An AI model might first go through a 'naive schemer' phase where it attempts deception clumsily before it becomes competent enough to hide its scheming