When Hedging Hides Heterogeneity: A Measurement and Identification Framework for Uncertainty in Large Language Models
Year: 2026
Type: article
Abstract: Large language models routinely hedge on contested questions, but the surface uniformity of such hedging conceals a heterogeneous mixture of underlying latent drivers. We distinguish four latent drivers: epistemic uncertainty, hermeneutic openness, alignment-induced policy effects (which can suppress or amplify evaluative stance), and cross-context interpretive divergence. We argue that verbalised output is best understood as a measurement-and-identification problem in which these heterogeneous latent drivers of uncertainty and stance are compressed into a single output channel by a noninjective verbalisation map: similar hedging or balancing language may reflect genuine uncertainty or instead policy and contextual effects, and a user cannot reliably distinguish these drivers from a single output. We formalise this compression, characterise the decision-theoretic regret of conflating drivers whose appropriate downstream interventions differ, and observe that this regret does not vanish under additional factual evidence when the dominant driver is non-epistemic. Empirical claims are anchored in a pre-specified replication study. We then introduce four families of identifying probes, namely factual-evidence augmentation, counterfactual rephrasing, standpoint-conditioned elicitation, and cross-lingual consistency, which provide directional identification of the dominant latent driver of an observed hedge. We close with implications for professional practice in journalism, policy analysis, and comparative research; design principles for uncertainty-disclosure interfaces; and a benchmark proposal for driver-stratified evaluation of LLM uncertainty.
Author: Yao Zhao
Institution: Temple University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225989453
