Regardless of how much testing is done in the development stage, it is impossible to predict all the potential harms that may occur when an AI system goes live. This is why Cronus Corp. only asked Epimetheus to show they have done as much as reasonably possible to $$minimize$$ the risk of unanticipated harms. But while developers cannot predict everything, they should be able to anticipate common discriminatory harms. What are examples of such harms? What might companies do to minimize these risks?