A conversational artificial intelligence platform was associated with reductions in anxiety and improvements in well-being compared with face-to-face group therapy and a waiting list control following 12 weeks in a randomized clinical trial of 995 university students with psychological distress.
The study, published in JAMA Network Open, enrolled students aged 18 to 35 years in Israel who reported current psychological distress. Participants were randomly assigned to a 12-week conversational artificial intelligence (AI) platform, face-to-face group therapy, or a waiting list control group.
The AI intervention used Kai, a mobile and web-based platform delivering individualized psychological support through text-based exchanges. The program incorporated elements of cognitive behavioral therapy, acceptance and commitment therapy, dialectical behavior therapy, mindfulness, and positive psychology. Participants had unrestricted access and were encouraged to engage at least three times per week.
Participants assigned to group therapy attended 12 weekly 90-minute sessions led by licensed psychologists in groups of approximately 20 participants. The control group received no active intervention during the study period.
Eligibility required psychological distress on a brief screening measure, along with Hebrew fluency and internet access. Participants were excluded for active suicidal ideation or psychiatric crisis, current psychotherapy or psychiatric medication use, or a history of severe mental disorders.
Modest Symptom Differences, With Worsening in Comparator Arms
The primary outcomes were anxiety, depression, post-traumatic stress disorder (PTSD) symptoms, well-being, and life satisfaction at 12 weeks.
At the end of the intervention, anxiety scores were lower in the AI group than in both comparator groups. The between-group difference corresponded to approximately a 2-point reduction on the 21-point Generalized Anxiety Disorder-7 scale—below the commonly cited minimal clinically important difference of about 4 points.
Importantly, anxiety and depression scores increased over 12 weeks in both the group therapy and control groups, while scores in the AI group declined. This pattern contributed to the observed between-group differences.
Depression scores were also lower in the AI group compared with control, although differences between AI and group therapy were not statistically significant following adjustment for multiple comparisons. PTSD outcomes did not differ across groups.
Positive Functioning and Clinical Transitions
Measures of positive functioning favored the AI intervention. Participants using the platform reported higher well-being scores than those in both comparator groups and higher life satisfaction scores than those in the control and group therapy groups.
Among participants with clinically elevated baseline symptoms (Generalized Anxiety Disorder-7 score of 10 or greater), 58% in the AI group moved to the nonclinical range at 12 weeks, compared with 14% in the group therapy group and 10% in the control group. Similar, though smaller, differences were observed for depression.
At 3-month follow-up, improvements in anxiety, well-being, and life satisfaction remained greater in the AI group than in both comparator groups. Differences in depression persisted compared with control but not compared with group therapy.
Engagement and Therapeutic Alliance
Participants in the AI group sent an average of 18.6 messages per week across approximately 3 active days. However, only 61% of participants remained active through week 12, and engagement metrics reflect those who continued using the platform rather than all randomized participants.
Within the AI group, greater perceived therapeutic alliance was associated with higher engagement, which in turn was associated with symptom improvement. These findings were based on associative analyses and do not establish causality.
Interpreting the Group Therapy Comparator
The face-to-face comparator consisted of relatively large groups of approximately 20 participants, which may differ from smaller or more intensive therapy formats commonly used in clinical practice. This design may have influenced the relative performance of the group therapy arm.
Notably, group therapy did not show statistically significant improvements compared with the control group for anxiety or well-being outcomes.
No Effect on PTSD
No statistically significant differences were observed across groups for PTSD symptoms at either the end of treatment or follow-up. The researchers suggested this may reflect the absence of trauma-focused, exposure-based interventions, which are typically required for meaningful improvement in PTSD.
Limitations and Clinical Context
The researchers noted several limitations, including reliance on self-reported outcomes, lack of masking, and substantial attrition at follow-up. Baseline symptom severity was generally mild to moderate, and participants with more severe psychiatric conditions were excluded, which may limit generalizability to clinical populations.
“These findings suggest that conversational AI may serve as a scalable resource within mental health frameworks, although its role is likely best suited as an adjunct or early intervention tool,” wrote lead researcher Anat Shoshani, PhD, of Baruch Ivcher School of Psychology, Reichman University in Israel, and colleagues.
The findings suggest that conversational AI may improve anxiety and well-being outcomes compared with group therapy and control in this population. However, the magnitude of change was modest, comparator groups showed increases in symptom scores over time, and the study population was limited to university students with mild to moderate distress.
Disclosures can be found in the published study.
Source: JAMA Network Open
