(13th-January-2021)
• The Ethics of Learned Environment Models, For an AI system like Google's self-driving car, the environment model is specified by engineers as part of the system design. The engineers can generally anticipate the possible situations the car will encounter and design safety into the car's responses. If the car is unable to recognize objects in the road, it can simply stop. However, more complex AI systems in the future will need to learn their environment models and hence their designers will be less able to anticipate the situations they will encounter or to specify safe resonses.
• When future AI systems learn to use human languages fluently and learn about human society, then they will learn about human laws, morality and ethical theories. Will they use this knowledge to guide their own ethical behavior? Not necessarily. For example, people with antisocial personality disorder (American Psychiatric Association 2013) know about social norms but often use them to predict and manipulate other people rather than to guide their own behavior. Knowledge of law, morality and ethics will contribute to an AI agent's environment model but not directly to its utility function. However, if human approval increases the value of an agent's utility function, then the agent may use its knowledge of social norms to choose behaviors that will please people.
Comments