How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study

Key Findings

• — No significant effects of interaction mode or conversation type were detected on any of the four psychosocial outcomes at the group level, despite measurable differences in AI and human conversational behaviour across conditions.
• — Participants spent an average of 5.32 minutes a day with the chatbot (range 1.01 to 27.65), more with voice than text, most with the engaging voice, and more in open-ended than in personal or non-personal conversation.
• — Within-condition daily duration predicted higher loneliness (beta 0.02, p = 0.027), less socialization with people (beta -0.05, p = 0.0019), more emotional dependence (beta 0.06, p < 0.001) and more problematic use (beta 0.02, p = 0.017); mediation analysis found duration carried modality and task effects onto socialization and dependence.
• — Baseline loneliness and socialization showed negligible correlation with later usage duration (Spearman rho 0.1 and -0.09), so lonelier participants did not simply use the chatbot more.
• — Higher trust in and social attraction toward the chatbot were associated with higher emotional dependence and problematic use; the authors name daily duration as a signal to monitor and soft caps or break nudges as levers to test in future trials.
• — The authors state that duration was not manipulated, so the direction of the duration-outcome relationship cannot be settled by this design.

Methodology Notes

Preregistered four-week randomized controlled trial, n = 981 adults recruited for daily use, nine conditions (3 modes x 3 conversation types), weekly self-report scales (loneliness, socialization, emotional dependence via an ADS-9 craving subscale, problematic use), plus automated classifier analysis of conversation content. Version 1 posted to arXiv 2025-03-21; version 2 posted 2025-10-02 (70 pages with supplement). A Research Square preprint of the same study (10.21203/rs.3.rs-8148142/v1, posted 2026-04-27) is marked 'Under Review' and not peer-reviewed as of 2026-09-15. Companion study to the platform-scale affective-use analysis published by the same OpenAI and MIT team in April 2025.

Sources