A digital shoulder to cry on: Understanding why large language models can be effective in extrinsic interpersonal...

Do large language models (LLMs) provide emotion regulation as effectively as humans? This research investigates when and why LLMs might be preferred as extrinsic interpersonal emotion regulators in nonclinical contexts. Across three studies, we systematically compared the regulatory effects and perceived effectiveness of LLM- versus human-generated comforting messages. Study 1a (N = 279) used thematic analysis to show that LLM-generated responses largely mirrored human extrinsic interpersonal emotion regulation strategies, with some discrepancies between different LLMs. Study 1b (N = 390) demonstrated that LLMs sometimes produced stronger regulatory effects, though effects varied by emotional context (no differences for sadness) and outcome (no regulatory effects on fear). Study 2 (N = 196) again showed LLM regulatory advantages in some scenarios and outcomes, but emotional validation (i.e., acknowledgment of the target's emotional response) did not explain the higher emotional improvement achieved by LLMs. Study 3a (N = 188) identified actionable support (i.e., specific and implementable regulatory tactics) as a feature that enhanced regulatory effectiveness across sources. This was confirmed in Study 3b (N = 180) while controlling for message length. These findings suggest that, when evaluated using experimentally generated messages based on prototypes, LLMs can effectively mimic or sometimes surpass human prototypes in extrinsic interpersonal emotion regulation. However, advantages seem to be context and outcome dependent. Importantly, actionable support emerges as a key component for effective interpersonal emotion regulation across sources. Implications are discussed for theories of interpersonal emotion regulation and artificial intelligence-human interaction. (PsycInfo Database Record (c) 2026 APA, all rights reserved).