A working paper ties an AI label to lower expressiveness ratings

The audio task

Anyone building a catalogue tag, a chart rule, or a marketing line around “you can't tell the difference” is making a claim about listener discrimination that needs its own evidence, separate from whether a track sounds technically competent.

What the documents show

A working paper titled Humans Perceive AI-Generated Music as Less Expressive than Comparable Human-Made Content, posted to SSRN on 8 January 2025 by Christopher William White of the University of Massachusetts Amherst and four co-authors, reports two experiments. In the first (N=120), participants heard the same excerpt and were told, alternately, that it was human-composed or AI-generated; the AI label produced lower ratings for expressiveness and emotional response, though the excerpt was rated equally enjoyable either way. In the second, larger experiment (N=657), participants tried to tell human and AI excerpts apart across four genres. When a pair genuinely came from different sources, participants used audible technical cues to identify them, which the authors take as evidence real acoustic differences existed. But when both excerpts in a pair secretly came from the same source, participants still projected more emotion onto whichever one they had guessed was human. The paper frames this as evidence that belief about creative agency shapes perceived expressiveness independent of the audio itself.

Rights status

The document is a working paper posted to SSRN's repository; its own DOI record classifies it as posted-content of subtype preprint, meaning it has not gone through peer review at the point of this retrieval, and it makes no legal or licensing claim. It should be read as an early research finding open to revision, not a settled scientific result, and it says nothing about who owns or may license AI-generated output.

What to check before you use it

This is an editorial checklist. Verify the paper's peer-review status has not changed since its January 2025 posting before citing it as settled research. Note the two experiments measured different things: Experiment 1 measured belief-driven bias on the same excerpt, and Experiment 2 measured actual discrimination ability across genuinely different sources, so a summary that blends the two risks overstating what either one shows. Check the reported sample sizes, 120 and 657, and the four tested genres before generalising to material the paper did not cover.

• Is the claim about listeners telling AI and human music apart, or about a label changing how they rate it?
• Has this working paper since appeared in a peer-reviewed venue, and did the findings change?
• Do the four tested genres resemble the material the marketing or policy claim is actually about?

Read together, the two experiments suggest a listener's belief about who made a piece of music does real work in how expressive it sounds to them, but that belief effect is a different claim from being unable to tell AI and human music apart at all.