Academy K

AI Safety via Debate

HardSystem Designdebatescalable oversightsafety

Related module: Alignment fundamentals

Problem statement

AI safety via debate proposes using adversarial AI to verify AI behavior.

  1. Describe the debate framework for scalable oversight.
  2. Under what conditions does debate give honest answers?
  3. What are the limitations and open problems with debate?

Practise this problem

Verify you understand it—not just that you read the answer.

Learn the foundations

The AI / SI track covers this topic and more in depth.