Scalable oversight is a research agenda in AI safety concerned with how humans can reliably supervise AI systems whose capabilities, in some domains, may exceed those of the supervisors. The core problem is that as models become better at complex reasoning, writing, or code generation, human evaluators struggle to tell whether a given output is actually correct or merely plausible-sounding — which undermines the reward signal that training techniques like RLHF rely on. Proposed solutions include debate (where two AI agents argue opposing positions for a human judge to evaluate), recursive reward modelling, and using a weaker AI to assist the human reviewer.

Think of it like a hospital quality-review board that must certify surgical procedures it lacks the expertise to perform itself. Rather than abandoning oversight, the board develops protocols — second opinions, checklists, outcome tracking — that let non-surgeons still catch systemic problems.

# Debate-style oversight sketch
Claim: "Protein X causes disease Y"
Agent A argues FOR the claim, citing evidence
Agent B argues AGAINST, pointing out confounds
Human judge evaluates quality of arguments
  → rewards whichever agent reasoned more honestly

Scalable oversight is tightly connected to RLHF, Chain of Thought, and Hallucination — better reasoning transparency is a prerequisite for meaningful human review. It also shapes how Foundation Model developers think about evaluation at increasing capability levels.

You will encounter scalable oversight in AI safety research papers, in policy discussions about frontier model governance, and increasingly in technical blogs from labs like Anthropic and DeepMind as they scale models past human expert ability in narrow domains.