🚨New #ACL2025 paper! Today’s “safe” language models can look unbiased—but alignment can actually make them more biased implicitly by reducing their sensitivity to race-related associations. 🧵Find out more below!