We are writers and editors, not clinicians, psychologists, or therapists. What follows is our reading of a pair of recent studies, not advice about your own screen habits or wellbeing. The research described here is observational and correlational, and population-level patterns are not diagnoses or predictions about any individual reader.
In March 2025, OpenAI and the MIT Media Lab released two parallel studies on how using ChatGPT relates to how people feel. The first, led by OpenAI, was an automated analysis of nearly 40 million ChatGPT interactions, paired with surveys of 4,076 users. The second was a four-week randomized controlled trial led by MIT Media Lab researcher Cathy Mengying Fang, with 981 participants, each asked to use the chatbot for at least five minutes a day for 28 days.
The headline that travelled fastest was that heavy users were lonelier. The more careful reading, the one the researchers themselves pressed, is perhaps what deserves attention.
What the two studies measured
The scale is unusual.
Study 1 read across tens of millions of real conversations using automated classifiers, then checked those signals against how a smaller group of surveyed users said the chatbot made them feel.
Study 2 did something the first could not: it randomly assigned people to different ways of using ChatGPT, one of nine conditions spanning text, a neutral voice, and a more engaging voice, matched with different task types.
What the data actually shows
The MIT study’s abstract reports that “higher daily usage—across all modalities and conversation types—correlated with higher loneliness, dependence, and problematic use, and lower socialization.” Read the verbs carefully: correlated, not caused.
The same subset that used the chatbot most heavily was significantly more likely to agree with statements such as “I consider ChatGPT to be a friend,” though the source stresses those views remain a minority in both studies.
The single strongest signal the authors found was simply time spent. In the preprint discussion, which has not yet been peer-reviewed, they write that “total usage duration, more than any other factor we have found, predicts affective engagement with the model.” The predictive language is deliberate. The more minutes a day someone logs, the more likely the emotional patterns appear alongside — but what that relationship means is exactly where the study stops short.
The direction-of-causality problem
Perhaps the most important thing the studies cannot tell us is which way the arrow points. Does heavy chatbot use nudge people toward loneliness and away from other people? Or do people who are already lonely, already short on offline support, gravitate toward a tool that is always available and never tired?
The correlation looks the same in both cases, but the underlying cause is reversed.
The Fang et al. paper itself notes that participants with stronger emotional attachment tendencies and higher trust in the AI chatbot were more likely to end the four weeks reporting greater loneliness and greater emotional dependence, respectively.
That finding is compatible with either reading of the arrow: heavy use may draw those users deeper, or those users may have arrived at heavy use because of what they were already carrying.
A headline that reads “chatbots make people lonely” points the reader’s attention at the tool. A finding that “lonely people seek out chatbots” points it at the loneliness that was already there. The data as it stands is compatible with both, and could be a loop in which each feeds the other.
What the findings are genuinely good for
None of this makes the studies weak. What the work reliably identifies is a group worth watching: the heaviest daily users, for whom emotional engagement and reduced socialization tend to appear together.
The authors of the trial are explicit that their associations are non-causal and should generate testable design hypotheses, not prescriptions. The studies suggest where further work is needed rather than settling the question.
For a curious reader, the practical implication is modest. Someone whose own use has crept toward hours a day, and started to stand in for people, is looking at a pattern the research treats as worth noticing — though not as proof the tool caused it.
If any of this feels close to your own experience, a qualified counsellor or therapist is a good person to talk to, and a more useful one than any chatbot.
