Abstract

Despite significant progress, the AI alignment literature has largely overlooked a philosophical dilemma:\textit {humans and machines represent and communicate uncertainty in fundamentally incompatible ways.} To formalize the arguments around this dilemma, we introduce a four-quadrant taxonomy of uncertainty that partitions uncertainty according to two dimensions: whether uncertainty arises from a human or machine agent, and whether it remains internal or is communicated externally. The taxonomy comprises\textit {human inherent uncertainty}(Bayesian credence),\textit {human self-reported uncertainty}(linguistic hedges),\textit {model inherent uncertainty}(frequentist risk), and\textit {model self-reported uncertainty}(preference-optimized language). We show that this taxonomy reveals an epistemological gap: humans assign uncertainty through grounded, causal understanding, while models assign it through risk-minimization over training data. When models trained with RLHF generate hedged language, humans interpret it as Bayesian belief vis-à-vis frequentist risk-penalty. We close with a practical case study and argue that bridging this epistemic gap is essential-but unfortunately neglected in current discourse-for establishing trust in complex human-AI collaboration modes.

Date

2026

Authors

Mayank Kejriwal