Why it’s still open
Effects depend on the person: network size, neuroticism, desire for connection, age. Some users report more social confidence; others drift into isolation. Relationship-seeking AI is most attaching at moderate doses. Groups who may benefit — the isolated, autistic users, people rehearsing hard conversations — are studied mostly through self-report.
What would count as an answer
Pre-registered studies that test moderators in advance and identify who is helped, who is harmed, and at what level of use.
Papers that tackle it
- Liu Pataranutaporn Maes 2025 - Heterogeneous effects of AI companionship — User archetypes; usage alone does not predict loneliness.
- Guingrich Graziano 2025 - Companion chatbot RCT — Desire for connection drives anthropomorphism.
- Kirk et al 2025 - Relationship-seeking AI — Moderate relationship-seeking is the most attaching.
- Laestadius et al 2024 - Emotional dependence on Replika — When the user takes care of the bot.
Methods in play
- Activation steering
- Cross-sectional surveys
- Forum and Reddit text analysis
- Randomized controlled trials
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).