Studies how to keep powerful AI systems aligned with human intentions when people cannot easily judge their work. Its research tests model debate as a tool for oversight, ways to train reliable behavior, and risks that emerge when multiple AI agents interact.
See something inaccurate or outdated?
