Getting AI Systems (Collectively) to Behave the Way We (Collectively) Want Them To

It is often hard to get an AI system to do what someone really wants it to do—a problem referred to as "aligning" the AI. But even if we can solve that problem reasonably well, other problems remain. One is the problem of accounting for multiple stakeholders: we generally want to take into account the preferences, values, or judgments of more than one party, even if they conflict with each other. Another is the problem of multiple interacting AI systems: even if each of these systems individually is, on its own, reasonably well-aligned, interactions among them can produce disastrous outcomes. I will give an overview of these problems, and will argue that we can address them using social choice theory and game theory. (No previous background required in any of this.)