Any LLM won’t have the right architecture to implement that kind of math. They are built specifically to find patterns, even obscure ones, that nobody knows of. They could start flagging random shit indirectly associated with gender like relative timing between jobs or rate of promotions, etc, and you wouldn’t even notice it’s doing it