Tag
The paper introduces a causal framework to analyze occupational bias in language models, revealing that representational biases can persist even when behavioral metrics show no disparity, and these biases may influence downstream behavior under intervention.