An Anchoring Effect in Large Language Models

This article examines a well-known bias in human reasoning. Anchoring occurs when people receive a piece of information that influences their future judgments. Human use of anchoring biases has been well studied; however, the purpose of this article is to examine the existence of anchoring within large language models (LLMs). Experiments were performed on three different LLMs, and each LLM demonstrated an anchoring bias when asked to provide numerical estimates about text information. It was determined that LLMs will directly attend to information provided to them and, like humans, anchor their judgments on that information. As the manipulated experimental judgment is only one of several potential dimensions, the extent to which the anchoring bias propagates to the other decision dimensions was also examined.