Implicit bias in language models

Reasoning models spend extra tokens when a prompt cuts against a stereotype

A team adapting the implicit association test to reasoning traces found four of five models working harder on association-incompatible prompts, which puts a measurable bias signal in the process rather than only in the answer.