The problem of supervising a system that is better than you at the task. If a model writes ten thousand lines of code or a hundred-page analysis, checking it properly costs more than doing it yourself, and that gap widens as capability grows. Techniques here try to give a weaker supervisor leverage.
Technical view
Approaches include decomposing a task into checkable parts, using models to assist evaluation, adversarial arrangements such as debate, recursive reward modelling where models help train the next stage’s evaluators, and prover-verifier framings where generating a verifiable argument is required. The core assumption is that verification is easier than generation, which holds for many tasks and not all. It is the practical bottleneck on aligning systems past human expert level.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.