People offload empathy when they think AI can give it instead

Key Findings

• — Study 1 (n = 249): perceiving ChatGPT as capable of empathy predicted willingness to offload empathy to it (b = 0.44, SE = 0.05, t(247) = 9.64, p < .001); mean perceived capacity was 3.24 on a 7-point scale
• — Study 3 (n = 743): after learning an acquaintance had gone to ChatGPT for emotional support, participants asked fewer questions about how to help (M = 0.70 vs 0.98 with no prior support) and scored lower on an emotion-reading task (M = 7.13 vs 7.73 correct, p = .012), while explicit ratings of personal responsibility did not differ (F(2, 740) = 1.32, p = .27)
• — Study 4 (n = 737, exploratory after a lost preregistration): when a real student was described as getting empathy from an AI chat platform, participants selected the fewest empathy words (M = 18.68 vs 21.84 for friends) and the most indifference words (M = 2.39 vs 0.90 for friends) of all conditions
• — Participants rated the student who turned to AI as more weird, antisocial and desperate than when the student turned to friends or a therapist; the authors suggest such users may receive less human empathy because of how they sought support

Methodology Notes

US adults recruited on CloudResearch Connect; Studies 1-3 preregistered, Study 4 treated as exploratory because its preregistration was overwritten. Vignette-based and quasi-behavioural measures (question writing, Reading the Mind in the Eyes Test, weighted word-cloud selection); no real support interactions. The Study 4 AI condition alone cited 'a recent research paper from Harvard' (Zao-Sanders, 2025) on therapy and companionship as top AI uses, an asymmetry the authors acknowledge. Total n about 2,223 across four studies (249; 494; 743; 737). Published online 28 October 2026 in Computers in Human Behavior vol. 187 (February 2027 issue), open access under CC BY 4.0. The journal abstract (read via OpenAlex) carries no figures; ScienceDirect full text was not reachable (captcha), so the figures come from the authors' preprint (PsyArXiv bsj9f, version 3 posted 23 October 2026, whose manuscript title matches the journal title). Data, code and preregistrations on OSF.

Sources