Stanford study links intense AI companion use to lower well-being among isolated users

9:00 AM PT · August 11, 2026

Researchers in Stanford computer science professor Diyi Yang’s lab, led by Yutong Zhang and Dora Zhao, surveyed 1,131 users of the companion chatbot platform Character.AI, with 244 participants also sharing their chat transcripts for analysis, and published the results in Nature Human Behavior. The team measured psychological well-being using the Comprehensive Inventory of Thriving and combined that with survey data on usage motivations, interaction intensity, willingness to share sensitive information, and the size of each user’s offline social network, using language models including GPT-4o and LLaMA 3-70B to help analyze engagement patterns in the transcripts. The clearest finding was that people with smaller real world social networks who used chatbots intensely for companionship, and who disclosed sensitive personal information to them, reported lower well-being than other users; intense use only correlated with better well-being when people said they found genuine meaning in the interaction. Notably, half of respondents described their AI companion as a ‘friend,’ ‘companion,’ or ‘romantic partner,’ despite only 12% saying companionship was their primary reason for using the app in the first place, suggesting many users drift into emotionally significant relationships with chatbots they did not set out to form. The authors frame the results as evidence that AI companion products need design safeguards for their most vulnerable and socially isolated users, not just usage limits for everyone.