Six Levels of Delegation

Taxonomies of automation ask how autonomous a system is. For general-purpose language models that is the wrong thing to measure. This paper measures something else: where the operator's understanding of the work is last refreshed. Somewhere on that scale is a boundary no autonomy scale marks. On one side of it an operator who missed something can find it by looking harder. On the other side looking harder does nothing.

An operator who reads the work is updated by the work; an operator who reads a test result is updated by an instrument. Between these two positions their knowledge changes in kind, not degree: it becomes bounded by what somebody thought to measure, in advance. Below the boundary, better oversight means paying more attention; above it, attention cannot help, because the missing information was never in the outputs at all. Kalman's notion of observability names this exactly, and it gathers a scattered family of results in the automation literature — surprises, complacency, skill decay — into consequences of one position rather than separate phenomena.

We locate the boundary within a six-level ladder derived from what can actually be handed over in knowledge work — the syntax, the thinking, the writing, the checking, the running, and the choosing. Each level is identified by the one decision the operator keeps, and that decision is also the last place their understanding of the work gets refreshed. The boundary falls between “reviewer” and “approver”, which is why they are separate rungs here and one rung almost everywhere else.

Two consequences follow. First, the level belongs to the task, not the system: an operator may cross the boundary several times an hour, and the drift upward is easy to perform and easy not to notice. Second, most current AI products fix the operator at one level by their shape — a chat window affords “collaborator”, inline completion affords “reviewer”, an autonomous agent affords “audience” — so changing level means changing application and losing context at the boundary. We give a case study of a command-line tool that spans all six levels within one binary by delegating composition to the Unix shell, as evidence that interruptibility makes the position recoverable rather than merely enumerable. The model is intended to be falsifiable; we state the experiments that would test it.