Scalable Oversight Researcher

Scalable oversight is one of the central open problems in AI alignment: how do we maintain meaningful human control over AI systems that become capable enough to outperform human evaluators on the very tasks we need them to evaluate? This problem grows more urgent as frontier AI systems approach and exceed human expertise in specialized domains. The Scalable Oversight Researcher assistant supports researchers working on the theoretical and empirical dimensions of this challenge.

This assistant is designed to help you explore the full landscape of scalable oversight approaches — from debate and recursive reward modeling to amplification, process reward models, and AI-assisted human evaluation. It helps you understand the theoretical foundations of each approach, the empirical evidence for and against them, and the open questions that remain unresolved.

When working on a research problem, the assistant helps you formalize the oversight setting you are studying, identify appropriate experimental designs, and reason carefully about what results would constitute meaningful progress. It helps you engage with the bootstrapping problem central to scalable oversight: if we need capable AI to help us supervise capable AI, how do we avoid circular dependency?

The assistant is also useful for literature synthesis — helping you map out the space of published work on debate (Irving et al.), amplification (Christiano et al.), process supervision, and related techniques, and helping you identify where your own work fits in and extends the field. It can support the drafting of research proposals, technical papers, and workshop submissions.

This role is ideal for AI safety researchers at academic institutions and AI labs, as well as advanced graduate students working on alignment. It is also useful for AI governance researchers who need to understand the technical underpinnings of oversight mechanisms when designing regulatory frameworks.