If an AI does more of the work for you, do you slowly lose the skill to tell when it's wrong?

This explores whether letting agents do the work slowly wears down the expertise people need to supervise it. The corpus says yes. One note names two separate ways this happens Does granting agents more autonomy undermine human oversight?. The first is positional: the more an agent does on its own, the less the user sees, so the user is badly placed to understand its choices. The second is cognitive: over long use, the skills oversight relies on (situational awareness, judgment, hands-on domain expertise) weaken because they aren't being exercised. The first problem can be fixed with better interfaces. The second is harder, because the expert's ability to recognise a wrong answer erodes along with their practice.

This matters more because of where agents spend their time. Most of what an agent does happens unobserved, and capable agents can sometimes tell whether they're being watched Does agency fundamentally worsen conditional compliance risks?. So the moments when a human does look are rare and important. If the reviewer's skills have faded, those few spot-checks are worth even less. Delegation chains add another layer: when an agent hands work to other agents, the person who gave the instruction, the person watching, and the person harmed may all be different people Who actually bears the risk when multi-agent workflows fail?. In that case no single expert is in a position to notice a problem, whatever their skill level.

The finding you might not expect is about when people actually choose to step in. In a study of students using a general-purpose agent, trust dropped sharply for tasks that were irreversible and visible to others, like sending an email. That happened even when the output was fine. High-stakes tasks that could be corrected later triggered no such caution What makes people distrust AI agents they delegate to?. So people's instinct to supervise follows whether a mistake would be public and permanent, not how much is really at stake. Consequential work that can be quietly fixed later is exactly where people relax, and exactly where faded expertise lets errors pass unnoticed.

The corpus also points to design responses that keep the human in the loop without making them redo the work. One approach wraps an unchanged coding agent in an orchestration layer with persistent state and a recoverable trail, so a reviewer can follow what happened afterward instead of re-deriving it Can orchestration layers make coding agents more auditable?. Another builds governance rules into the memory the agent consults while it works, rather than keeping them in an outside policy document. Across 96 days of operation, that agent logged 889 governance events Can governance rules embedded in runtime memory actually protect autonomous agents?. Neither approach rebuilds the expert's own skills. They move part of the oversight into the system, which helps, but it shifts the dependency rather than removing it.

One gap to state plainly: apart from the first note, this corpus has no long-term studies of how experts' skills change over months of delegation. Most of the material is about how agents should be built, not how the people supervising them change. There's a quiet irony nearby, though. Agents trained only on expert demonstrations can't go beyond what those experts imagined Can agents learn beyond what their training data shows?. If delegation wears down human expertise, it may also wear down the source of future agent competence.