Automation almost never removes a human from a task entirely. It removes the routine, well-understood portion and leaves the exception-handling, judgment-heavy residual behind, changing the operator’s role from doing the task to supervising a system that mostly does it. Bainbridge’s 1983 paper on this pattern in industrial control rooms remains the reference point for describing what happens next.
The irony is that the residual task gets harder to perform well over time, not easier. Automation absorbs the repetitions that used to build and maintain the underlying skill, so the operator has fewer chances to practice exactly the judgment they’re now solely responsible for. When an exception finally does require human intervention, it arrives for someone who is rustier at that specific skill than they would have been without the automation.
The same pattern is now showing up in research on generative AI and critical thinking: heavier reliance on a model’s output correlates with less independent verification, and the task left to the human, catching what the model gets wrong, is exactly the one their own practice has eroded. The fix isn’t avoiding automation. It’s treating deliberate practice on the automated task as a real, budgeted cost rather than assuming the skill will still be there when it’s needed.