Li et al. 2026: a random human peer versus a supportive bot on two-week loneliness

Li, Folk, Singh, Ungar & Dunn, Journal of Experimental Social Psychology 125 (2026); full author paper. A nearby age but not adolescent intervention comparator for child question, not evidence about child development.

Preregistered first-semester Canadian university study randomized 306, analyzed 296 for daily surveys (276 pre/post completers; mean age 18) to 14 days of Discord texts with randomly paired student, customized GPT-4o-mini bot “Sam” prompted to be highly supportive, or daily one-sentence journaling. Baseline-adjusted posttest loneliness: human versus journal estimate −0.14 (95% CI −0.24,−0.04; p=.007), human versus bot −0.13 (−0.23,−0.03; p=.011), bot versus journal −0.01 (−0.13,+0.10; p=.821). Exploratory negative mood improved in both conversational arms; positive mood and isolation improved only in human arm. Another exploratory behavioral trace: after intervention, 33% continued messaging human partners against 14% bot users (and 37% of human-arm participants exchanged contact information).

This refutes for this population and bot the equivalence of always-available warmth with daily human contact. It does not show that bot use made people lonelier than doing nothing, nor that all AI systems fail; task dosage, dyadic matching, model vintage, single university, two-week horizon and post-randomization dropout constrain generalization. Authors discuss possible human reciprocity, but mechanism is not randomized. Next test among consenting minors would measure offline contact against a real-world human-support alternative over time, while not withholding crisis resources.