The Perfection Trap: How Formally Aligned AI Systems May Create Inescapable Ethical Dystopias — LessWrong