Creating Resilience in AI

David Bau
Northeastern University

White paper thumbnail
NDIF Resilience White Paper

How Will Society Withstand the AI Revolution?

Large-scale artificial intelligence has begun to confront humanity with an historic challenge: how to deal with a technology that is designed to surpass the boundaries of human understanding and control? The solution lies in resilience: As a society, we must develop the ability to sustainably adapt to unexpected challenges in AI.

The Challenge of Large-Scale AI

Large-scale AI poses fundamental challenges that are poised to reshape society:

  • Generative models that are trained to fool human perception and shape human behavior.
  • Reasoning models that are trained to think in novel ways unseen in recorded human history.

The gravity of these challenges demands solutions that do not lie in the technology alone, but in a societal investment in the science, engineering, and culture of resilience around AI.

Resilience and Interpretability

To develop a resilient AI ecosystem, we must invest in three things:

  1. The science of understanding AI so that humans can know why AI goes wrong when it does and what should be fixed.
  2. The technical practice of control of AI so humans have the ability to fix the things that need fixing.
  3. A culture of power over AI so that humans have the right, responsibility, and expectation of controlling AI.

These are all goals of the field of AI interpretability, which is the focus of my lab.