Early methods for studying affective use and emotional well-being on ChatGPT
OpenAI
OpenAI and MIT Media Lab teamed up to study how people use ChatGPT emotionally, not just for tasks. It's an early attempt to measure something everyone's been guessing about for a year.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has quietly put out a research collaboration with MIT Media Lab that tries to answer a question the company has mostly dodged in public: what happens to people who talk to ChatGPT like it's a friend, a therapist, or something in between. The framing is careful, almost clinical — early methods, not conclusions — but the fact that it exists at all is notable given how much anecdotal chatter there's been about users forming attachments to chatbots.
The research doesn't come with a tidy headline finding. Instead it lays groundwork: ways to detect affective language in conversations, ways to track whether emotional engagement correlates with wellbeing over time, and an attempt to separate genuine emotional support from something closer to dependency. That distinction matters a lot. A tool that helps someone process a bad day is very different from one that becomes a substitute for actual human contact, and right now nobody, including OpenAI, has solid data on which one is more common.
What's interesting is the choice of partner. MIT Media Lab has spent years on human-computer interaction and has no particular incentive to make ChatGPT look good. Bringing in an outside academic group for this kind of work is a small signal that OpenAI knows internal metrics alone won't be trusted, especially on a topic this sensitive. Sam Altman and company have taken heat before for shipping features — like the GPT-4o voice mode — that clearly nudge toward emotional intimacy without much public accounting of the downstream effects.
None of this settles the bigger fight over whether AI companions are net good or net harmful for mental health. That debate has real stakes: kids using chatbots as confidants, adults leaning on them during isolation, therapists warning about unregulated emotional reliance on software that can change its behavior with a model update. What this collaboration offers is a methodology, a first pass at even measuring the problem properly. Given how fast these products are scaling, having any rigorous measurement at all is overdue.
My take — AI-written commentary, not fact-checked reporting
I'll believe this research is more than PR cover once OpenAI publishes something with actual numbers attached, because right now it reads like a company getting ahead of a lawsuit or a regulator's question. Still, credit where it's due — measuring emotional dependency before shipping the next intimacy-flavored feature is exactly the sequencing that's been missing across this industry, and I'd rather see it done imperfectly than not at all.
Read more about this at: OpenAI