Wiki · Reference entry
Oversight
Humans checking what an AI system is doing.
The default answer to every safety question, and it works while a person can keep up. Scalable oversight is the research problem of what to do when they cannot, either because the volume is too high or because the reasoning is beyond the reviewer. Note that both failures arrive quietly: an overwhelmed reviewer still signs things.
Why it matters
Every control you currently trust probably assumes a reviewer who can keep up.