Evaluating the Moral Beliefs Encoded in LLMs

https://arxiv.org/abs/2307.14324

This paper presents a case study on surveying large language models to elicit their encoded moral beliefs. It introduces statistical methods to quantify LLMs' choices, uncertainty, and consistency. By administering a survey with 1,367 moral scenarios to 28 LLMs, the study finds that in unambiguous cases, models align with commonsense, while in ambiguous cases, they often express uncertainty or show varied preferences, with closed-source models tending to agree.

Scherrer et al.