A new randomized controlled trial involving nearly 1,000 participants reveals complex, often counterintuitive insights into how generative AI interactions via text and voice impact human mental health and behavior.

Key Points

• Researchers utilized a 3x3 factorial design to test ChatGPT interactions across three modalities—text, neutral voice, and engaging voice—and three conversation types.
• Contrary to common assumptions, lonelier individuals did not voluntarily spend more time interacting with the AI chatbots during the study.
• The study found that text-based interactions elicited higher levels of emotional outpouring from users compared to voice-based communication.
• Increased time spent interacting with AI, regardless of the modality or conversation type, was associated with worsening psychosocial outcomes for participants.
• The experiment involved 1,000 U.S.-based adults recruited via CloudResearch, with each participant receiving $100 for completing the study.

Why it Matters

These findings challenge prevailing assumptions about how humans form emotional attachments to AI and suggest that increased usage may carry hidden risks for mental well-being. As millions of people turn to chatbots for guidance, this research highlights the urgent need for rigorous clinical oversight to prevent potential long-term societal harm.