Societal Impacts of AI: Challenges and Opportunities

The rapid adoption of data-driven technologies has outpaced our foundational understanding, leaving critical questions unanswered; yet, these technologies are deeply embedded in our society, driving high-stakes decisions in policy, science, business, and beyond. To ensure these systems serve society effectively, we must confront key questions: What challenges and opportunities arise in these decision-making contexts? How can we mitigate the challenges and unlock the opportunities?

In this talk, I will present my work addressing these questions. In the first part, I will discuss prediction-powered and active inference??"principles for drawing valid conclusions based on AI outputs with provable guarantees and efficiency in cost, data, and computation. I will highlight the potential benefits of these methods in the context of an ongoing collaboration that enables the use of large language models for studying social science questions. In the second part, I will discuss performative prediction??"a conceptual framework that formalizes the impacts that algorithms can have on society and the natural equilibria they create. Performativity is a well-studied phenomenon in economics and social sciences, but it has largely been absent from machine learning discourse. I will share key insights enabled by the framework and principles for achieving desirable equilibria between algorithms and society.

Speaker: Tijana Zrnic, UC Berkeley