AI is quietly taking over the small starter tasks juniors learn from, and seniors and juniors don't see the same problem.

This explores why senior and junior engineers read AI's effect on junior skill-building so differently, and how that mismatch keeps the problem from being fixed. The corpus has one direct study on this. In interviews with 14 South Korean software engineers, generative AI pulled entry-level tasks into senior-AI workflows. The small, foundational jobs that used to land on juniors now get done by a senior working with an AI. That removes the hands-on struggle through which juniors historically built expertise, and the gap widens because the two groups perceive the problem differently (Does generative AI prevent juniors from getting entry-level work?). The summary tells us the perceptions diverge, but not the detail of each side's view.

The neighboring notes suggest a reason: the two groups are looking at different things. A senior sees the work product, and it is getting done faster. A junior sees what they are no longer doing. Automation research points at why the senior's view feels complete. Polished outputs hide errors instead of removing them, so the finished result looks fine while what is missing never shows up in it (Does more automation actually hide rather than eliminate errors?). That note is about scientific integrity, not junior developers, so applying it here is my reading. A lost learning opportunity doesn't appear in a pull request, a ticket, or a delivered feature. It shows only in the person who didn't get to do the task.

This also means the answer isn't simply that AI stops people learning. In a randomized experiment with 1,174 adults, AI shrank the higher-education advantage on a business problem-solving task from 0.548 to 0.139 standard deviations. Lower-education participants kept part of their gain after the AI was taken away (Can AI narrow the education performance gap?). It's a different population and task from junior engineers. Still, it shows AI assistance can leave learning behind when the person is doing the work with the AI. In the software case, the work moves to the senior and the AI, and the junior never gets that chance.

Two more findings explain why the senior's optimism is easy to hold. In a 535-person study, people using LLM help captured only about half the gain the model's accuracy improvement made available. Complementarity creates potential, not guaranteed synergy (Why does assisted accuracy capture only half the LLM gain?). Benchmarks have a similar blind spot: the field optimizes what it measures, and it has measured contests instead of real work (Why do agent benchmarks not predict real economic value?). By analogy, a team measured on shipped output looks healthier with AI, while the pipeline that produces future seniors isn't measured at all.

The corpus explains the perception gap only by inference. It has one interview study, and everything else is adjacent evidence about hidden costs and partial gains. It doesn't yet offer tested ways to make juniors' lost learning visible to seniors, or to keep the productive struggle alive once AI absorbs the easy tasks.