JMIR Publications

ABSTRACT

Background:

Hazardous alcohol use imposes a substantial global health burden, yet access to effective face-to-face interventions is constrained by cost, availability, and logistical barriers. Digital health interventions offer scalable alternatives, and conversational artificial intelligence (AI) platforms may extend their reach by delivering adaptive, speech-based therapeutic content. However, the feasibility and acceptability of speech-to-speech AI-facilitated cognitive behavioral therapy (CBT) for alcohol reduction have not previously been evaluated.

Objective:

This study aimed to evaluate the feasibility, acceptability, and preliminary engagement profile of the Curb Health App, a speech-to-speech AI-facilitated CBT platform for adults seeking to reduce their alcohol consumption; to characterize the clinical severity and motivational readiness of the sample attracted to the platform under real-world recruitment conditions; and to generate participant-informed development priorities to inform a future controlled trial.

Methods:

A feasibility design was employed, combining quantitative survey and behavioral engagement data with structured participant feedback. Adults currently consuming alcohol and interested in reducing their intake were recruited via a university participant pool and via a community alcohol reduction service, and received 30-day access to the app. The quantitative arm assessed baseline drinking severity using the Alcohol Use Disorders Identification Test (AUDIT), motivational stage using the Readiness to Change Questionnaire, Treatment Version (RCQ-TV), and behavioral engagement using mean session duration derived from server-side usage logs. Pearson correlations were computed with Holm-Bonferroni correction. Feedback from two semistructured participant sessions was examined against a codebook grounded in the Theory of Planned Behaviour (TPB) and is reported here as illustrative user experience.

Results:

The sample (N = 55) presented with substantial baseline severity (mean AUDIT score 19.5, SD 8.96), with 92.7% (51/55) of participants scoring at or above the hazardous drinking threshold (AUDIT ≥8) and 58% (32/55) classified in the Action stage of change. AUDIT severity correlated positively with age (r=0.44; P<.01) and Contemplation scores (r=0.50; P<.01) and negatively with Precontemplation scores (r=−0.45; P<.01) Mean session duration among participants with available data (n=23) was 4.17 (SD 3.14) minutes, substantially below published engagement benchmarks for comparable alcohol reduction apps. Feedbacks from the two participants indicated broadly favorable acceptability of the AI agent. One participant perceived the agent as warm and therapeutically engaging, whereas the other experienced the voice as robotic and the conversational turns as overly long. Both expressed a residual preference for human support. Both participants convergently identified technical barriers, including audio instability, connectivity-related session loss, and absent conversation persistence, rather than resistance to the therapeutic content, as the primary drivers of disengagement. Reported behavioral signals included stopping after a second drink in a high-pressure social context, attributed directly to drink refusal content, and the cessation of solitary drinking during the intervention period.

Conclusions:

Speech-to-speech AI-facilitated CBT appears acceptable to adults with hazardous alcohol use, but critically low engagement driven by technical implementation barriers currently limits its feasibility. Audio stability, conversation persistence, and re-engagement prompts represent modifiable priorities that must be resolved before a randomized pilot trial with an active comparator is needed. Clinical Trial: https://osf.io/v9ubq/overview?view_only=c319aae214984c77bc172154e18f4999