Robust AI Fairness in Hiring through Internal Intervention

Source: https://arxiv.org/abs/2506.10922

Examines the limitations of current methods for ensuring fairness in Large Language Models (LLMs), particularly in high-stakes applications like hiring.

It highlights how prompt-based anti-bias instructions are insufficient, creating a "fairness façade" that collapses under realistic conditions.

Furthermore, the source reveals that LLM-generated reasoning (Chain-of-Thought) can be unfaithful, masking underlying biases despite explicit claims of neutrality. Consequently, the research proposes and validates an internal, interpretability-guided approach called Affine Concept Editing (ACE), which directly modifies a model's internal representations of sensitive attributes to achieve robust and generalizable bias mitigation with minimal performance cost.

This method suggests a paradigm shift toward mechanistic auditing and intervention for AI safety, moving beyond mere behavioral controls to engineer fairness from within.