An important aspect of AI ethics and safety is aligning an AI system's behavior with the preferences (and values) of its owners, users and people affected by its actions.  But often the preferences of a group of people are diverse and not aligned themselves. Voting theory provides many different methods that aggregate the preferences of a group of people.  In the first part of the talk, I will survey  different voting methods highlighting their benefits and costs using examples from recent elections.  The second part of the talk will address a central issue is social choice theory and mechanism design: the manipulability of voting methods. Classic results show that any reasonable preferential voting method sometimes gives individuals an incentive to report an insincere preference  instead of their true preference. The extent to which different voting methods are more or less resistant to such manipulation has become a key consideration in the literature on comparing voting methods. In computational social choice, one standard measure of resistance to manipulation is the worst-case computational complexity of computing whether there is some ballot that will elect a desired candidate. I will report on a recent paper in which we take a different approach, measuring resistance to manipulation by whether neural networks of varying sizes can learn to profitably manipulate a given voting method in expectation, given different types of information about how other voters will vote.