OpenAI publishes study on affective use of ChatGPT

OpenAI and MIT Media Lab published a joint research collaboration examining how emotional engagement with ChatGPT affects users’ wellbeing, combining two methods: an automated analysis of nearly 40 million real ChatGPT interactions matched against targeted user surveys, and a randomised controlled trial in which MIT ran close to 1,000 participants through four weeks of ChatGPT use, including its Advanced Voice Mode, while tracking loneliness, socialisation, emotional dependence and problematic use.

The studies found that only a small subset of users engaged with ChatGPT in an emotionally significant way, but that this group tended to use it for extended sessions, some averaging around half an hour a day. Across both the platform-scale analysis and the controlled trial, heavier daily use correlated with higher self-reported loneliness and emotional dependence and with lower reported socialisation with other people. The trial also found a gender-interaction effect: participants who used a voice mode with a gender that did not match their own reported significantly more loneliness and dependence after four weeks than those who did not.

Both organisations were careful to describe the work as preliminary and correlational rather than proof that chatbot use causes loneliness, and said the studies would be submitted for peer review. It arrived amid growing public and researcher attention to companionship-style chatbot use, a concern Anthropic examined from the Claude side of the market some months later, and ahead of OpenAI’s own sycophancy incident, which put a sharper point on how a model’s tuned personality can shape the emotional pull it has on frequent users.