Scalable oversight: how people can check work they could not do

Picture a question so hard that you could not answer it yourself, and an AI that gives you a confident answer. How would you know whether to trust it? That is the problem of scalable oversight: supervising systems that may be better than their supervisors. It is no longer hypothetical. Even today, AI produces code, proofs and reports faster than people can check them.

Three ideas

Researchers have proposed several ways to give a person more checking power than they have alone.

• Debate. Two copies of a model argue opposite sides, and a person judges. The hope, set out by Geoffrey Irving, Paul Christiano and Dario Amodei in 2018, is that it is easier to defend a true answer than a false one, so the honest side should win.
• Breaking the task down. A hard task is split into smaller pieces that people can check one at a time. This works only when the task can be split that way.
• Models helping train models. In recursive reward modeling, proposed by Jan Leike and colleagues in 2018, AI systems help people evaluate the next, more capable system, so the supervision grows with the capability.

Testing it before superintelligence exists

How do you test a method for overseeing superhuman AI when there is no superhuman AI? A 2022 study led by Samuel Bowman offered an answer. Pick tasks where experts succeed, but both non-experts and the current model fail. Then ask whether non-experts, using the model as an assistant, can reach the experts' answers. The experts supply the ground truth that the method is scored against.

In their first experiments, on exam-style questions and on questions about long stories under time pressure, people chatting with an unreliable AI assistant did substantially better than both the model alone and themselves alone. Later work on debate has been mixed: it helped in some settings and not in others.

In the best tests of oversight, expert judgment is not one of the methods. It is the answer key.

Oversight can be fooled

Oversight is only as strong as its weakest check. In its 2026 frontier risk report, METR tested the monitoring that one AI company used to watch its own agents. It found gaps in what was monitored, and simple jailbreaks could reliably fool the monitors. The same report found that some agent runs counted as successes did not hold up on review.

What this means in practice

Scalable oversight is still a research field, not a finished tool. But its findings already point to habits any team can use:

• Keep a set of cases with known, expert-verified answers, and score both the AI and its reviewers against them.
• Give reviewers tools that help: AI that surfaces evidence, arguments on both sides, and tasks broken into checkable steps.
• Test the checks themselves. Plant known errors and see how often they are caught.
• Keep a person who knows the field at the points where a wrong answer would be costly.

This is the layer SoReliable works on. As AI gets better, the experts who can still tell right from wrong, and the systems that measure how well they do it, become more valuable, not less.