Answers are grounded in this paper and cite their sources.
As large language models (LLMs) are increasingly integrated into decision-support systems, understanding how users interpret uncertainty expressions is critical for responsible deployment. Prior work has shown that verbal hedging (e.g., “I'm not sure, but …”) can reduce overreliance on incorrect outputs. We examine whether combining verbal hedging with visual confidence indicators improves trust calibration in a pre-registered 2 × 2 × 2 mixed-design experiment ( N = 495 ). Response accuracy was the dominant driver of trust and reliance across all conditions. Visual confidence indicators improved users’ subjective sensitivity to accuracy, reflected in higher per-participant discrimination indices and steeper confidence-accuracy slopes, yet simultaneously amplified behavioral agreement with incorrect outputs. Verbal hedging reduced agreement and satisfaction without improving discrimination. Combining both cues produced the highest overreliance on incorrect responses alongside the lowest rates of independent source verification. These findings reveal a dissociation between subjective and behavioral calibration as users can become more sensitive to accuracy signals while simultaneously making worse decisions. Simply adding more uncertainty cues does not produce more trustworthy or interpretable LLM systems; what matters is whether signals translate into behavior-relevant calibration.
No takes yet. Share an insight, caveat, or question.
Ojewale et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: