AI-written guided imagery added to music: anxiety did not budge, negative mood did

Guided relaxation usually pairs music with a spoken script, and the script is usually written for a Western listener. Li and colleagues tested whether a script written to fit the listener’s culture does more, and used a large language model to write it. After a demanding 40-minute mind-mapping task, 113 Chinese undergraduates were randomized to Chinese film music with AI-written narration using Eastern imagery, a standard music-therapy recording with conventional narration, or piano music with no narration at all.

Two of the three outcomes showed nothing. State anxiety did not differ between groups (p = .51), and neither did positive affect (p = .36). Negative affect did differ: the Eastern imagery group scored lower than both comparison groups after adjusting for baseline, with a partial eta-squared of 0.08, which is a small-to-moderate effect. Heart rate also differed, but in a direction that does not fit the story, because the lowest adjusted heart rate was in the standard music-therapy group, not the Eastern imagery group.

The authors are unusually straight about what this does and does not show. They write that the findings do not establish efficacy or the independent effect of the culturally adapted suggestions, because the three conditions differed in musical material and duration as well as in the narration. In other words, the one group that improved on one measure was also listening to different music for a different length of time. That is the confound, the authors named it themselves, and it is why a 113-person randomized trial with no dropout still cannot answer its own question.

Limitations. The arms were not matched on anything except the broad idea of “music plus words”, so nothing can be attributed to the AI-written script. The sample was mostly female. The stress-induction task was never validated, so it is not established that participants were anxious to begin with, which matters when the anxiety measure came out flat. One significant result across several outcomes, in a pilot explicitly labelled exploratory, is the kind of finding that often fails to repeat. We grade this mixed rather than moderate, which is where its rubric score of 3.8 would put it, because the headline outcomes were null and only a secondary one moved. An argument for unsupported is reasonable: the primary anxiety outcome was flat, and the authors disclaim efficacy themselves.