Researchers develop BADx metric to detect how large language models amplify biases when adopting different personas

Existing bias testing methods (CEAT, I-WEAT, I-SEAT) fail to capture how LLM biases shift dynamically when models take on social roles

New BADx metric combines differential bias scores, Persona Sensitivity Index, and volatility measurements to quantify persona-induced bias amplification

BADx integrates LIME-based explainability analysis to provide transparent insights into how and why biases emerge in persona-driven contexts

Study reveals that LLMs embed and amplify intersectional biases especially when operating under persona-driven scenarios