Does automation actually remove humans from the control loop?

Lisanne Bainbridge's 1983 paper "Ironies of Automation" argues that automating an industrial process to remove an operator seen as "unreliable and inefficient" does not remove the human from the system — it leaves the operator "an arbitrary collection of tasks" that the designer could not figure out how to automate, with "little thought... given to providing support for them." The leftover tasks are the hardest ones: monitoring that the automatic system is working, and taking over manual control or diagnosing a fault when it is not. Bainbridge names this a double irony: "designer errors can be a major source of operating problems," and the attempt to eliminate the operator "still leaves the operator to do the tasks which the designer cannot think how to automate."

The mechanism she gives is skill decay through disuse. Manual control skills "deteriorate when they are not used, particularly the refinements of gain and timing," so "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one" who, on takeover, "may set the process into oscillation." Cognitive skills decay the same way: long-term knowledge of the process depends on "frequency of use" and fades like unused classroom learning, and the "working storage" an operator needs for on-line decisions — not raw data but built-up predictions about the process — "takes time to build up" and is absent without active control. Because takeover is needed precisely when something is already wrong, Bainbridge argues the operator "needs to be more rather than less skilled" at the moment automation has made this least likely, and generalizes the point via Bibby and colleagues' 1975 paradox that the more advanced a control system is, "the more crucial may be the contribution of the human operator."

The vault holds the same mechanism argued for AI agents four decades later. Does granting agents more autonomy undermine human oversight? makes nearly Bainbridge's claim about agent systems, that oversight capacities are "undermined by the act of using these systems." Bainbridge is the earlier, narrower case that later claim generalizes from: here the erosion is named concretely as decay of manual control timing and long-term process knowledge in industrial operators, not AI overseers. Can organizations lose scrutiny capacity while keeping oversight forms? describes the same capacity loss at an organizational level without naming a mechanism; Bainbridge supplies one of the oldest mechanisms on record, loss of practice. Does AI augmentation protect workers from skill erosion? states the modern version of the warning Bainbridge raises about "machine-minding operators."

The excerpt is a 1983 journal article about industrial process control — power plants, batch chemical processes, aircraft — and says nothing about software agents, language models or knowledge work; its cited studies (Edwards and Lees 1974 on step changes, Bainbridge 1975 and 1981 on operator knowledge) concern manual and supervisory control skill, not oversight of autonomous decision-making, and no figure quantifies how much skill decays or over what period beyond a general claim about disuse. The defensible implication is scope-limited but durable: a system designed to remove a human because the human is the unreliable part should expect the human to remain, doing a residual task that is harder and less supported than the one removed, and the system itself should be examined for whether it is removing the practice that made the human capable of that residual task.