Degradation only shows when takeover is needed
Aliases: latent deskilling · discovered at takeover · no alarm on the way down
What it is
Day to day the automation works, the person’s scorecard is green, degradation has no display. Until the machine cannot, and the person must put hands on, the scorecard turns red for the first time. Skill loss shows only at takeover is a misaligned observation window: inventory is taken on the hardest beat, the one least practised.
Fluency at the end of training, and rust on the first fault three years later, have no alarm in between.
Why it happens
The success path does not ask the person to do; the failure path does — and it is the machine’s failure sample, harder than the average task. Measurement therefore happens in the tail of the distribution; practice happened in the middle that automation deleted. Feedback people get day to day is “the system is steady,” not “how well can I still do this.” Complacency makes people look less; degradation makes them unable even if they look. The two often stack on the takeover beat: not looking, and not able.
A product that only calls the person on failure has scheduled the skill physical at the moment the patient has already fallen.
Studying it
After long automation, compare two tests: everyday supervisory scores (spot a fault, click confirm) versus sudden-takeover manual scores (complete an anomalous item oneself). Independent variables: time since last manual work, whether takeover comes with startle, whether the anomaly sits in the training distribution. Dependent variables: the gap between the two scores, whether people are surprised they cannot, how long error lasts after takeover.
The gap is the quantity. Supervisory scores good, manual scores bad, is exposure postponed.
Where it stops holding
A shift that still has a mandatory manual stretch every day will not postpone exposure until a fault. A purely advisory system has the person doing the work anyway; degradation is slow and exposure is early. Whether skill is dropping is the previous claim; this one only treats when it becomes visible. Whether to keep a manual path is the next prescription.
Applying it
- Do not treat everyday confirm rates and fault-detection rates as proof of skill. Schedule undeclared manual spot-checks on anomalous items, not on the easy items the machine can do.
- Run takeover drills on failure samples, and record the gap between “thought I could” and “actually could.” Treat the gap as a quality signal, not as someone’s shame.
- Check: in a fault-free month, do one sudden manual spot-check. If those scores are substantially worse than end of training, and everyday gauges are all green, exposure is already postponed. Waiting for a real fault measures a tail stacked with startle, and cannot be a baseline.