Vincent Conitzer (Duke) - Moral Artificial Intelligence: How to Learn Objectives from People

Moral Artificial Intelligence: How to Learn Objectives from People

Algorithms in machine learning (ML) and artificial intelligence (AI) generally require us to specify an objective function, which formalizes what it is that we want our algorithm to optimize. While these algorithms were confined to the laboratories in which they were developed, the exact objective function often did not matter much. But now that they are being broadly deployed in the world, we find that simplistic objectives often have unintended side effects. For one, AI systems increasingly need to make decisions with a moral component.  E.g., should a self-driving car prioritize the safety of its passengers over that of others, and to what extent? I will briefly discuss some general approaches to such problems.

I will then go into detail on the application of these techniques to kidney exchanges (no prior familiarity required). A kidney exchange allows patients who are in need of a kidney transplant, and who have willing but incompatible donors, to exchange donors. Some real kidney exchanges use algorithms to determine an optimal matching. Should such an algorithm take features such as the patient's age into account, and to what extent? What would be the consequences for a reasonably large exchange?

Finally, if time permits, I will discuss in more detail the problem that, generally, not everyone will agree on what the morally preferred option is, and how this can be addressed using techniques from computational social choice and computational learning theory.

Host: Rebecca Willett