First Year CS Students Exploring And Identifying Biases and Social Injustices in Text-to-Image Generative AI

Abstract

Generative AI is a recent breakthrough in AI. While it has become a hot topic in computing education research (CER), much of the recent research has focused on e.g. issues of plagiarism or academic integrity. One problem spot with Generative AI is its susceptibility to various kinds of algorithmic bias. In this study, we collected data from an introductory computing course, where students experimented with text-to-image generative models and reflected on their generated image sets, in terms of biases, related harms, and possible fixes. Data were collected in Fall 2023 (pilot data in Fall 2022). Data included reports from 163 students. The results show (1) a variety of bias types observed by students related to gender, ethnicity, age, as well as a variety of bias types not observed by students, (2) two major types of attributions for the source of bias: bias caused by biases in the society and bias caused by data or algorithms, and (3) a number of potential harms associated with the biases, as well as attributions of those harms in specific contexts and use cases.