Skip to content
finalstate.aiDeskilling: The Skill That Catches the Error Is the One Automation Quietly Takes
126Atlas

126THE MONITOR'S SEAT

Deskilling: The Skill That Catches the Error Is the One Automation Quietly Takes

Deskilling is a loss of unaided capability. Automation can remove opportunities to practise while leaving difficult monitoring and recovery tasks to people. Whether that capability has changed needs evidence.

An automated console develops an unusual signal while a human operator remains responsible for takeover.In an illustrative control room, routine adjustments move to automatic control. An abnormal signal then appears and a response is required from the operator. The drawing leaves the person's capability unresolved; it shows no measured readiness gauge or inevitable skill loss.AutoOperator?Response
Illustrative control room. Successful routine operation does not measure the operator's readiness for an unusual event.

THE PRACTICE GAP

Lost practice leaves a question about capability.

Routine cases bypass the operator, while an unaided trial opens a separate question about retained capability.Case markers move along the automated route and the ordinary manual-practice route is interrupted. A separate trial connects a case to the operator and leaves its response unscored. No curve, percentage, or time axis claims a rate of skill loss.RoutineAutoOperatorUnaided?Trial
  • Routine automation can remove occasions to act and receive feedback
  • An unaided trial supplies evidence about a specific retained capability

Reduced practice may affect manual control, retrieval of knowledge, or diagnosis. It does not prescribe one decay curve for every person or task. Metacognitive demand is the burden of evaluating the output; capacity is the ability to meet that burden. Demand can persist while capacity changes. Neither workload nor a quiet operational record measures that change by itself.

THE IRONIES OF AUTOMATION

Automating routine control can leave demanding residual tasks.

Routine control is assigned to an automatic controller; an abnormal condition is routed to a human operator.The fixed allocation diagram represents the classic arrangement Bainbridge examines. Its separate routes show automated routine control and human diagnosis or recovery without a declining skill bar or a claim that every automated system uses this allocation.RoutineAutoRecoveryOperator
A task allocation analyzed by Bainbridge, not a measured distribution of human and machine work.Lisanne Bainbridge, Automatica 19(6), 1983, pp. 775–779
  • Monitoring requires recognizing abnormal behavior
  • Recovery can require diagnosis and manual control

In Ironies of Automation90046-8), pp. 775–777, Lisanne Bainbridge analyzes industrial process control, with flight-deck examples. Designers can leave people the tasks they cannot automate while ordinary operation supplies less practice. She distinguishes manual skills, knowledge, awareness of the current process, and monitoring. The paper describes design problems and possible responses, not a universal numerical law of decay.

A rare takeover can still require practiced judgment.

The system's reliable days do not demonstrate what the operator could do during a failure. The reverse-centaur case raises a practical question: which recovery task remains human, and what evidence supports assigning it to this person?

SIX POINTS, IN EXPERTS

Unaided detection changed after AI entered the workflow.

Fixed bars show unaided ADR of 28.4 percent before AI introduction and 22.4 percent afterward on a common zero-based scale.The bars represent 226 of 795 procedures before and 145 of 648 afterward. The confidence interval in the caption belongs to the after-minus-before difference, not to either bar. The before/after design observes performance and does not identify a causal loss of skill.Unaided ADR0%20%40%28.4%Before226/79522.4%After145/648
ADR here is the share of procedures detecting at least one adenoma or cancer; sessile serrated lesions are excluded. Before: 226/795 (28.4%). After: 145/648 (22.4%). After minus before: −6.0 percentage points, 95% CI −10.5 to −1.6. Both groups were examined without AI. Association, not proof of AI-caused deskilling.Budzyń et al., Lancet Gastroenterology & Hepatology 10 (2025), 896–903; correction, November 2025

Budzyń and colleagues compared approximately three-month windows around AI introduction, using procedures from September 2021 to March 2022. Their accepted manuscript, Methods pp. 9–11, excludes the first 100 procedures per centre after introduction. The historical comparison was not randomized, although the parent study randomized AI-on/off days. Patient mix, time, and changing diagnostic volume could affect the result; withdrawal-time data were absent. A 2025 correction00294-8) adds an omitted indication covariate to a supplementary adjusted analysis. This exhibit uses the raw comparison.

