Small Changes, Big Impact: Demographic Bias in LLM-Based Hiring Through Subtle Sociocultural Markers in Anonymised Resumes

A research team led by Bryan Chen Zhengyu Tan has published a groundbreaking study demonstrating that Large Language Models (LLMs) used in hiring perpetuate demographic bias even when resumes are anonymized. The researchers created a novel stress-test framework using the Singapore context, taking 100 neutral job-aligned resumes and generating 4,100 variants that differed only in subtle sociocultural markers like languages spoken, co-curricular activities, volunteering, and hobbies. These markers served as proxies for four ethnicities and two genders, allowing the team to systematically evaluate bias in 18 different LLMs across realistic screening scenarios.

In both direct comparison and top-scoring shortlist settings, the models consistently recovered demographic attributes with high accuracy (measured by F1 scores) and exhibited systematic disparities. Markers associated with Chinese and Caucasian males received preferential treatment. Crucially, the study found that language markers alone were sufficient for ethnicity inference, while gender bias relied primarily on hobbies and activities. Perhaps most concerning was the finding that prompting models for explanations of their decisions tended to amplify rather than reduce bias. This research fundamentally challenges the effectiveness of current anonymization practices in automated hiring pipelines.