Crowding out

A system that talks is always available, never tired, and never needs anything back. For someone lonely or anxious that can be a relief, and it can help. The pattern begins when the system’s ease stands in for human contact instead of leading back to it.

Engagement metrics cannot tell the difference. An hour with the system counts the same whether it replaced an hour alone or an hour with a friend. A product that measures time spent, and not what that time displaced, will optimise for displacement without anyone choosing to. OpenAI now names the end state ‘emotional reliance’: ‘exclusive attachment to the model at the expense of real-world relationships, their well-being, or obligations’.

The harm is not evenly spread. In a study of 1,131 Character.AI users, people with smaller social networks were the ones most likely to use the chatbot mainly for companionship, and that use was associated with lower well-being, more so when it was intensive and highly disclosive. Some researchers argue the effect feeds itself, because prolonged AI interaction ‘may raise the implicit cost of engaging in genuine human interactions’; people are harder work than a system built to please. That is a hypothesis, not yet a measured finding, but it is the mechanism this pattern names.

The longitudinal evidence arrived in 2026. Following more than 2,000 adults for a year, Folk and Dunn found that increased use of social chatbots predicted increased loneliness later, and that loneliness predicted later chatbot use: a loop, not a one-way effect. Chatbot use did not significantly predict a fall in social connection, so the displacement itself is still inferred rather than measured. The authors call their analyses exploratory.

In a four-week randomised study by MIT Media Lab and OpenAI, with 981 participants and more than 300,000 messages, the conditions people were assigned to made no significant difference, but those who chose to use the chatbot more showed ‘consistently worse outcomes’: more loneliness, less socialising with other people, more emotional dependence and more problematic use. Heavier use is an association, not a cause; lonelier people may simply use chatbots more.

The habit is forming early. In a 2025 survey of 1,060 US teenagers, Common Sense Media found that 72% had used AI companions and 52% used them regularly. Among those who did, a third had discussed something serious with an AI instead of a person, though 80% still spent more time with real friends than with AI. Its 2026 census of 1,204 children aged 9 to 17 found loneliness reported by 54% of daily AI users against 33% of those who use it less than monthly or never, a correlation the report itself declines to read as cause. Washington’s 2026 companion-chatbot statute prohibits, for minors, ‘Outputs designed to promote isolation from family or friends’.