The quality of an AI outcome depends heavily on the institution in which it is used: its incentives, standards, and capacity to correct mistakes.