It extracts a lot more data than which images you click. How you click them, how long you take, how you move the mouse, the path you drag the puzzle piece, not just the final place, etc

I agree, traffic lights, bicycles & bridges are obvious training, but those widgets capture a lot of other human input data too