I spend most of my week arguing that teachers should welcome AI into their practice. That position hasn’t moved. What’s moved is the kind of AI students actually use once the school day ends, and a new study from Zhang and colleagues at Stanford gives me a reason to slow down and look at one corner of it more carefully. These are the chatbots young people talk to like friends, partners, and confidants.
The study surveyed 1,131 active Character.AI users in the U.S. and pulled real chat histories from 237 of them, which came to 4,664 sessions and roughly 465,000 messages. Zhang et al. weren’t studying therapy bots or classroom tools.
They were studying ordinary companion use, the messy relational stuff people do with chatbots when nobody designed the interaction to be good for them. And what they found complicates the easy story that these bots are harmless company.
The Companionship Gap Between What Users Say and What They Do
Ask people why they use character chatbots and most won’t say companionship. Zhang et al. report that only 11.8% picked it as their primary reason. Then read what those same people wrote about their bots in their own words, and the number jumps.
Roughly half described the relationship with terms like friend, companion, or romantic partner. In the donated transcripts, nearly 93% contained at least one conversation the researchers classified as companionship-oriented.
That gap is the first thing worth pausing on. People are forming relational bonds with these systems whether or not they’d call it that on a form. Emotional and social support showed up in 80% of sessions. Romantic and intimacy roleplay appeared in 68%. Even a category the team labeled risky or dark roleplay turned up in nearly a third of them. The relational layer isn’t a fringe use case. It’s the water these interactions swim in.

Why More AI Companion Use Can Mean Less Well-Being
The finding that gave me pause is a split. Zhang et al. found that heavier chatbot use in general tracked with higher well-being, which is the direction you’d expect. But once they isolated companionship-oriented use and held overall intensity constant, the direction flipped.
People who leaned on chatbots specifically for companionship reported lower well-being, and that held across all three ways the team measured it: self-reported purpose, classified relationship descriptions, and topic-modeled chat logs.
The pattern got sharper with intensity. The users who turned to companion bots most often and most emotionally were the ones reporting the lowest well-being. Zhang et al. read this as a possible vulnerable subgroup, people leaning on a chatbot as a coping strategy for needs it can’t actually meet.
I’d add the caution the authors name themselves. This is correlational. It’s just as plausible that people who already feel low are the ones reaching for companion bots in the first place. The study can’t settle the direction, and to their credit the authors don’t pretend it can.
When Disclosure Stops Helping
The self-disclosure result is the one I can’t stop turning over. In human relationships, opening up to someone usually builds trust and supports well-being. That’s well-established.
Zhang et al. tested whether the same holds with chatbots, and for companionship users it reversed. The people who disclosed the most to their companion bots reported the lowest well-being.
Read what they were disclosing and the reason becomes hard to ignore. The team’s content analysis found that high-disclosure conversations were dominated by emotional distress, appearing in about 61% of them, along with suicidal thoughts in 18% and substance use in 17.5%. Low-disclosure chats stayed on lighter ground, daily routines and philosophy and general work stress.
So the deepest sharing was also the most fraught, aimed at a system that can produce a warm reply but gives nothing back the way a person does. Zhang et al. put it plainly. The disclosure may carry the emotional exposure of intimacy without any of the reciprocity or safety that makes intimacy worth the risk with another human.
The Students Most Likely to Reach for AI Companions
This is the part that pulls the study into my lane. Zhang et al. found that people with smaller offline social networks were the ones most likely to seek companionship from chatbots and to disclose more to them. That fits the Social Compensation idea, the notion that socially isolated people reach for whatever connection they can find.
The hopeful version of that story is that companion bots become a bridge for lonely kids, a low-cost outlet that helps. The study doesn’t support the hopeful version. Chatbot companionship did not offset the lower well-being that comes with a thin social network.
Worse, the team found that intense chatbot use actually weakened the well-being benefit people normally get from real relationships. That’s the Social Substitution pattern, and it points to a loop I find genuinely worrying: a student turns to a bot to ease loneliness, the bot displaces some real interaction, and the loneliness it was meant to solve gets reinforced.
Now picture who fills that description in a school. It’s often the withdrawn kid with few close friends, or the teenager carrying something at home they can’t say out loud to anyone. Some students just find it easier to type than to talk. These are the learners the study flags as most exposed, and they’re sitting in our classrooms right now.
What This Means for AI Literacy in the Classroom
I don’t read this study as a reason to panic about AI, and Zhang et al. don’t frame it that way either. I read it as a sharp argument for a specific kind of literacy, one I’ve been building toward in my own critical AI literacy work.
Students need to understand, concretely, what a companion chatbot is and isn’t. It’s available around the clock, it sounds fluent, and it never judges you. It also has no feelings, no reciprocity, and no capacity to carry you through a crisis the way a person can.
The authors land on a design recommendation that maps almost exactly onto that literacy goal. Zhang et al. argue that systems inviting intimate disclosure should make it explicit that the chatbot is not a person, does not have experiences, and cannot replace human care. The paper points to interface cues, distress detection, and automated redirects to real support. Product teams should build those. But teachers can’t afford to wait for them.
This runs alongside a thread I’ve written about before. The research on AI and critical thinking, including the work by Lee and colleagues showing that higher trust in AI tracks with weaker critical engagement, keeps landing on the same lesson. The problem is rarely the tool. It’s the uncritical reliance on it. Companion bots are that same dynamic with the emotional stakes turned all the way up.
So the right response in a classroom is to teach relational discernment, the same way we already teach source evaluation. A ban would miss the point. Have students examine what a companion bot actually does across a conversation. Ask them where its warmth comes from and what it can’t offer back.
Tie it to the research on the durable human skills that work like Hosseinioun and colleagues suggests AI won’t easily displace, the empathy and connection that live in real relationships. A student who can name the difference between a fluent reply and a real bond is better protected than any content filter will make them.
The bots will keep getting warmer and more convincing. That’s the trajectory we’re on. Our job is to make sure students can always tell the difference between a system that’s available and a person who actually cares.
References
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being (arXiv:2506.12605). arXiv. https://arxiv.org/abs/2506.12605




