Can an AI be “unbiased”?
New research shows that even so-called “aligned” AIs — trained to follow human values and avoid harmful outputs — still reflect stereotypes. Even GPT-4, one of the most advanced models, repeats the very biases it was meant to suppress.