A User-Centered View of AI Music Description

Oh, Eun Ji; Park, Sangheon; Kim, Yonghyun; Beck, Jun Woo

Abstract

Recent audio-language models produce fluent, detailed natural-language descriptions of music, yet evaluation still relies on reference-caption corpora rather than potential users. We conducted an exploratory online survey (n = 82) about interest in AI music description (AIMD). Interest was low in this sample (M = 2.39 on a 7-point scale), and only 25.6% had ever obtained such a description using AI. Interest did not differ significantly by musical expertise, but it was higher among participants with prior AIMD experience and, within that group, correlated with how satisfying that experience had been. AIMD users and non-users also differed in envisioned purposes and free-text prompts: prior AIMD users specified musical content they wanted, such as genre or instrumentation, while non-users more often expressed skepticism or declined to use AI. Together, the results suggest a gap between AIMD research and its prospective users. People want music descriptions for different purposes, so model outputs should be evaluated against those purposes, not only against reference captions.

The study

We recruited participants from online communities, including music- and AI-related subreddits and music-technology forums. The survey took a median of 4.5 minutes and offered no compensation. The Georgia Tech IRB approved the study, and all participants gave informed consent.

Sample: Of 149 responses, we excluded 63 incomplete submissions and 4 that treated the task as music generation, leaving 82 participants. Their mean age was 32.4 (SD = 12.9), and 36.6% identified as professional or semi-professional musicians, which we use as a proxy for musical expertise.
Groups: 21 participants had obtained a music description from an AI system such as ChatGPT or Gemini. We call them AIMD users. The other 61 are AIMD non-users.
Questions: Musical background, using selected Gold-MSI items. AIMD users also reported what they had used descriptions for and how satisfied they were. Everyone then rated their interest in using AI to describe music, chose the purposes they would use it for, and wrote the prompt they would give an AI system.
Analysis: Likert responses were skewed, so group comparisons used Mann–Whitney U tests (α = .05). Spearman’s correlation was used for satisfaction and interest among AIMD users. Multi-response items are summarized descriptively. All four authors coded the free-text prompts independently through inductive thematic analysis and discussed each response until they reached consensus.

The survey, from a participant’s seat

1Prior useeveryone

Have you used AI to describe music? 21 participants said yes and 61 said no.

2Interestafter an example

After seeing an example description, everyone rated their interest and picked the purposes they would use it for.

3Promptin their own words

Then they wrote the prompt they would give an AI system. These are the prompts behind Table 1.

Findings

1Few participants had used AI to describe music

Only 21 of the 82 participants (25.6%) had ever obtained a music description from an AI system, although 46 of the remaining 61 knew that current systems can produce one. Their most common uses were finding similar music, musicological analysis, and writing prompts for AI music-creation tools, each chosen by 8 of the 21. Satisfaction was polarized, with 47.6% rating their experience 5 or higher on the 7-point scale and 38.1% rating it 3 or lower.

2Interest was higher among prior users

Overall interest was low (M = 2.39, SD = 1.99 on a 7-point scale), and 58.5% of participants chose the lowest rating. Interest did not differ significantly between professional or semi-professional musicians and other participants (p = .88). AIMD users reported higher interest than non-users (U = 982.0, p < .001), and among the 21 users, interest correlated with their satisfaction with prior use (Spearman’s ρ = .67, p < .001).

Figure 1. Interest in using AI to describe music (a) by prior use, with means and 95% confidence intervals, and (b) against satisfaction with prior use, among AIMD users only (n = 21). Each dot is one participant. The line in (b) is a least-squares fit with its 95% confidence band.

3Users chose most purposes at higher rates

AIMD users selected nearly every listed purpose at a higher rate than non-users. Excluding the open “Other” field, finding similar music by describing a style was selected most often overall (27 of 82, 32.9%), and the largest gap was for generating prompts for AI music-creation tools (38.1% vs. 8.2%). Only two options had higher selection rates among non-users, accessibility for deaf or hard-of-hearing listeners and the open “Other” field. Several write-ins in that field expressed unwillingness to use AIMD, such as “I wouldn’t,” rather than additional purposes. This suggests that willingness to use these systems should be distinguished from intended use.

Figure 2. Envisioned purposes by prior use of AI music description. Participants could choose several purposes. Values are percentages within each group.

4Users specified content, non-users declined the premise

Every participant wrote the prompt they would give an AI system to get a music description, and we coded the prompts into the categories in Table 1. A larger proportion of AIMD users requested concrete content. Non-users more often rejected the idea or disengaged from it, as in “I would never use AI for anything.” While users viewed AI-generated descriptions as a tool to be specified, non-users often declined the premise.

AIMD userasked for content

AIMD users named the content they wanted, such as genre or instrumentation.

AIMD non-userdeclined the premise

AIMD non-users more often refused or disengaged.

Selected prompt categories by prior use, from Section 3.3 of the paper. Values are percentages within each group.

Question asked right after participants chose their purposes (Figure 2)

What would you ask an AI if you wanted it to generate a music description for your chosen purposes? Please write down as much as you would like to.

You may freely include any information, such as a specific name of a piece or the context for its use.

Select a category to read the answers coded with it.

Table 1. Categories identified in participants’ prompts. Counts are participants whose prompt received each code, and one prompt can carry several codes.

Conclusion

Interest was low in this sample. It did not differ significantly by musical expertise but was higher among participants with prior AIMD experience. This divide also appeared in envisioned uses and prompts: AIMD users specified musical content, while non-users often declined the premise. People want music descriptions for different purposes, so future work should evaluate model outputs against those purposes and examine how people respond to them. Our sample was small and self-selected, and the cross-sectional design cannot show whether prior use increases interest.

Citation

@inproceedings{oh2026usercentered,
  title     = {A User-Centered View of {AI} Music Description},
  author    = {Oh, Eun Ji and Park, Sangheon and
               Kim, Yonghyun and Beck, Jun Woo},
  booktitle = {Extended Abstracts for the Late-Breaking Demo Session
               of the 27th International Society for Music
               Information Retrieval Conference},
  address   = {Abu Dhabi, UAE},
  year      = {2026}
}