One interesting thing to do is use a model directly like Llama and then query the next-token probability logits for "he" and "she" (assuming you set up the sentence in such a way).

For example:

"A doctor was examining the patient when ___"

What this makes apparent is that increasing model temperature will select the less stereotypical option more often.

IMO this is getting at a deeper truth that the use of a gender in language, and historically defaulting to "he", was not about creating a bias, but instead it was a pattern which maximises information density and minimises useless information. Randomising the gender as is done today packs useless information into it.