JMIR Publications

ABSTRACT

Background:

Capitalization refers to sharing positive personal events with others. When met with active and constructive responses, this can enhance positive affect, but people may not always have access to such a listener. Large language models can produce supportive responses, yet the immediate mood effects of artificial intelligence (AI)-based capitalization dialogue and the role of presentation modality remain unclear.

Objective:

This study aimed to examine whether a single 10-minute session of AI-based capitalization dialogue improved immediate mood relative to quiet rest, whether outcomes differed across presentation modalities, and whether baseline depressive symptoms and loneliness moderated mood change.

Methods:

We conducted a four-arm pilot randomized experiment with 80 university students recruited offline in Japan. Sessions were conducted in person. Participants were randomly assigned to Spatial Display (SD), Flat Display (FD), Text-Only Display (TD), or quiet-rest control conditions (n=20 per group). SD and FD presented a 3D character with voice interaction on a depth-rendering display and a standard laptop, respectively. TD used laptop-based text exchange. The three conversational agent (CA) conditions used the same GPT-4o mini dialogue system as a capitalization partner. The dialogue was fully automated, with researcher involvement limited to setup and technical assistance. Mood states were assessed before and after the session using the POMS2. CA participants also completed measures of perceived responses to capitalization attempts, perceptions of the agent, and perceived intimacy. Linear mixed-effects models tested time × group interactions. Multivariate analyses of variance compared CA conditions on subjective evaluations. Baseline depressive symptoms (PHQ-9) and loneliness (UCLA-10) were tested as moderators.

Results:

All 80 randomized participants completed the assigned 10-minute session. The time × group interaction for Total Mood Disturbance (TMD) was significant (F(3,76)=3.16, P=.03). Relative to control, TMD decreased significantly more in the FD (b=−12.90, P=.005) and TD groups (b=−10.60, P=.02). The SD group did not differ significantly from control (b=−8.55, P=.06). Mood change did not differ significantly across the three CA conditions (P=.70). Both FD and SD elicited greater social presence than TD. FD also elicited greater perceived enjoyment than TD, and SD elicited greater perceived intimacy than TD. Perceived capitalization responses did not differ across modalities. Higher depressive symptoms predicted greater mood improvement (b=−5.00, P=.02), whereas higher loneliness predicted smaller improvement (b=5.29, P=.01). Overall, 78% (45/58) would use the application, and 69% (40/58) would recommend it.

Conclusions:

A single 10-minute session of AI-based capitalization dialogue was associated with greater immediate mood improvement than quiet rest in two of three modalities. Visual embodiment improved some user experience outcomes without altering overall mood effects. These findings support further testing with active control conditions, repeated use, and larger samples.