The “human in the loop” is meant to be a backstop to prevent flawed AI models from causing harm. But does it work?