Most of what is known about humans supervising machines was learned in cockpits, control rooms and nuclear plants, and the AI conversation borrows it wholesale without asking which parts survive the move. Some of it transfers exactly. Some of it does not, and the difference is not decoration.
The answer, in one line
Four things. The irony of automation itself, where the easy parts are automated and the hard residue is left to a person with decaying practice.
What transfers, and why#
The irony itself. Automating the easy parts and leaving the hard residue to a human whose practice is decaying is a structural result, not an artefact of aviation. It holds anywhere a person is left to catch what the machine cannot.
Complacency and bias. Parasuraman and Manzey's review describes an attentional mechanism driven by learned trust. Nothing in it depends on the system being deterministic, and the effect appears in experts.
The out-of-the-loop problem. A supervisor who has not been doing the task loses the picture needed to take it back. Endsley's levels of situation awareness are as applicable to an analyst reading a generated summary as to a pilot.
The stage model. Splitting a workflow into acquisition, analysis, selection and action is a design vocabulary that works anywhere. See levels of automation.
What does not transfer cleanly#
A stable level of automation. A control system sits where it was designed to sit. A language model moves between analysis and decision selection depending on the prompt, the reader and the day, so the level is a property of the interaction rather than of the system.
Error behaviour you can learn. An autopilot fails in characteristic ways an operator eventually recognises. A model's errors are fluent, various and often most convincing in the places it is most wrong, which is a different detection problem and a harder one.
The bounded task. Control-room work has a defined envelope, alarms and clear outcomes. Knowledge work usually has none of the three, and the measures built for the first, including SAGAT for situation awareness, have no equivalent in the second.
Training and licensing. The aviation answer to automation dependency is recurrent training, simulator hours and a licence to lose. Nothing in the knowledge economy has the same enforcement, so the aviation comparison is useful as an argument and weak as a plan.
Why this matters for a policy#
Because a borrowed frame comes with borrowed confidence. When a board is told that human oversight is a solved problem in safety-critical industries, the claim is half true and the missing half is the one that decides whether the policy works: those industries hold the level fixed, define the envelope, measure the operator and fund the practice. An organisation that adopts the vocabulary and none of the four has bought the reassurance without the mechanism.
The position this research takes is that the human-factors literature gives this territory its best mechanisms and none of its measurements. What follows from that is on the models of judgement.
Key sources
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6).
- Endsley, M. R. (1995). Toward a theory of situation awareness in dynamic systems. Human Factors, 37(1).
- Parasuraman, R. and Manzey, D. H. (2010). Complacency and bias in human use of automation. Human Factors, 52(3).
- Rasmussen, J. (1983). Skills, rules and knowledge. IEEE Transactions on Systems, Man and Cybernetics, 13(3).
Essay · SS-2026-316
Hirji, R. (2026). What the control room literature does not transfer. The SuperSkills evidence base, SS-2026-316. https://thesuperskills.com/research/what-the-control-room-literature-does-not-transfer. Last reviewed 26 September 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work