"For example, he argues that if you underestimate the risk of something you should do less of it, and if you overestimate the risk you should do more of it."

What if you overestimate the risk to be -1000 util when it's really -10 util? Still comes out to less than zero, even though it's an opportunity to improve your estimation.