The social intelligence reliance halo: perceived social understanding, epistemic trust, and reliance quality in human-AI decision-making

Abstract

As generative AI systems become more socially responsive, personalized, and perspective-aware, a central question is whether such features improve human–AI decision quality or simply make AI advice more influential. This study introduces the Social Intelligence Reliance Halo as a proposed social-to-epistemic spillover pattern in which perceived social understanding functions as a non-diagnostic cue for epistemic evaluation: an AI that appears to understand the user may also be granted greater epistemic authority or reliability even when the relational cue provides no evidence of factual accuracy. A randomized two-condition experiment was conducted with 462 working adults with recent generative-AI experience. Participants interacted with either a High Social-Intelligence (High-SI) relational framing package or a Neutral AI across six repeated decision trials. Post-interaction measures assessed perceived social intelligence, epistemic authority, epistemic reliability, suspension of critical judgment, verification behavior, and AI literacy; trial-level behavioral reliance was analyzed with generalized estimating equations. The randomized High-SI package increased perceived social intelligence and increased advice-taking both when participants were initially correct and the contingent AI recommendation was wrong (OR = 1.507, p < .001) and when participants were initially incorrect and the recommendation was correct (OR = 1.800, p < .001). The Condition × participant-dependent advice-correctness interaction was nonsignificant (OR = 1.195, p = .289), providing no reliable evidence that the randomized treatment effect differed across these two states. Because recommendation correctness was mechanically determined by participants’ initial accuracy rather than independently randomized, this interaction is interpreted as exploratory treatment heterogeneity across participant-dependent correctness states rather than evidence that the High-SI package altered sensitivity to AI accuracy. Post-interaction perceived social intelligence was positively associated with epistemic authority and reliability, and AI literacy was associated with lower suspension of critical judgment and greater verification; these associations do not establish temporal or causal mediation. The study therefore provides experimental evidence that a bundled relational framing package can amplify both beneficial and harmful reliance, while the proposed social-to-epistemic pathway and accuracy-sensitive reliance remain questions for stronger future designs.