Do Personal AI Conversations Reduce Loneliness? Evidence from a Pre-Registered Experiment

Key Findings

• — Treated participants increased the number of days with a personal AI conversation by about 5 (first-stage coefficient 4.77 with the pre-specified missing-value indicator; 5.16 on complete cases); a broad AI-use index rose by 0.43 SD and its socio-emotional subscale by 0.46 SD.
• — Assignment to daily personal AI conversation increased single-item loneliness by 0.163 points on a 0-10 scale (0.056 SD; control mean 3.47; Romano-Wolf p = 0.001).
• — UCLA-3 loneliness (+0.025 SD), life satisfaction (-0.029 SD), happiness (-0.013 SD) and depression (+0.029 SD) all moved in the adverse direction but were not statistically significant (Romano-Wolf p 0.344-0.495).
• — Pre-specified heterogeneity analysis gave larger point estimates for respondents with a partner; exploratory analysis suggested adverse effects concentrated among respondents in smaller urban units.
• — The authors read the pattern as consistent with substitution, rumination and AI validating grievances, and contrast it with single-session and purpose-built-companion studies that found loneliness reductions.

Methodology Notes

Individual-level randomized controlled trial pre-registered on the AEA RCT Registry (AEARCTR-0016749). Adults resident in France recruited from the Bilendi online panel; baseline wave 5-19 January 2026, intervention 22 January to 18 February, endline 19-28 February (a later cohort recruited through a second partner followed a three-week-delayed schedule and added 106 participants). Analysis sample 7,286 (3,642 control, 3,644 treatment) who passed the baseline attention check and completed both waves; endline completion 74.8% vs 74.7%. ANCOVA intention-to-treat estimates with baseline outcome and pre-specified covariates, Romano-Wolf correction within outcome families, Lee bounds for attrition. Compliance self-reported at endline; the sample is older, more educated and more often partnered than the French population. Working paper, not peer-reviewed. Dates: HAL version 2 released 2026-10-02 (working paper dated October 2026); SSRN posting 2026-10-17. Verification: the HAL PDF is behind an Anubis bot challenge and SSRN behind Cloudflare; the full 30-page PDF was read from a Wayback Save Page Now capture made 2026-10-19 (20261019020413); abstract matched via the HAL API and Crossref.

Sources