A preprint ties believed authorship to a track's expressiveness

The tool moment

A working paper titled “Humans Perceive AI-Generated Music as Less Expressive than Comparable Human-Made Content” was posted to SSRN on 8 January 2025 by Christopher White and four co-authors. The paper describes two experiments. In the first (N=120), participants heard the same musical excerpt while being told, alternately, that it was composed by a human or by AI. In the second (N=657), participants tried to distinguish human-composed from AI-generated excerpts across four genres, sometimes comparing genuinely different sources and sometimes comparing two excerpts from the same source under a deceptive label.

What the documents show

The abstract states that in the first experiment, AI-attributed excerpts “reported the music to be less expressive and to evoke less of an emotional response, but also to be equally as enjoyable as the human-made content,” a more limited effect than a straightforward preference for human-labeled music. In the second experiment, when the two excerpts genuinely came from different sources, participants “relied on technical musical characteristics for identification,” but when forced to compare two excerpts from the same source, they “projected more emotional qualities onto the music they labeled as human.” A Crossref record confirms the paper is posted content on SSRN, not a peer-reviewed journal article, as of this writing.

What stays with the musician

The paper's own framing is that beliefs about creative agency, not necessarily audible differences, shape emotional engagement with music. That leaves a musician's disclosure choices doing real work: how a track is labeled or credited may shape a listener's felt response independent of anything in the audio itself, according to this preprint's own account. Because the paper has not yet cleared peer review, a musician or outlet citing it should treat its specific effect sizes and genre comparisons as provisional rather than settled science.

Judge it by listening

The second experiment's own design is instructive for a reader: the paper reports that listeners could pick up on audible differences when two excerpts truly came from different sources, using technical musical cues rather than a gut feeling about humanness. That suggests any home comparison should keep the source pairing honest, genuinely different tracks against genuinely different tracks, rather than relying on a single labeled pair, before concluding that a listener's ear or their assumptions did the discriminating.

• Is a given comparison testing whether AI-generated and human-made music sound different, or testing what a label alone does to perception?
• Has this specific finding been replicated in a peer-reviewed venue since this preprint was posted?
• Which genres, if any, does the reported effect hold up across, based on the paper's own four-genre design?

A preprint with a five-author byline and two experiments is a real data point, not a settled result. Its own abstract draws a careful line between audible difference and labeling effect that is worth preserving in any summary of it.