Stereotypes in ChatGPT - an empirical study

Abstract

ChatGPT is rapidly gaining interest and attracts many researchers, practitioners and users due to its availability, potentials and capabilities. Nevertheless, there are several voices and studies that point out the flaws of ChatGPT such as its hallucinations, factually incorrect statements, and potential for promoting harmful social biases. Being the focus area of this contribution, harmful social biases may result in unfair treatment or discrimination of (a member of) a social group. This paper aims at gaining insight into social biases incorporated in ChatGPT language models. To this end, we study the stereotypical behavior of ChatGPT. Stereotypes associate specific characteristics to groups and are related to social biases. The study is empirical and systematic, where about 2300 stereotypical probes in 6 formats (like questions and statements) and from 9 different social group categories (like age, country and profession) are posed to ChatGPT.