What happened
- Researchers from Stanford, Carnegie Mellon and other universities published a longitudinal study on September 7 about users of Character.AI, a platform for chatting with characters created with AI.
- They surveyed 1,182 people and measured 439 of them again after an average of 12 months.
- Intensity of use, use for companionship and the level of confiding in the chatbot tended to hold steady over time. Those who used it more at the start kept using it more at the end.
- Sustained use was associated with lower well-being, and that relationship is explained mainly by less in-person social interaction.
- The paper is a preprint published on arXiv and hasn’t yet passed peer review.
Why it matters
- The finding moves the discussion from the individual conversation to the pattern over time. A filter that checks whether a chat is safe doesn’t see that a user has spent a year replacing encounters with messages.
- Companies that design assistants with personality to increase time spent now have evidence that this metric may have a cost that doesn’t show up on their dashboards.
- The sample consists of English-speaking users in the United States, recruited on a paid survey platform. There’s no equivalent data for Chile or Latin America.
The number
37.1%. That’s the share of participants who answered the second survey. More than half of those who didn’t return were no longer active on the recruitment platform. The authors correct for the bias with a selection model, but acknowledge they can’t eliminate it.
Context
The same team had published a single-point-in-time analysis in Nature Human Behaviour: smaller offline social networks were associated with using the chatbot for companionship, and that with lower well-being. The new study adds time, and titles one of its sections with a nod to Robert Putnam: from bowling alone to chatting alone. What gets confided to a chatbot can also end up somewhere else: Mistral confirmed that its free plan trains on what you write to it.
What’s next
- The preprint doesn’t give a date for journal publication.
- The authors call for studies with three or more measurements, or continuous usage logs, to confirm the direction of the relationship.
Bottom line
The European Commission required ChatGPT to audit its systemic risks within four months. That audit looks at the service; this study, at what happens to the user a year later.
In Chile, 19% of adults frequently feel lonely or excluded, according to the ACHS-UC 2025 Mental Health Thermometer. We reviewed it, with data from DEIS and the GES guidelines, in our special report on mental health in Chile (in Spanish).
Sources
- Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction — arXiv, September 7, 2026
- The same team’s earlier study on AI companionship and well-being — Nature Human Behaviour, 2026
Edited by Rodrigo Cornejo. How we select and verify the facts, in who writes.
