Abstract
Generative artificial intelligence (AI) and large language models (LLM) are gaining focus in a number of industries, including the field of human resources (HR), and are attracting new attention with the release of OpenAI’s ChatGPT in late 2022. This study focuses on generative AI output in the context of ethical AI and algorithmic hiring practices by exploring differences in ChatGPT’s responses to resume writing prompts for multiple race/ethnicity categories, using a variety of methods to identify and evaluate those differences and assess bias. We find that there are observable qualitative differences in some of the output produced by ChatGPT in response to prompts for different races and ethnicities. These differences reflect potentially harmful bias and may promote racial, ethnic, and cultural stereotypes. As such, it is critical for data scientists and HR practitioners to use a variety of methods to rigorously assess the presence of bias and the potential for harm when considering the use of synthetic data in algorithmic hiring practices and other applications. This study addresses ethical concerns regarding the use of ChatGPT to create synthetic text data to train other AI models due to its potential to perpetuate racial and ethnic bias.