Evhub proposes monitorability red lines for AI labs
The post sorts monitorability into four types and argues that trusted monitors are the main open problem.
AGI & Safety#06
4 types
Types of monitorability the post categorizes
Image: AI Alignment Forum
4 typesTypes of monitorability the post categorizes
The short version
- The post categorizes four types of monitorability and discusses how trustedness affects them.
- It notes that a monitor using the same model as the policy could in principle get a perfect score, so trusted monitors are needed.
- It proposes evaluating the ratio of unmonitorability to capabilities, which it calls the monitorability externality, and rejecting model changes that push it too high.
- It proposes Control and Training Monitorability red lines, with Auditing Monitorability implicit in how monitors are judged trusted.
Why it matters
Monitoring is one of the tools proposed for overseeing advanced models, so red lines on it would limit how labs build and change models. The post argues that absolute red lines alone do not constrain actions today, so it pairs them with limits on changes that degrade monitorability. For developers, it offers a concrete way to think about how oversight holds up as capabilities grow.
“a monitorability externality would be a change that makes monitorability even worse than would otherwise have previously been predicted for a given level of capabilities”