At best, it has probabilistic biases. OpenAI had to train newer models to not favor the name "Lily."

They have to do this manually for every single particular bias that the models generate that is noticed by the public.

I'm sure there are many such biases that aren't important to train out of responses, but exist in latent space.