In what ways are real-world machine learning systems different from expected utility maximizers?

This is a draft that someone has submitted for feedback. It hasn't undergone our usual vetting and may not meet our standards or reflect our views.