How Language Model Hallucinations Can Snowball

TL;DR

This paper investigates how language models' hallucinations can escalate through a process called snowballing, where initial errors lead to more mistakes, and finds that models can often recognize their own errors.

Contribution

It introduces the concept of hallucination snowballing and demonstrates that models like ChatGPT and GPT-4 can identify a significant portion of their own mistakes.

Findings

ChatGPT recognizes 67% of its mistakes

GPT-4 recognizes 87% of its mistakes

Hallucination snowballing leads to cascading errors in LMs

Abstract

A major risk of using language models in practical applications is their tendency to hallucinate incorrect statements. Hallucinations are often attributed to knowledge gaps in LMs, but we hypothesize that in some cases, when justifying previously generated hallucinations, LMs output false claims that they can separately recognize as incorrect. We construct three question-answering datasets where ChatGPT and GPT-4 often state an incorrect answer and offer an explanation with at least one incorrect claim. Crucially, we find that ChatGPT and GPT-4 can identify 67% and 87% of their own mistakes, respectively. We refer to this phenomenon as hallucination snowballing: an LM over-commits to early mistakes, leading to more mistakes that it otherwise would not make.