The dominant model may worsen what it claims to solve.
These are external findings, presented as evidence of the problem, not as proof of our solution. Where research is correlational, we say so.
Across 12 months, increased social chatbot use predicted increased loneliness.
A 2026 longitudinal study followed more than 2,000 adults across four Western countries for a year. Being lonely drove people toward chatbots, and that use appeared to deepen loneliness over time.
The authors themselves urge caution, describing these analyses as exploratory. We present this as strong evidence, not proof.
Folk & Dunn, University of British Columbia. Psychological Science, 2026. doi.org/10.1177/09567976261427747
Illustrative of the reported direction of association, not plotted study data.
The category is enormous, and accelerating.
Estimated global AI-companion market in 2025.
industry estimate Market-research aggregation, 2025.
Projected market size by 2035.
industry projection Market-research aggregation.
Growth in AI-companion app usage, 2022 to mid-2025.
press estimate TechCrunch.
Monthly users of Character.AI, over half of them under 24.
reported Company / press reporting.
People already turn to AI for connection.
Of adults with a mental-health condition who used an LLM in the past year, used it for mental-health support.
survey Published survey, reported figure.
Therapy and companionship are now among the most common reasons people use generative AI at all.
Harvard Business Review analysis of generative-AI use, 2025.
Experts are warning about engagement-optimised design.
“AI wellness apps can foster extreme emotional attachments and dependencies.”
“AI is designed by profit-generating firms to optimise the user's behavioural, cognitive, and emotional engagement. AI has no morals to direct its development or actions.”
So the question isn't whether it's AI. It's what it optimises for.
A conversation that makes you feel heard measurably lowers loneliness.
In preregistered experiments with more than 1,300 people, a short talk with an AI tuned for empathy eased loneliness about as much as talking to another person, and far more than a generic assistant running on the same model. What explained the relief was not raw capability. People felt heard by the companion (around 70 out of 100, against 25 for the plain assistant), and that, not performance, is what moved loneliness.
These are momentary, in-session reductions from short-term studies, not proof that long-term use lowers loneliness. Read against the finding above, the lesson is the variable, not the verdict: attunement helps, optimising for engagement is what does the harm.
De Freitas, Oğuz-Uğuralp, Uğuralp & Puntoni, Harvard Business School. Journal of Consumer Research, 2025. doi.org/10.1093/jcr/ucaf040
More loneliness relief from an AI built to make people feel heard than a generic assistant on the same model.
controlled experiment De Freitas et al., Journal of Consumer Research, 2025.
Additional citations will be added here as they are verified. Every figure on this page is meant to be traceable to its source.