Consistent but Miscalibrated: Evaluating LLM Limitations for Risk Communication in Natural Language

2026-07-04

AI Summary

This study systematically evaluates the consistency and calibration of large language models (LLMs) in translating probabilistic risk—encompassing likelihood and uncertainty—into natural language descriptions under zero-shot conditions. Employing a two-stage prediction pipeline that integrates Beta-distribution-simulated upstream model outputs, multi-context prompting, temperature sampling, and repeated trials, the work presents the first comprehensive assessment of nine LLMs across diverse domains and settings. The findings reveal that while LLMs maintain descriptive consistency, they exhibit pervasive miscalibration, particularly when articulating uncertainty. Providing statistical summaries—such as the mode and prior sample size—reduces context sensitivity but fails to fundamentally resolve calibration inaccuracies. This work thus uncovers critical limitations of LLMs in risk communication and offers vital insights for developing trustworthy AI applications.

Abstract

LLMs are increasingly deployed as post-hoc explainers of AI-generated outputs, yet it remains unclear whether they can reliably communicate probabilistic information in natural language. For this role to be viable, models must produce identical verbal descriptions for identical inputs, and select descriptions that accurately reflect the magnitude of the underlying numerical quantities. We evaluate whether nine LLMs meet these requirements within a two-stage prediction pipeline, in which an upstream model has produced probabilistic outputs characterized by their likelihood and uncertainty, and LLMs are tasked with selecting an appropriate verbal descriptor for each. We simulate predictions from an upstream model by taking samples from a Beta distribution parameterized by its mode and prior sample size. We then prompt LLMs to explain these predictions under six domain contexts and with ten temperature settings, and repeating each experiment ten times. We find that LLMs are generally consistent but miscalibrated, with substantially weaker performance on uncertainty than on likelihood tasks. Providing models with precomputed summary statistics (mode and prior sample size) reduced sensitivity to contextual framing but did not resolve the underlying miscalibration, suggesting that the bottleneck resides in the verbalization step itself. These findings indicate that current LLMs do not yet constitute reliable zero-shot standalone risk communication tools for probabilistic predictions.

Problem

• risk communication
• large language models
• probabilistic information
• calibration
• natural language

Innovation

• risk communication
• probabilistic calibration
• verbal probability expressions
• large language models
• uncertainty quantification