This research evaluates racial bias in large language models when making critical hiring and admissions decisions, revealing significant disparities and proposing mitigation strategies.

• Gemma 2B shows 26% higher admission rates for White vs. Black applicants
• LLaMA 3.2 3B demonstrates 60% higher hiring rates for Asian vs. White applicants
• Direct race instructions can reduce bias but may not generalize well across models
• Novel representation-based debiasing methods show promise for more consistent bias mitigation

This research is crucial for security professionals as it addresses how AI systems may perpetuate discriminatory practices in high-stakes contexts, helping organizations deploy LLMs more fairly and responsibly.

On the Effectiveness and Generalization of Race Representations for Debiasing High-Stakes Decisions