If you shut your eyes, you do not see anything in the room. This does not mean there is nothing in the room.

If you randomize your seeds, you do not deliberately bias your model. This does not mean your model is unbiased.

Intelligence does not start from a random state, but from a pretrained one. Random state is likely to roll an artificial overfit or if it just fits, you won't know why it is better than another state - or rather why your training method sucks from other initial states.

Proper methods of bias analysis are network compression or random surgery, and expanding test data set. Proper methods of training are robust to varied initial states.