MIT and OpenAI found that heavier chatbot use tracks with more loneliness, not less

Published in March 2025, this was two studies run in parallel rather than one. MIT Media Lab researcher Cathy Fang and colleagues ran a four week randomized controlled trial with close to 1,000 participants using ChatGPT daily. Separately, OpenAI analyzed roughly 40 million real ChatGPT conversations for patterns, without any human reviewer reading the content, to see how often people engaged emotionally with the chatbot in normal use.

• ~1,000 participants in a 4-week randomized controlled trial (MIT)
• ~40 million real ChatGPT conversations analyzed (OpenAI)
• Neither study used human reviewers reading chat content

The topline finding across both studies: higher daily usage correlated with more loneliness, more emotional dependence, more problematic use, and less time spent socializing with actual people. That held across every conversation type and interaction mode the researchers tracked. Voice mode looked protective at first, reducing loneliness better than text at low usage levels, but that advantage disappeared once people used it heavily, particularly with a neutral rather than expressive voice.

• Higher usage → more loneliness, emotional dependence, problematic use
• Less time socializing with actual people
• Voice mode's protective effect disappeared with heavy use

The nuance that got lost in a lot of the coverage: emotional engagement with ChatGPT turned out to be rare overall. Most conversations in OpenAI's dataset had no emotional content at all. The people who did bond with the chatbot, trusted it, and talked to it about personal topics were the ones who showed the loneliness and dependence effects.

• Most conversations had no emotional content at all
• Bonding with the chatbot correlated with loneliness and dependence
• Researchers describe it as correlation, not causation

Neither study has completed peer review as of this writing.