Seminar Artificial Intelligence - Dr. M. Revel, Harvard Law School

Title: Humans and AI Governance

Abstract:

Governance (the art of accommodating a plurality of potentially irreconcilable views) is a daunting challenge in human cooperation. Its intricacies have now extended into the realm of AI alignment. This talk explores humans and AI governance, delving into the theoretical foundations of voting theory and a real-world application: Reinforcement Learning from Human Feedback (RLHF).

We will first navigate the insights computational social choice offers to craft legitimate decision-making frameworks capable of aggregating human preferences and expertise. We will delve into the mathematical underpinnings of liquid democracy (a delegation-based system deemed fair and efficient) and surface theoretical and empirical methods used to study quantitatively the concept of governance.

Our focus will then shift towards the intricate interplay between RLHF and the alignment of AI models with human values. Through an examination of the AI alignment pipeline, we will discuss a novel approach aimed at constructing Minimal Viable Alignment Datasets (alignment datasets whose entries are optimized to be learned by a reward model). We will see how such approaches may foster explainable and responsible AI developments.

In closing, I will share my vision for the AI and Democracy project, wherein researching foundational principles of AI alignment serves as the cornerstone to advance AI-driven decision-making within democratic processes. I intend to build my scholarship on these two pillars towards seamlessly intertwining AI and human governance, ushering in a new era of augmented, responsible, and explainable decision-making.