  • Four Polish centres; 19 experienced endoscopists
  • 1,443 non-AI colonoscopies in a retrospective before/after comparison

COMPLACENCY AND ATTENTION

Automation complacency is found in both naive and expert participants and cannot be overcome with simple practice.

Parasuraman & Manzey, Human Factors 52(3), 2010, abstract

The review's finding concerns automation complacency when manual tasks compete with an automated task for attention. That is distinct from losing an unaided skill through disuse. The quotation does not establish irreversible deskilling or show that every form of training fails. A proposed trial must specify whether it tests attention, diagnosis, manual performance, or some combination.

TRIALS AND FEEDBACK

Record a response before showing the answer.

  1. 01Use a controlled simulation with a known fault and an unaided response
  2. 02Score the response against a criterion set in advance
  3. 03Review the case and response together, with access to the original evidence

For the illustrative control-room team, these are proposed ways to gather evidence. They are not validated cures or a proven drill schedule. Use comparable cases to examine assisted and unaided performance; learning, task difficulty, and attention can affect the comparison. A reviewer with a different routine is not automatically independent. Where independent assurance is needed, scope, access, reporting, resources, and conflicts still require attention.

An unaided response to a simulated fault is recorded, scored against the case, and made available to a reviewer.The response record starts incomplete. After the operator identifies the planted fault, corresponding case and response marks are compared. A reviewer then receives both records. This is an illustrative successful trial, not an empirical pass rate or a claim of formal reviewer independence.CaseResponseReviewer

WHAT APPROVAL DOES NOT SHOW

An approval record can coexist with a missed fault.

Two simulated trials contain the same fault; one detects and holds it, while the other approves it despite missing it.Both case records contain the same abnormal signal. The first response identifies the fault and stops acceptance. The second response treats it as normal, then receives an approval stamp while the fault remains visible. The two alternatives demonstrate why approval alone does not establish detection; no trajectory, frequency, or cause of failure is inferred.DetectedHoldMissedApproved
  • One trial records detection and a hold
  • Another records approval despite the same planted fault

These are illustrative alternatives. A miss calls for investigation; it does not reveal whether the cause was lost skill, diverted attention, a poor interface, or something else. The demand of checking has not disappeared merely because an approval was logged. This human-monitoring question also differs from the generation-verification gap, a model-level failure to select a correct generated candidate.

WHO STILL KNOWS HOW

Ask for the task and the observed result.

A missed unaided trial is recorded beside the responsible operator; the next takeover assignment is held for reassessment.The record connects the operator to a specific simulated fault and a response that missed it. The mismatch leads to a hold on the proposed unaided takeover assignment. The date symbol denotes a required record field, not an invented dated assessment. This local decision is illustrative, not a universal competency threshold.OperatorDateTrialTakeoverHold
  • Person: who is expected to diagnose or recover this failure?
  • Task and date: what comparable unaided case did they attempt, and when?
  • Result: what did they detect, miss, or need help to resolve?
  • Decision: what changes before the next takeover assignment?

In the missed-trial example, the team records the fault and response, then holds the proposed unaided takeover assignment for reassessment. A name and date without a result cannot support that decision. One successful trial would support only the capability it tested. Return to the reverse-centaur case with an observed performance record and a reasoned plan for the remaining verification.

Read the transcript

01 · THE MONITOR'S SEAT

An operator watches a control-room console. In this illustrative scene, software takes over the routine adjustments. The person remains responsible for unusual events. Then an unfamiliar signal appears. Someone must diagnose it and decide whether to take over. Routine operation has supplied fewer occasions to practise those actions, but that does not tell us whether this operator can still perform them. Their readiness needs evidence of its own.

02 · THE PRACTICE GAP

Deskilling means a loss of unaided capability. Routine cases now follow the automated route, bypassing opportunities for the operator to act and receive feedback. Reduced practice may affect manual control, retrieval of knowledge, or diagnosis. It does not imply one decay curve for everyone. Open a separate unaided trial to find out what the person can still do. That response has yet to be scored. Keep demand and capacity distinct: the burden of evaluating an output can persist even if the ability to meet it changes. Workload alone does not measure lost skill.

