MIT and OpenAI study reveals who's most at risk from AI friends

A new joint study from MIT and OpenAI confirmed what many of us on the beat suspected: for the vast majority of people, ChatGPT is a tool, not a therapist. But for a small, heavy-usage group, the emotional lines are blurring, and the consequences are complicated. The research pair, which analyzed nearly 40 million interactions and ran a controlled trial with 1,000 participants, found that emotional engagement is remarkably rare on the platform. Affective cues — language indicating empathy or support — were essentially absent in most conversations. That’s the good news if you were worried about a mass replacement of human bonds.

But drill down into the data on heavy Advanced Voice Mode users, and a different picture emerges. A tiny subset is doing a lot of emotional heavy lifting. These users are far more likely to agree with statements like “I consider ChatGPT to be a friend.” This isn’t a broad trend you’d spot in platform averages; it’s a concentrated phenomenon that makes studying the impact a real headache. The controlled experiment revealed that how you talk to the model matters immensely.

Voice mode, particularly, showed a split personality when it came to well-being. Brief voice chats correlated with users feeling better, but the effect soured with prolonged daily use. Interestingly, the study found that a more engaging, personable voice didn’t lead to more negative outcomes than a neutral one over time. The content of the conversation proved more potent. Personal chats were linked to higher loneliness but also lower emotional dependence, while non-personal chats had the opposite association. It’s a messy, human result that defies a simple moral panic.

This work is clearly a starting point, not a final word, but it injects some much-needed data into a conversation dominated by vibes. It suggests we should worry less about the model seducing the masses and more about a vulnerable slice of users forming intense bonds. Platform design can’t just be about safety rails; it might need to actively guide people away from patterns that look, according to this data, like a path to isolation.