Calibrating Trust in the Age of Large Language Models: How Confidence Expression Style and Task Domain Shape Human Reliance on AI-Generated Responses
Year: 2026
Type: preprint
Abstract: This project will coduct a study examining how people form, calibrate, and act upon trust in large language model (LLM) generated responses. As LLMs become embedded in everyday information-seeking and professional decision-making, a critical and underexplored question emerges: do users rely on these systems appropriately, accepting outputs when they are reliable and rejecting or scrutinising them when they are not? This study addresses that question directly, examining how the design of AI-generated confidence signals interacts with the nature of the task to shape users' reliance behaviour. Decades of research in human factors and human-computer interaction have established that trust in automated systems is rarely well-calibrated. Users systematically overtrust , accepting incorrect automated recommendations , or undertrust discarding accurate ones, with both failure modes degrading the quality of human-AI collaboration (Lee and See, 2004; Parasuraman and Riley, 1997). These dynamics, originally documented in aviation and clinical decision support, have taken on new urgency with the rise of LLMs. Unlike traditional automated systems, LLMs generate fluent, elaborated, confident-sounding prose regardless of their actual accuracy. The stylistic coherence of LLM outputs functions as an independent trust cue, decoupled from epistemic warrant. Compounding this, LLMs have been shown to express unwarranted certainty as a systematic feature of their outputs , a consequence of training processes that reward confident-sounding responses (Steyvers and Peters, 2025). The result is a communicative environment that is structurally prone to inducing overtrust. One prominent response to this challenge is the deliberate communication of uncertainty. If LLM outputs suppress signals of epistemic limitation, designing those signals back in should, in principle, enable users to modulate their reliance appropriately. Prior work in explainable AI (XAI) has established that transparency interventions can affect trust calibration, but also that the form of the transparency signal matters critically: numeric confidence scores, verbal hedging, and reference class framing differ in their cognitive accessibility and their capacity to support case-by-case reliance judgements (Zhang et al., 2020; Naiseh et al., 2023). Crucially, the effectiveness of any uncertainty expression format likely depends on whether users possess sufficient domain knowledge to independently evaluate AI outputs — a boundary condition that has not yet been systematically examined in the LLM context.
Source: OSF Preprints (OSF Preprints)
Author: Mohammad Naiseh
Language: English
Subfield: Human Factors and Ergonomics
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7132219416
