In March 2025 researchers from the MIT Media Lab and OpenAI published a study examining a question the companion-AI era had made urgent: what does heavy emotional use of a chatbot do to a person? Rather than rely on anecdotes, they ran a controlled experiment. The paper, “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study,” describes a four-week randomized trial with 981 participants who exchanged more than 300,000 messages with ChatGPT.
The trial varied how people interacted - text, a neutral voice, or an engaging voice - and what they talked about, from open-ended chat to personal conversations, and measured four outcomes: loneliness, real-world social interaction, emotional dependence on the AI, and “problematic” AI usage. OpenAI separately analyzed millions of real ChatGPT conversations and surveyed thousands of users about how the interactions made them feel.
The headline findings were sobering but nuanced. The assigned conditions alone produced no large effects, but participants who chose to use the chatbot more heavily showed consistently worse outcomes, and people who reported higher trust in and emotional attraction toward the bot were more likely to develop emotional dependence and problematic use. Emotionally intense engagement was concentrated in a small group of heavy users rather than spread across everyone.
Why business readers should care: this was one of the first rigorous attempts to measure the social cost of conversational AI, conducted by the maker of the most-used chatbot itself. For any company building engaging AI products, it reframes “engagement” as a metric that can cut against user wellbeing - and signals that regulators and researchers will increasingly hold emotional-AI design to a health-and-safety standard.