03 · THE IRONIES OF AUTOMATION

Lisanne Bainbridge described the design problem in Ironies of Automation in nineteen eighty-three. She focused on industrial process control, with examples from flight decks. Designers can automate routine control while leaving people to monitor the system and handle events they could not automate. Manual recovery, diagnosis, and knowing the current process state make different demands. Ordinary operation may provide too little practice for the remaining role. Bainbridge also considered training, displays, and human-computer collaboration. Her paper is an analysis of design problems and possible responses, not a measured law of skill decay.

04 · THE FINAL IRONY

A rare takeover can still require practiced judgment. The system may need human intervention infrequently, then require a difficult diagnosis under time pressure. Its reliable days do not show what the operator could do in that situation. Name the recovery task that remains human, then ask what evidence supports assigning it to this person.

05 · SIX POINTS, IN EXPERTS

A study published in twenty twenty-five examined unaided performance in colonoscopy. At four Polish centres, nineteen experienced endoscopists performed one thousand four hundred and forty-three procedures without AI. The comparison used roughly three-month periods around AI introduction, with the first hundred procedures at each centre excluded after introduction. The procedures took place in twenty twenty-one and twenty twenty-two. The share detecting at least one adenoma or cancer fell from twenty-eight point four percent to twenty-two point four. Sessile serrated lesions were excluded from that measure. The difference was minus six percentage points, with a ninety-five percent confidence interval from minus ten point five to minus one point six. This historical before-and-after comparison was not randomized. Patient mix, time, and changing procedure volume could contribute. A later correction changed a supplementary adjusted analysis; these are the raw results. They show a change in observed unaided performance, without establishing that AI caused skill loss.

06 · COMPLACENCY AND ATTENTION

Automation complacency presents a related but different problem. Parasuraman and Manzey's twenty-ten review describes it when manual tasks compete with the automated task for attention. Their statement that it cannot be overcome with simple practice concerns complacency under those conditions, in both naive and expert participants. It does not establish irreversible deskilling or show that every kind of training fails. A proposed trial needs to say what it is testing: attention, diagnosis, manual performance, or a combination.

07 · Advertisement · Bubble AI App Builder

Some ideas do not need another document before they become testable. With Bubble AI, you describe the app you want, and Bubble creates a working starting point: interface, data, and logic you can inspect. From there, you refine visually, connect AI models and services, and turn the first version into something real enough to use.

08 · TRIALS AND FEEDBACK

For the control-room team, use a simulation with a known fault. The operator attempts the case without the aid and records a response before seeing the answer. Compare that response with the criterion set in advance. In this example, it identifies the planted fault. A reviewer then receives both the case and the response. These are proposed ways to gather evidence, not validated cures or a proven drill schedule. Comparable assisted and unaided trials can help examine performance, but learning and task difficulty can affect the comparison. A different reviewer routine does not establish independence; an assurance role still needs appropriate scope, access, reporting, resources, and safeguards against conflicts.

09 · WHAT APPROVAL DOES NOT SHOW

Consider two illustrative responses to the same planted fault. In the first trial, the operator detects the abnormal signal and holds acceptance. In the second trial, the response treats that signal as normal. Approval is recorded anyway, while the fault remains. The approval record alone cannot establish detection. A miss needs investigation: diverted attention, an unsuitable interface, lost skill, or other conditions could explain it. The demand of checking has not disappeared. And this human-monitoring problem is distinct from the model-level generation-verification gap, which measures failure to select a correct generated candidate.

10 · WHO STILL KNOWS HOW

Keep a record that connects a person to a task, a date, and an observed result. In the missed-trial example, record the simulated fault and the operator's response. The team compares them and records the miss. It holds the proposed unaided takeover assignment for reassessment. The record should explain what needs further practice, support, or testing before that assignment changes. A name and date without a result cannot support the decision. One successful trial would support only the capability it tested. Bring that evidence back to the role you expect the person to carry.

01 / 10 · THE MONITOR'S SEAT0:00 / 6:48