“In a study by Stanford and Carnegie Mellon, AI models affirmed a user was not at fault in 51% of cases where a human consensus determined the user had committed a moral transgression.”