An Evaluation of Estimative Uncertainty in Large Language Models

TL;DR

This paper evaluates how well large language models like GPT-4 understand and communicate estimative uncertainty compared to humans, revealing both alignments and divergences across contexts and languages.

Contribution

It provides a comparative analysis of estimative uncertainty in LLMs versus humans, highlighting strengths and gaps in LLMs' ability to interpret and map uncertainty.

Findings

GPT-3.5 and GPT-4 align with human estimates for some WEPs

Divergence occurs with gendered roles and Chinese contexts

GPT-4 can map between statistical and estimative uncertainty

Abstract

Words of estimative probability (WEPs), such as ''maybe'' or ''probably not'' are ubiquitous in natural language for communicating estimative uncertainty, compared with direct statements involving numerical probability. Human estimative uncertainty, and its calibration with numerical estimates, has long been an area of study -- including by intelligence agencies like the CIA. This study compares estimative uncertainty in commonly used large language models (LLMs) like GPT-4 and ERNIE-4 to that of humans, and to each other. Here we show that LLMs like GPT-3.5 and GPT-4 align with human estimates for some, but not all, WEPs presented in English. Divergence is also observed when the LLM is presented with gendered roles and Chinese contexts. Further study shows that an advanced LLM like GPT-4 can consistently map between statistical and estimative uncertainty, but a significant performance gap remains. The results contribute to a growing body of research on human-LLM alignment.