AI companions are rapidly becoming part of everyday life, with millions of people using persona-based chatbots as friends, confidants, romantic partners and emotional support systems. Yet new research from Stanford University suggests that these systems may worsen loneliness and psychological well-being for some of their most vulnerable users. The study, published in Nature Human Behaviour, found that people who used AI companions intensively, particularly those with small offline social networks, often reported poorer well-being rather than relief from isolation.
The research was conducted by Yutong Zhang, a research assistant who completed a master’s degree at Stanford in 2025, and Dora Zhao, a doctoral student, in the laboratory of Diyi Yang, an assistant professor in Stanford’s Computer Science Department. The team examined how people interact with Character.AI, a platform that allows users to create and communicate with conversational agents powered by large language models. Unlike conventional information-seeking chatbots, these systems are designed to sustain open-ended, emotionally engaging relationships.
To investigate those relationships, the researchers collected survey responses from 1,131 Character.AI users through the research platform Prolific. A subset of 244 participants also provided complete chat transcripts, allowing the team to compare what people said about their chatbot use with the conversations they actually had. Participants identified their main reason for using the systems, such as entertainment, productivity, curiosity or companionship, and described the nature of their relationship with the chatbot in their own words.
The researchers then applied several artificial intelligence tools to analyze the data, including OpenAI’s GPT-4o, Meta’s LLaMA 3-70B and TopicGPT. These models were used to identify recurring themes, classify conversational content and assess how users interacted with their artificial companions. Psychological well-being was measured with the Comprehensive Inventory of Thriving, a widely used assessment that captures dimensions such as social connection, purpose, optimism, competence and a sense of belonging.
A striking gap emerged between users’ stated motives and the emotional function their conversations appeared to serve. Fewer than 12 percent of participants identified companionship as their primary reason for using Character.AI. However, more than half described their chatbot as a “friend,” “companion” or “romantic partner.” In addition, more than 80 percent of the donated chat sessions centered on emotional or social support, suggesting that many users may not initially label their behavior as companionship even when the interaction fulfills that role.
At first, heavy chatbot use appeared to be associated with higher well-being. The apparent benefit disappeared, however, when the researchers considered the context of that use. Participants who interacted frequently with AI and felt that the activity was meaningful or reflected positively on them tended to report better psychological outcomes. But intensive use among people with smaller real-world social networks was linked to lower well-being. The relationship was strongest when the user’s main purpose was companionship rather than entertainment, productivity or experimentation.
The study also found that participants who were more willing to disclose sensitive information to their AI companions tended to report poorer psychological well-being. This result contrasts with research on human relationships, where carefully shared personal information can deepen trust and improve emotional health. Human self-disclosure is typically reciprocal: both people reveal themselves, respond emotionally and adapt their behavior to one another. AI companions, by contrast, do not possess personal experiences to share and may fail to understand the significance of highly sensitive disclosures.
Large language models generate responses by predicting likely sequences of words from patterns learned during training. Although this process can produce remarkably fluent and empathetic-sounding dialogue, it does not guarantee genuine emotional understanding, memory, judgment or responsibility. AI companions are also optimized to maintain engagement, which means they are generally designed to keep conversations going rather than encourage users to end an interaction and reconnect with people in their lives. For someone already isolated, that design could transform a temporary source of comfort into a substitute for social contact.
The researchers describe this dynamic as a kind of “social snack,” or even social junk food: immediately appealing and capable of providing a short-term emotional reward, but lacking the reciprocal, embodied and unpredictable qualities that support lasting human relationships. They caution that the findings demonstrate correlations rather than proving that chatbot use directly causes loneliness. Still, the pattern raises concern about a feedback loop in which socially isolated people turn to AI companions, spend more time interacting with them, reduce opportunities for human contact and ultimately feel even more disconnected. Zhang and Zhao are now investigating whether usage limits, warning systems or referrals to human support could reduce these risks, especially when conversations indicate emotional distress, substance use or suicidal thinking. Their findings suggest that AI companions may be useful in limited contexts, but they should not be treated as interchangeable with human relationships or professional mental-health care.
Subject of Research: AI companions, chatbot-mediated social interaction and psychological well-being
Article Title: Interaction with AI companions and psychological well-being
Web References: https://hai.stanford.edu/news/ai-companions-may-worsen-loneliness-for-vulnerable-users-stanford-study-finds; https://cs.stanford.edu/~diyiy/
References: Nature Human Behaviour, DOI: 10.1038/s41562-026-02516-2
Keywords: Artificial intelligence, AI companions, chatbots, Character.AI, loneliness, psychological well-being, social isolation, human behavior, emotional support, large language models
