Conversational AI Agent Eases Psychiatric Symptoms In Randomized Trial

Key Takeaways

• Among college students, anxiety was associated with reduced anxiety with a conversational AI platform compared to face-to-face therapy.
• Perceived digital therapeutic alliance in the AI arm was associated with engagement and indirectly with symptom improvement.

The JAMA Network Open randomized clinical trial compared a 12-week conversational AI platform, face-to-face group therapy, and a waiting list control group. Participants were reassessed 3 months after treatment ended, and the study also examined digital therapeutic alliance alongside engagement and follow-up symptom trajectories.

Researchers conducted a 3-arm parallel-group randomized clinical trial from April 1 to October 27, 2025. The study randomized 995 university students aged 18 to 35 years with psychological distress in a 1:1:1 ratio: 336 to AI, 331 to face-to-face therapy, and 328 to control. Recruitment, screening, consent, and baseline assessments were remote. Prespecified symptom outcomes were anxiety by GAD-7, depression by PHQ-9, and PTSD symptoms by the brief PTSD Checklist for DSM-5. Researchers also measured well-being by WHO-5, life satisfaction by the Brief Multidimensional Students’ Life Satisfaction Scale, and intention to seek therapy.

At 12 weeks, anxiety scores were lower in the AI arm than in face-to-face therapy, with an adjusted mean difference of -2.17 (95% CI, -2.67 to -1.67). Anxiety scores were lower than control, with an adjusted mean difference of -2.15 (95% CI, -2.65 to -1.65). Depression also fell more in the AI arm than in control, with an adjusted mean difference of -1.99 (95% CI, -2.63 to -1.35). Well-being also favored AI, with differences of 5.72 (95% CI, 2.71 to 8.73) versus face-to-face therapy and 9.16 (95% CI, 6.14 to 12.18) versus control. Life satisfaction differences were 1.19 (95% CI, 0.23 to 2.14) versus face-to-face therapy and 2.58 (95% CI, 1.62 to 3.54) versus control; PTSD differences were not significant.

At 3 months, anxiety scores remained lower in the AI arm than in face-to-face therapy, with a mean difference of -1.65 (95% CI, -2.42 to -0.87). Anxiety also remained lower than control, and depression remained lower than control, at mean differences of -2.08 (95% CI, -2.85 to -1.30) and -1.79 (95% CI, -2.82 to -0.76). Well-being remained higher with AI, with differences of 6.08 (95% CI, 3.07 to 9.10) versus face-to-face therapy and 10.13 (95% CI, 7.11 to 13.15) versus control. Life satisfaction also remained higher, at 2.17 (95% CI, 1.05 to 3.29) versus face-to-face therapy and 2.79 (95% CI, 1.67 to 3.91) versus control. PTSD differences remained nonsignificant at the 3-month posttreatment assessment.

In the AI arm, therapeutic alliance was assessed with 15 adapted Counselor Rating Scale items covering warmth, empathy, and professionalism. Participants averaged 18.6 messages, 3.0 active days, 1645 words, and 83.1 minutes per week, and 61.0% remained active through week 12. Alliance was associated with engagement, at β = 0.31 (95% CI, 0.16 to 0.43; P < .001). Engagement was associated with symptom improvement (β = -0.58; 95% CI, -0.69 to -0.46), and the indirect path from alliance through engagement to symptom improvement was significant (β = -0.18; 95% CI, -0.27 to -0.10), both P < .001. No serious adverse events were reported; minor transient distress increases occurred in 4 of 336 AI participants and 5 of 331 face-to-face participants, with no discontinuations due to adverse events.