Scaling human oversight of AI systems for tasks that are difficult to evaluate.