Large Language Models Show Human Behavior

Neural networks can approximate any function given sufficiently many hidden units, which implies that they, in theory, can approximate human behavior. Recently, natural language processing has advanced rapidly due to increases in the amount of hidden units and in the size of the datasets. With these advances in natural lan- guage capabilities, we wondered whether state-of-the- art Large Language Models show human behavior. In this article, we demonstrate that these models show language comprehension and communication skills to solve problems, which are considered to be key features of human behavior. Moreover, the process by which such AI-based models encode information leads to er- rors which are also common in humans, such as be- ing vulnerable to misleading questions, source amne- sia, and being sensitive to small changes in wording. Given the similarities with human behavior, we dis- cuss the potential applications of LLMs in social sci- ence research. We conclude that LLMs and their close alignment with human behavior may provide a valu- able source of information that can be studied to gain a better understanding of human behavior.