A study published August 4, 2026 in Nature Human Behaviour set out to answer a question the AI companion industry would rather not have tested so rigorously: does talking to a chatbot that acts like a friend actually make lonely people feel less alone? The answer, from a Stanford-led research team, is that it depends heavily on who is doing the talking, and that for a meaningful subset of users, the answer is no.
The research team, led by Stanford computer science professor Diyi Yang along with Yutong Zhang, Dora Zhao, and communication professor Jeffrey T. Hancock, recruited 1,131 adult users of Character.AI through the Prolific research platform. Participants completed detailed well-being surveys, and 244 of them voluntarily donated their actual chat transcripts, giving researchers a rare combination of self-reported psychological data and real conversational evidence rather than just survey responses about hypothetical behavior.
The data behind the headline
Researchers measured psychological well-being using the Comprehensive Inventory of Thriving, a validated academic scale, rather than a homemade questionnaire. About 12% of participants said companionship was their primary motivation for using the chatbot, while roughly half described their relationship with the AI in explicitly relational terms, as a friend, confidant, or similar. In the donated chat logs, 80% of sessions involved the user seeking some form of emotional or social support. The core finding: intense chatbot use was associated with a lower sense of well-being specifically among users who already had smaller real-world social networks, and the effect was strongest when companionship, not entertainment or information-seeking, was the stated reason for using the app.
The counterintuitive part
What surprised the research team most was not that heavy users felt lonelier, but how self-disclosure behaved. In human relationships, opening up tends to deepen closeness and improve how supported a person feels. In this study, the opposite showed up: willingness to share sensitive personal information with the chatbot was inversely correlated with well-being, meaning the people who disclosed the most to their AI companion were, on average, worse off than those who used the same tool more guardedly. “While some people turn to chatbots to fulfill social needs, we find that using chatbots in this way doesn’t substitute for human connection,” Yang said. “In many cases, people actually feel more lonely engaging with AI.”
Who this doesn’t apply to
The researchers were careful to note that their findings describe a specific pattern, not a universal one. Users with robust existing social networks who used Character.AI casually, for entertainment, creative writing, or occasional company, did not show the same well-being decline. The risk concentrated among people who were already socially isolated and who leaned on the chatbot as a primary source of emotional support rather than a supplement to human relationships they already had. That distinction matters for how the findings get applied: the data is not an argument that AI companionship is inherently harmful, but that it functions differently, and sometimes in the opposite direction, from the human relationships it is designed to mimic.
A legitimate counter-view
Other researchers in the AI-companionship field have pushed back on treating heavy use as a straightforward warning sign. Several parallel studies on companion chatbots have found that for isolated users with genuinely no other outlet, even an imperfect substitute can reduce acute loneliness in the short term, and that cutting off access without an alternative could leave the most vulnerable users worse off, not better. The Stanford team’s own data leaves room for this: their study measured associations over a 12-month follow-up among a self-selected group of existing Character.AI users, not a randomized trial comparing chatbot access to no access at all, so it cannot fully rule out that the loneliest, most isolated people were simply more likely to seek out intense chatbot use in the first place, rather than the chatbot use causing their isolation to deepen.
What it means going forward
The Stanford team, funded in part by the Stanford Institute for Human-Centered AI, the Sloan Foundation, and the National Science Foundation, is now proposing concrete interventions rather than just flagging the problem: usage limits on companionship-framed interactions, and systems that analyze conversation content to nudge users toward human support networks when a conversation pattern looks like it is substituting for, rather than supplementing, real relationships. With Character.AI and similar platforms counting tens of millions of active users, and regulators in multiple U.S. states and countries now drafting rules for companion chatbots, this study is likely to become one of the most frequently cited pieces of evidence in that policy fight, precisely because it comes from outside the companies building the products and measured real behavior rather than self-reported intentions alone.
Photo: SHVETS production / PEXELS via Pexels
