Checking AI's work takes expertise — but that expertise is normally built by doing the junior work AI now does instead.

This explores why handing beginner-level tasks to AI could leave us without enough people skilled at checking AI's work, since that checking skill has usually been learned by doing the beginner tasks first. The core argument is a loop that eats its own tail. Good verifiers are experts, and experts were once juniors who learned by doing routine work, making mistakes and getting corrected. When AI takes over that routine work, the training ground goes away while demand for expert checking keeps rising. How does automating junior work destroy verification oversight? names this the "missing junior loop." It pairs it with a second failure, the "codifier's curse": experts who label data and correct AI outputs are turning their own unwritten judgment into training material for the systems that replace them. That leaves fewer and narrower chances for anyone else to learn the same judgment.

This isn't only theory. Interviews with South Korean software engineers in Does generative AI prevent juniors from getting entry-level work? show it happening. Foundational coding tasks no longer go to juniors. They go to seniors working with AI, because that's faster. Juniors lose the hard, hands-on effort that used to build their intuition. Seniors and juniors also see the problem differently: seniors see efficiency gains, while juniors feel shut out. That gap in perception is part of why the pipeline can erode without anyone deciding to let it.

There's also a quieter side effect. Juniors who do get to use AI may not notice what they're failing to learn. How do AI tools trick users into overestimating their own skills? describes how fluent AI output, unclear authorship and offloaded thinking lead people to count the AI's competence as their own, and each of these makes the others worse. So the pipeline doesn't just shrink. It can also produce people who feel ready to verify but lack the judgment to do it.

The broader AI research explains why this matters so much now. Does task verifiability determine what AI systems will learn to solve? argues that AI gets good at a task roughly in proportion to how easily answers can be checked. That makes verification the bottleneck for AI progress. Some of it can be automated: Can verification accuracy scale without training models? shows machine verifiers improving without any retraining, and Where do reasoning agents actually fail during long traces? finds that most agent failures show up in the intermediate steps, not the final answer. But spotting a bad step means knowing what a good process looks like, and that's exactly the knowledge juniors used to pick up on the job. Automated checkers cover tasks with clear right answers. The judgment calls with no answer key still need human experts.

One direction the collection suggests: if expertise is going to be captured anyway, it can at least be made visible and correctable. Can person-grounded skills remain auditable without hidden prompt state? treats an expert's captured knowledge as versioned files that people can inspect and roll back, rather than hidden prompt state. That doesn't rebuild the apprenticeship, but it turns codified judgment into something a learner could study instead of a black box. The collection identifies the problem clearly but has little on fixes. No paper here tests ways to deliberately keep junior learning loops alive alongside automation. That gap is worth noticing.