Using a simulation or model allows a large number of training runs to be packed into a short period of time, and constrains the output of the AI. But the AI is then tainted by the model, it effectively becomes an extension of the model.