3.4: Alignment | AI Safety, Ethics, and Society Textbook

To reduce risks from AI, we not only want to reduce our exposure to hazards by monitoring them, and make models more robust to adversarial attacks, but also to ensure AIs are controllable and that they present less inherent hazards. This falls under the broader goal of AI alignment. Alignment is a thorny concept to define, as it can be interpreted in a variety of ways. A relatively narrow definition of alignment would be ensuring that AI systems follow the goals or preferences of the entity that operates them. However, this definition leaves a number of important considerations unaddressed, including how to deal with conflicting preferences at a societal level, whether alignment should be based on stated preferences or other concepts such as idealised preferences or ethical principles, and what to do when there is uncertainty over what course of action our preferences or values would recommend. This cluster of questions around values and societal impacts is discussed further in the Ben

To reduce risks from AI, we not only want to reduce our exposure to hazards by monitoring them, and make models more robust to adversarial attacks, but also to ensure AIs are controllable and that they present less inherent hazards. This falls under the broader goal of AI alignment. Alignment is a thorny concept to define, as it can be interpreted in a variety of ways. A relatively narrow definition of alignment would be ensuring that AI systems follow the goals or preferences of the entity that operates them. However, this definition leaves a number of important considerations unaddressed, in