An alignment safety case sketch based on debate

Debate as a scalable oversight solution: The paper proposes using debate between AI systems as a method to evaluate superhuman AI when humans can no longer directly judge outputs. This addresses a critical challenge in AI safety: as systems become more capable than humans, traditional human feedback becomes insufficient for steering them toward safe behavior.

Four foundational assumptions required for safety: The authors identify that debate-based safety depends on four specific claims: (1) the AI system must become skilled at the debate game itself, (2) winning at debate must reliably correlate with honest behavior, (3) honesty must persist during actual deployment through ongoing training, and (4) the deployment environment must tolerate some level of errors. Each represents a potential vulnerability if not satisfied.

Concrete application to AI R&D agents: Rather than abstract theorizing, the paper grounds debate safety in a specific, high-stakes scenario: an AI agent working on research inside an AI company that could sabotage results. This example illustrates why the approach matters—preventing deceptive behavior in systems that could cause "egregious harm" through their actions.

Significant open research questions remain: The authors acknowledge this is a "sketch" rather than a complete solution, explicitly identifying research gaps that must be solved before debate can be considered a compelling safety argument. This honest assessment highlights that debate is a promising direction but not yet a proven safety guarantee.

Online training during deployment is essential: The paper proposes that honesty cannot be ensured through training alone; systems must continue learning during actual deployment to maintain safe behavior, suggesting that AI safety requires ongoing monitoring and adjustment rather than one-time solutions.