Recalled Emotion Regulation in Matched Human and LLM Chatbot Support Episodes

Key Findings

• — 583 Prolific participants who had talked to both a person and an LLM chatbot about the same or a comparable personal topic in the past week reported both episodes in randomised order.
• — Both episode types were associated with large recalled distress reductions: AI dz = 1.22, human dz = 1.00.
• — The difference met the preregistered equivalence criterion (dz = 0.20; pTOST = .009), so the two sources were statistically equivalent on recalled distress change.
• — Contrary to the authors' prediction, attunement, emotional expression, processing and same-source help-seeking intention showed no detectable difference by source, with 95% confidence intervals within dz = plus or minus 0.12.
• — Conclusions held when the 479 excluded cases were restored and when topic domain was covaried.
• — Felt closeness to the source was higher for humans and predicted attunement in both episode types, processing in human episodes and distress reduction in AI episodes.
• — A reflexive thematic analysis of 619 topic descriptions identified four needs participants brought to both sources.

Methodology Notes

Preregistered (osf.io/pesrh) within-subject retrospective design; recalled outcomes reported after the fact for both episodes, so this measures memory of support rather than measured affect change, and recall of a chatbot conversation may be differently biased from recall of a human one. Participants are recent dual-source users, a selected group. The authors state that whether the underlying relational processes or longer-term outcomes are similar remains open. Preprint, not peer-reviewed; v2 posted 2026-09-09, DOI 10.31234/osf.io/y65d4_v2. No institutional affiliations are listed for any of the three authors, which is why credibility is graded preliminary; the design and preregistration are otherwise clean.

Sources