Abstract
Rapid evolution of artificial intelligence (AI) based on deep neural networks has resulted in artificial systems such as generative pre-trained transformer 3 (GPT-3), which can generate human-like language. Such a system may provide a novel platform for studying how human perception is related to knowledge and the ability of language generation. We compared the frequency distribution of basic color terms in the answers of human subjects and GPT-3 when both were asked similar questions regarding color names associated with the letters of the alphabet. We found that GPT-3 generated basic color terms at a frequency very similar to that of human non-synaesthetes. A similar frequency was observed when color names associated with numerals were tested indicating that simple co-occurrence of alphabet and color word in the trained dataset cannot explain the results. We suggest that the proposed experimental framework using the latest AI models has the potential to explore the mechanisms of human perception.