Assessing Ethnic Bias in ChatGPT’s Recruitment Evaluation

Sammanfattning

Artificial Intelligence (AI) has evolved over the past few years and is now widely applied in decision-making systems. An area where AI adoption is growing is recruitment, where automated tools are used to evaluate job applicants. Large language models (LLMs) enable the automated evaluation of textual documents, such as curricula vitae (CVs), because they are trained on large amounts of data available on the internet. Previous studies have shown that applicants with non-Swedish names may face ethnic bias in recruitment processes conducted by human recruiters. This raises concern because there is a risk that the models may reproduce similar biases when evaluating applicants. This study investigated whether ChatGPT, a widely used LLM, shows systematic ethnic bias when evaluating identical CVs for a software developer position in Sweden. This was achieved through a controlled experimental design, in which ChatGPT (GPT-4o) evaluated identical CVs where only the applicant's name varied between Swedish names and non-Swedish names. The model’s output was collected through automated interactions with the ChatGPT API, enabling consistent and repeatable data collection. The study used both quantitative and qualitative methods. Statistical analysis was used to examine differences in suitability scores and interview recommendations, while qualitative analysis was used to analyze the differences in the written feedback generated by the model. The results showed significant differences in suitability scores, interview recommendations, and the written feedback between the two groups. Applicants with Swedish names received higher suitability scores, more interview recommendations, and more favorable written feedback compared to applicants with non-Swedish names. The differences observed in the evaluations may be linked to biases introduced at different stages of the large language model lifecycle.