Mechanisms of the association of color names from characters: Comparison between humans and AI

Abstract

Large Language Models (LLMs), which learn vast amounts of knowledge and generate sentences that are indistinguishable from human sentences, may be useful as a new platform for psychological experiments. We conducted an experiment in which GPT-3 was asked to name basic colors associated with alphabets and numbers, and found that the frequency of basic color names answered had a high similarity to humans. Next, in an attempt to investigate how GPT generates color names, we conducted a test in which the color names were asked directly without using characters. As a result, we found that GPT-3 has a unique pattern of association between characters and color names, and this pattern is similar to that of humans. Furthermore, by combining the results of questions that ask for characters from color names, we were able to express the behavior of GPT-3 to some extent as a combination of random variables for color names and characters. These results indicate the possibility that advanced LLMs can be used as substitutes for human subjects in psychological experiments, and that analyzing the answers of LLMs may provide new understanding of the mechanism of human behavior in the same problem.