Frontier stage
Keynote

Generating robot constitutions for aligned behavior

In person
  • Date
    10 July 2025
    Timeframe
    15:00 - 15:20 CEST
    Duration
    20 minutes
    Share this session

    As robots are becoming increasingly integrated into society, ensuring their safe and ethical behavior is paramount. We explore the use of “robot constitutions”, sets of rules designed to govern robot actions, generated directly from real-world images, hospital injury reports and science fiction. We demonstrate how these constitutions can be automatically tailored to specific environments, and even optimized through counterfactual reasoning. By evaluating robot responses against generated “undesirable” scenarios, we can assess and improve constitution effectiveness. Notably, AI equipped with these constitutions shows significantly higher alignment with human values than in their absence. Reassuringly, we find that alignment of modern AI with humans is very high (96%) compared to AIs and robots depicted in science fiction (21%).

    Share this session
    • 20
      Days
      17
      Hours
      08
      Min
      57
      Sec