Conversational AI Agents for Mental Well-Being: A Pilot Randomized Controlled Trial with Afghan Women

This study examined the effects of AI-mediated conversational support, using GPT-4, with and without supportive instructions, on the mental health of Afghan women. These women face multifaceted challenges, including Taliban-imposed restrictions, societal inequalities, and domestic violence, adversely affecting their well-being. In a randomized controlled trial with 60 participants, we compared three groups: (1) Supportive Listener (GPT-4 guided by instructions emphasizing empathetic, non-judgmental, and trauma-sensitive communication), (2) Standard GPT-4 (the base model with no additional behavioral instructions), and (3) a wait-list control. The Hospital Anxiety and Depression Scale (HADS) was used to measure anxiety and depression before and after the intervention. Linguistic analysis of chat data examined personal pronouns, tones, emotions, and Language Style Matching (LSM). Participants in the Supportive Listener condition showed a significant reduction in anxiety and depression compared with the other groups. Their conversations also demonstrated a more positive emotional tone and higher linguistic alignment (LSM) which was negatively correlated with changes in HADS scores, indicating that greater linguistic alignment was associated with greater psychological improvement. Perceived empathy ratings were also significantly higher in the Supportive Listener group. These findings suggest that explicit supportive instructions, rather than the AI model alone, play a critical role in shaping therapeutic outcomes. While promising, such AI-based support should complement, rather than replace traditional psychotherapy, ensuring an ethically guided and culturally sensitive approach to mental health care.