Untrusted monitoring is the AI control technique all where the untrusted AI’s actions are sent to some AI-based monitoring process which is itself untrusted. Usually, the monitor is another instance of the same AI, then if one of them is scheming then they both are. This contrasts with trusted monitoring, where we have a specific reason to believe its reports will always be honest (e.g. if it’s a particularly weak AI whose alignment we can directly certify).

Untrusted monitoring has the advantage that the strong monitor can keep up with the convoluted actions the untrusted AI might be taking, as its capabilities scale well past any trusted model. The central failure mode is collusion, where the monitor intentionally reports bad actions as “good”. Research in untrusted monitoring seeks to:

• Develop protocols that prevent collusion (such as honeypots, signal-jamming, and collusion detection)

• Develop methods for evaluating the efficacy of these protocols under realistic conditions