Systems that approximate the world without genuine representation of it cannot validate their outputs, making them vulnerable to both unintentional errors and intentional abuse, and this represents a poor foundation for AI safety.