When AI takes over more of a job, what actually keeps people getting better at it?

This explores what keeps people learning on the job once AI starts doing more of the work. One caveat first: the collection has little on institutions in the usual sense, like apprenticeships, schools, professional bodies or regulation. What it has is closer to the ground: evidence about when automation erodes skill, some proposed design fixes, and labor economics about which human work stays valuable. Put together, these suggest that learning is lost or kept mainly through how work is divided up and designed, not through any single institution.

The clearest finding is about where skill loss happens. A meta-analysis of 18 experiments found that automation improves routine performance but hurts people's ability to recover when the system fails and to notice what's going on around them. The damage is worst when automation moves from *analyzing information* to *choosing actions* When does automation harm performance after failures?. So the line that matters isn't how much gets automated. It's whether the human still makes the decision. A person who reviews an AI's analysis and then decides keeps practicing judgment. A person who only watches an AI decide gradually loses the ability to step in. Any organization that wants to protect learning has to decide deliberately where that line sits.

The nearest thing the corpus has to an institutional answer comes from Human Factors research, which has studied this problem in aviation and industrial control for decades. Simkute and colleagues apply five of its principles to generative AI: continuous feedback, personalization, interfaces that show the system's state, keeping tasks stable so people's mental models stay valid, and clear rules about who does what Can Human Factors design fixes restore GenAI productivity losses?. The last two act like institutions inside a team: agreed rules about which work stays human. These proposals are untested for generative AI, so treat them as a research agenda rather than a playbook.

The economics notes add a less obvious point. When a job's tasks depend on each other, so one weak link spoils the result, automating some tasks can make the remaining human tasks *more* valuable, because people's time and attention concentrate there Does automating some tasks actually increase the value of remaining work?. McKinsey similarly finds that over 70% of current skills apply to both automatable and non-automatable work, which points to skills being used differently rather than becoming obsolete Will most worker skills survive AI automation?. The open question these notes leave is how anyone gets *good* at the high-value bottleneck tasks once the easier tasks that used to serve as practice are automated.

This is where the "verifier's rule" becomes useful. Wei argues that AI solves tasks roughly in proportion to how easily answers can be checked Does task verifiability determine what AI systems will learn to solve?. This next point is an inference, not something the corpus states: easy-to-check tasks are often the ones beginners learn on, because clear right-or-wrong feedback is what makes practice work. If those tasks go to AI first, the traditional path from novice to expert breaks early. The structure that would preserve learning is one that deliberately keeps some of that checkable practice work for people, even when a machine could do it faster.