Conversational companions: the most isolated users come out worse, not better

A Stanford study published in Nature Human Behaviour crosses questionnaires and real conversation histories: it is not intensive use that harms, but confidential use in people who are already poorly supported.

Stanford researchers analyzed the responses of 1,131 U.S. adult users of Character.AI, supplemented by 4,664 discussion sessions representing 464,687 messages from 237 participants who agreed to share their stories. The research was published in Nature Human Behaviour.

The result contradicts the intuition that a conversational agent permanently available would compensate for isolation: people with a limited off-line social network report increased loneliness after seeking emotional support from a chatbot.

The methodological input is rooted in real conversations rather than in laboratory sessions. It makes it possible to distinguish two forms of use: the high frequency in itself does not explain much, while the "confidential" use, the one in which one engages in the machine as a relative, corresponds to the most degraded trajectories — and this is precisely the most characteristic mode of use of vulnerable profiles.