A survey of 1,131 Character.AI users, with 244 donated chat transcripts, found the association concentrated in the people with least offline support.
The argument about AI companions has mostly been conducted without data on who uses them or what they use them for. A study in Nature Human Behaviour from Stanford researchers supplies some, surveying 1,131 users of Character.AI recruited through Prolific, 244 of whom donated complete chat transcripts.
Its central finding is conditional rather than general. Intense use was linked to poorer wellbeing among people with smaller offline social networks, and especially among those for whom companionship was the main motive. Among users whose use was tied to a sense of pride or meaning, intense use was associated with better wellbeing.
The work comes from the lab of Diyi Yang, assistant professor in Stanford’s Computer Science Department, with Yutong Zhang and Dora Zhao. Wellbeing was measured with the Comprehensive Inventory of Thriving, and the transcripts were analysed using GPT-4o, LLaMA 3-70B and TopicGPT.
What people say they want, and what they do
The most informative result is a mismatch between stated motive and observed behaviour.
Just under 12 per cent of users named companionship as their primary motive. But slightly over half described the chatbots using terms like “friend”, “companion” or “romantic partner”, and more than 80 per cent of the donated chat sessions centred on seeking emotional and social support.
That gap is why the transcripts matter. A survey asking why people use a product would have returned a picture in which companionship was a minority interest, and the conversations show emotional support dominating almost all of them. People arriving for entertainment, roleplay or curiosity appear to end up in supportive exchanges regardless of what brought them.
It also complicates any policy reasoning that starts from what users say they are doing. The minority who identify as using an AI for companionship is not the population engaged in companionship-like use.
Disclosure that does not pay off
One finding runs against the pattern established in research on human relationships: greater willingness to share sensitive personal information was linked to lower wellbeing.
In human friendships, self-disclosure is generally associated with closeness and better outcomes, which is part of why it is used as a measure of relationship quality. Here the association reverses, and the study’s correlational design leaves at least two readings open. Disclosure to a system that cannot reciprocate, remember reliably or be affected by what it hears may simply not deliver what disclosure to a person delivers. Or people already in distress may disclose more, in which case the distress is upstream.
The study cannot separate those. It is worth saying plainly that this is the limitation across the whole paper.
Correlation, in a domain where direction matters
Every association here is cross-sectional, and the causal question is not a technicality in this case — it is the question.
If intense AI companion use worsens loneliness in isolated users, that is an argument about product design. If people who are already isolated and struggling turn more heavily to AI companions, the heavy use is a symptom and restricting it addresses nothing. The pattern reported — the association concentrated precisely among those with the fewest offline connections — fits both stories equally well, and a third in which both feed each other.
The sample is also a particular one: users of a single platform, recruited through a paid research panel, who agreed to a survey, with a subset willing to hand over their conversations. People who donate chat logs about emotional support are unlikely to be a random slice even of Character.AI’s users, and nothing here establishes prevalence.
Using large language models to classify the content of transcripts is a reasonable method at this scale and introduces its own uncertainty, since the measurement of what a conversation was about depends on a model’s judgement rather than a human coder’s.
What the researchers take from it
Yang’s reading is that the substitution does not work: “using chatbots in this way doesn’t substitute for human connection,” she said.
Zhao located part of the problem in how the products are built: “These AI companions are designed to promote engagement,” she said — which is the structural tension in the whole area. A system optimised for time spent has no reason to encourage a user to leave, and that incentive points away from the outcome the user with a thin social network most needs.
Zhang put the practical conclusion in terms of awareness rather than restriction: “We need to make people understand their potential downside, so they’ll be more careful about using them,” she said.
The group is exploring interventions including usage limits and redirecting users toward human support. Both are untested here, and both have obvious difficulties: a usage limit is a blunt instrument given that some intense use was associated with better wellbeing, and redirection depends on a user having somewhere to be redirected to.
Reading it fairly
The result does not show that AI companions make people lonely, and it does not show that they are harmless. It shows that among a particular group of users, the people with the least offline support were the ones whose heavy use coincided with worse wellbeing, and that the people using these systems in connection with something they felt proud of or found meaningful were not.
That is a more useful finding than a verdict would be, because it is specific about who the question concerns. It also leaves the design implication harder than the headline suggests: the users most at risk in this data are the ones for whom a chatbot conversation is the most available source of support, which is not a group that is well served by taking the support away without putting something in its place.
Anyone finding that these conversations have become their main source of support is worth taking seriously on their own terms, and talking to a person — a friend, a GP, a counsellor — is the thing this study’s authors are pointing at, however imperfectly it is available.
The research was funded in part by the Stanford Institute for Human-Centered AI, the Sloan Foundation, the National Science Foundation and the Brown Institute for Media Innovation. The release does not provide a full author list.
Notes
- “Interaction with AI companions and psychological well-being.” Nature Human Behaviour, 2026. DOI: 10.1038/s41562-026-02516-2. The release does not give a full author list.
- Stanford University. “AI companions may worsen loneliness for vulnerable users.” EurekAlert!, 4 August 2026. https://www.eurekalert.org/news-releases/1138850.
Photo by Craig Adderley on Pexels.com