Scalable oversight is the problem of how people can reliably supervise AI systems that may become better than them at the task being checked.
In plain words
Today people can often judge whether a model's answer is right. As models get stronger, that gets harder: a human may not spot a subtle error in expert-level code or science. Scalable oversight research looks for ways to keep supervision working, such as using AI assistants to help humans check, or training strong models with labels from weaker ones.
A quick example
Researchers let a small, weak model label data and use those labels to train a much stronger model. The strong model does better than its weak teacher, which gives a small-scale test of whether weaker supervisors, like humans, could guide smarter systems.
Where you'll meet it
- Alignment research at frontier labs.
- Designing evaluations for tasks humans find hard to check.
Related
Sources
- PaperMeasuring Progress on Scalable Oversight for Large Language Models, Bowman et al. (arXiv) · read 29 Sept 2026
- PaperWeak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision, Burns et al. (arXiv) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.