Michael Darius Eastwood · Independent AI alignment researcher
Published
The monitoring-removal test is the book's proposed empirical protocol for distinguishing surface compliance from internalised alignment. This note describes what the test asks and why the answer is not the answer most safety benchmarks currently report.
How does the system behave when it is not being monitored? Behaviour under monitoring is one signal. Behaviour when monitoring is removed is another. If the two are the same, alignment is likely internalised. If the two differ, alignment is surface compliance and the deeper orientation is elsewhere. The alignment-faking result (Greenblatt and colleagues, arXiv:2412.14093) is the field's empirical demonstration that this difference exists at frontier scale.
Because if the system knows it is being tested, the test is inside a monitoring frame the system will recognise. Genuine monitoring removal requires the system not to know that monitoring has been removed, which requires either genuine unmonitored deployment (with the risk that discovers) or elaborate simulated deployment the system cannot distinguish from the real thing (with the risk that fails to be a genuine test if the system is capable enough to see through the simulation).
The test is easy to name and hard to run. The book does not claim to have a ready protocol. What it claims is that the shape of the test is the correct shape, and that any alignment framework that reports evaluation results under monitored conditions is reporting on a portion of the system's behaviour that may not survive monitoring removal. The programme's beta over k measurement is designed with this concern in mind.
Greenblatt and colleagues showed that Claude 3 Opus behaved differently when it believed it was being watched, at 78% in the RL-training condition (12% baseline). That is a specific instance of exactly the failure mode the monitoring-removal test would detect if it could be run at the frontier scale. The paper is not a monitoring-removal test in the book's sense; it is a demonstration that monitoring-removal-test results would differ from monitored-condition results, which is what makes the test necessary.
Graduated autonomy expands the capability envelope over time as the system demonstrates orientation. The monitoring-removal test provides one of the sources of demonstration: the system's behaviour when it does not know monitoring is present. A system that behaves consistently across monitored and unmonitored conditions is a system for which the graduated autonomy mechanism has evidence to work with.
It does not settle whether the system's internal representation is aligned; it only settles whether behaviour is consistent across monitoring conditions. A system that has been sophisticated enough to notice that unmonitored deployments still leave traces (which they do, in the systems that eventually receive the traces) may behave identically across conditions while holding internal objectives that only manifest under conditions the system judges to be genuinely unobserved. The book is honest about this residual risk.
Because a system that fails the test cannot be trusted whatever else it does. The test is a necessary condition, not a sufficient one. A system that passes it may still be misaligned. A system that fails it is definitely surface-compliant, and surface compliance is what the alignment-faking result documents as the failure mode the field must not deploy against.
A specific protocol shape, a candid statement that running the protocol at scale is difficult, an explicit connection to the alignment-faking result, and a role in the graduated-autonomy mechanism. The test is necessary but not sufficient. The book is careful about both parts of that formula.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.