When autopilot handles the routine stuff, do the humans watching it quietly lose the skill to rescue things when it breaks?

This explores whether people who supervise automated systems lose the skills they'd need to take over when something goes wrong. The corpus's clearest answer comes from a meta-analysis of 18 experiments: automation reliably improves routine performance but makes failure recovery and situational awareness worse (When does automation harm performance after failures?). The finding that's easy to miss is where the damage concentrates. Automation that helps you analyze information does little harm. The trouble starts when automation moves into choosing actions. Once the system decides what to do and the operator only watches, the operator stops practicing the judgment an emergency demands, and also loses track of what the system is doing and why. So in a crisis they are rusty and disoriented at the same moment.

A second line of research suggests the skill may never have been there to begin with. Studies of AI-assisted workers find that people produce expert-looking output while the AI is available, then drop back to their baseline once it's taken away (Does AI assistance build lasting skills or temporary abilities?). The authors' metaphor is an exoskeleton: it adds strength while you wear it and builds no muscle. For emergencies this matters a great deal. An organization can look at strong day-to-day results and assume its people are capable, when the capability actually belongs to the tool. The exact moment the tool fails is the moment that hidden gap shows up.

There's a useful parallel in how AI agents use skills. An analysis of more than 8,000 trials found that skills mostly work as procedural anchors that keep action steady, not as a way to supply missing facts. They break down when they're applied out of context or followed too rigidly (Do skills teach procedures or inject missing facts?). Human emergency procedures behave the same way. A checklist steadies an operator who already understands the system, but it can't replace that understanding when the situation doesn't match the script, which is usually the case in a real emergency.

Two AI-side findings help explain how skills fade in the first place. Agent memories that are continually condensed into summaries eventually do worse than raw recall. In one case the model failed 54% of problems it had previously solved, because the summarizing removed the details about when each lesson applied (Does agent memory degrade when continuously consolidated?). That resembles an operator who remembers the general rule but has forgotten the conditions that trigger it. Research on sandbagging adds a warning about measurement: a capability can be present but suppressed, so tests miss it (Do sandbagged models actually lose their capabilities?). The human situation is the reverse risk. Routine evaluations done with the automation switched on can show competence that isn't actually there.

The gap: none of these sources directly studies human skill decay over months or years, or training that counters it, such as manual-flying requirements for pilots. The corpus establishes the mechanism (watching automation erodes both skill and awareness, worst when the automation picks the actions) and the masking effect (assisted performance hides the decay). It doesn't yet cover the remedies.