CHAI (Center for Human-Compatible AI)
CHAI (Center for Human-Compatible Artificial Intelligence) is an academic research center at UC Berkeley founded by Stuart Russell. CHAI focuses on developing AI systems that are provably beneficial to humans.
Overview
CHAI takes an academic approach to AI alignment, combining theoretical foundations with practical research. The center is known for work on Cooperative Inverse Reinforcement Learning (CIRL) and assistance games.
Research Approach
Human-Compatible AI
Rather than specifying goals for AI systems, CHAI advocates for systems that are uncertain about human preferences and actively work to learn them. This approach is designed to produce systems that defer to humans and remain corrigible.
Assistance Games
CHAI developed the assistance game framework, where AI systems are modeled as assistants trying to help humans achieve their (initially unknown) goals. This builds on CIRL.
Key Contributions
- Cooperative Inverse Reinforcement Learning
- Assistance games and off-switch games
- Stuart Russell's book "Human Compatible"
- Research on value alignment and preference learning