Artificial intelligence as a virtual coach in a cognitive behavioural intervention for perfectionism in young people: A randomised feasibility trial

We examined the feasibility and outcomes of Artificial Intelligence (AI) as a virtual coach in guided self-help (GSH-AI) compared to pure self-help (PSH). Participants ( N = 85 undergraduate university students; M age = 20.65 years [ SD = 2.38]; 84 % female) were randomised to PSH ( N = 42) or GSH-AI ( N = 43). The intervention was a brief 11-module online cognitive behaviour therapy for perfectionism intervention completed over 4-weeks. GSH-AI participants were given suggested questions to ask AI for guidance in completing the intervention. Data were collected at baseline, 4- and 8-weeks post-randomisation. Engagement was good, only one person in each group did not use any modules; module completion was equivalent across conditions (6.67, SD = 3.22 and 6.18, SD = 3.42 respectively). Between baseline and post-intervention people in the GSH-AI condition showed an almost 3.5 times increase in preferring support to be received from AI versus other modes of support. Only 52 % and 22 % of participants completed 4- and 8-week post-randomisation surveys, with no differences in psychological outcomes between the PSH and GSH-AI groups. Main effects of time indicated moderate to large within-group effect size improvements for disordered eating, stress, anxiety, and perfectionism. Qualitative feedback indicated that AI was initially acceptable as a guide and became even more acceptable after it had been experienced. Fully powered trials are required to determine the impact of AI guidance on outcomes, and whether type of AI platform (customised versus generic) and type of mental health disorder interact with its effects. • We examined the feasibility of AI-guided internet CBT for perfectionism. • Barriers included lack of AI familiarity and the impersonal nature of responses. • Facilitators included ease of gaining valuable, confidential, tailored information. • Preference for AI support increased strongly after experiencing it. • The addition of generic AI support did not improve psychological outcomes.