Scalable Oversight

Techniques that let humans reliably supervise AI systems that may be more capable than they are, e.g. AI-assisted evaluation or debate.