved via an archived SSRN abstract page after the live page returned a bot challenge. The abstract page does not state a peer-reviewed publication venue, so this should be read as a working paper as retrieved, not a confirmed journal result.

What the documents establish

Experiment 1 (N=120) played the same excerpt to different listeners while alternately labeling it human- or AI-composed. The AI label produced lower ratings of expressiveness and emotional response, but enjoyment ratings did not differ. Experiment 2 (N=657) asked listeners to distinguish human from AI excerpts across four genres, using a genuine 'solvable' pairing and a deceptive 'unsolvable' pairing built from two pieces by the same actual source. When sources genuinely differed, listeners relied on audible technical cues; when forced to choose between two excerpts from an identical source, they still projected more emotional quality onto whichever one they judged human. A second listening study, on functional emotional music, reported in its own preprint that participants similarly rated human-composed clips more effective at eliciting a target feeling, adding a second, independently authored data point to the same pattern.

Craft and rights

Separating capability from result matters here: the paper does not find that AI-generated audio is less enjoyable, only that labeling and belief about its origin change how expressive and emotionally engaging it is judged to be. That has a direct rights-adjacent consequence as disclosure requirements for AI-assisted tracks spread across platforms and standards bodies. If a truthful 'AI-assisted' tag depresses perceived expressiveness independent of the audio itself, transparency and commercial reception may pull in different directions, a tension worth naming rather than assuming away. This is an editorial reading of the paper's own belief-manipulation result, not a claim the authors make about disclosure policy.

Outcomes and open questions

As a working paper without a stated peer-reviewed outlet on the retrieved page, White and colleagues' results should be treated as one lab's measured effect, not a settled field finding. The stimuli, exact genres and full statistical results sit in the full paper rather than the abstract; a reviewer should consult it before quoting effect sizes. Whether the labeling effect holds outside a blind-test setting, where listeners usually already know a track's source before pressing play, is untested here.

• Am I rating this track on what I hear, or on what I was told about how it was made?
• Would disclosing a track's AI assistance change how I value or credit it, separate from its audio quality?
• Has this specific finding been replicated or peer-reviewed since the version I am citing?

Two independently authored studies now report the same direction of effect: people judge human-labeled or human-composed music as more emotionally effective, even when the underlying audio does not obviously differ, which keeps the rights and disclosure question live rather than resolved.