Does Model Uncertainty Track Human Ambiguity? Evidence from Multi-Annotator Vision Benchmarks

Models fail to recognize ambiguous cases that humans identify: On CIFAR-10H, 50.4% of images received multiple valid human classifications, yet models confidently settled on a single label. This represents a critical gap where humans would flag cases for expert review, but AI systems proceed with unwarranted confidence.

Standard uncertainty measures don't track human disagreement: Model confidence scores and entropy showed only weak correlation with human disagreement (correlation coefficients of 0.24–0.55 on tested datasets), meaning current methods for detecting uncertain predictions cannot reliably identify the ambiguous cases that humans struggle with.

Multiple models don't solve the problem: Even when using predictive multiplicity—measuring disagreement across eight different models—the improvement in identifying human-ambiguous cases was only modest. This suggests the issue runs deeper than individual model limitations.

Facial emotion recognition shows even starker misalignment: FER+ images showed 33.5% receiving multiple valid human labels, yet models converged on single predictions, indicating particular challenges in subjective domains where reasonable disagreement is common and expected.

Current AI systems are unprepared for genuinely ambiguous decisions: The findings reveal that widely-used uncertainty quantification methods cannot reliably flag instances where human judgment would call for caution or expert involvement, creating a hidden risk in AI-assisted decision-making systems.