Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction

Key Findings

• — Engagement is persistent: baseline interaction intensity predicted follow-up intensity (beta 0.43), companionship use (beta 0.32) and self-disclosure (beta 0.28), and baseline companionship use predicted follow-up companionship use (beta 0.49), all p below 0.001 among the 439 follow-up respondents.
• — Sustained social engagement on every dimension (intensity, companionship use, self-disclosure) was associated with lower well-being at follow-up.
• — The association was mainly explained by less time spent interacting with other people in person, supporting a social-displacement pathway over a direct-effect account.
• — Companionship-oriented descriptions of the chatbot relationship rose from 173 participants at the first wave to 217 at follow-up among the 331 free-text responders who continued using chatbots (version 2 figures; version 1 reported 203 and 227); 108 follow-up respondents reported having stopped using the chatbot.
• — Retention was 37.1% (dropout 62.9%); 59.54% of baseline participants were no longer active on the recruitment platform at follow-up, and the authors report attrition-bias checks.

Methodology Notes

Two-wave panel recruited on Prolific; baseline N = 1,182, follow-up N = 439 after a mean of 12 months (first-wave collection September 2024 to September 2025, follow-up August 2025 to May 2026); longitudinal regression and mediation models; well-being measured with the Comprehensive Inventory of Thriving; companionship use measured by baseline primary-use choice and by GPT-5.1-assisted clustering of free-text relationship descriptions; in-person social interaction time at follow-up as mediator. Self-report, Character.AI users only, high attrition; country of residence not stated in the abstract. arXiv 2609.07243 version 1, 7 September 2026 (announced 9 September); not peer reviewed. Revised to arXiv version 2 (submitted 2026-10-18): the abstract is unchanged; the companionship-description counts changed from 203 and 227 to 173 and 217, the wave-specific collection windows were added, and the IRB number and funding acknowledgment were removed. Coefficients, retention (37.1%), the 108 stopped users and the 331 continuing responders are unchanged.

Sources