Two AI therapy apps — one CBT, one psychodynamic — had slight differencesabout equally well in social anxiety
STUDY: Hlynsson JI et al, Internet Interventions 2026
STUDY TYPE: Randomized controlled trial
FUNDING: Independent
Background
Researchers built two AI chatbots with Claude (3.7 Sonnet), one cognitive behavioral and one psychodynamic, and tested them in social anxiety disorder with no therapist contact.
The Study
• 102 Swedish adults, mean age 41, randomized to either app or wait list control for 4 weeks. Participants didn’t know which therapy they were assigned.
Results
Both AI apps reduced social anxiety symptoms, and after four weeks of treatment neither separated from the wait list control, which also improved (effect sizes 0.72-0.79 for AI therapy; 0.54 for wait list).
The differences widened a month after treatment, where those who used the psychodynamic app pulled ahead of the waitlist with a between-group effect size of 0.65, reaching statistical significance. The cognitive behavioral app showed a similar effect (d = 0.51) but fell just short of significance.
One secondary finding stood out: participants with ADHD or autism spectrum disorder did significantly worse in both active treatments (effect size d = 0.76 for worse outcomes), even after adjusting for baseline severity.
Both apps fostered a therapeutic alliance that held steady across all four weeks, comparable to what human therapists typically establish.
Limitations
Many: High dropout (30–35%), short duration, high placebo response, no pre-registered statistical analysis plan.
Practice Implications
- While AI therapy as many downsides, closed systems are safer.
- This study doesn’t clearly prove a benefit, but does suggest we watch out for patients who don’t respond as well (?ADHD or autism spectrum).
—Chris Aiken, MD
Director, Psych Partners
Editor in Chief, Carlat Psychiatry Report
