They can’t improve the models in that way. They are non-determainistic and don’t listen to instructions. They can tinker with training data but straight up forcing a model to work a certain way take all of the magic away. Without the magic, it would be worse and everything would look more uniform.