How Do Students Advance When AI Does the Entry-Level Work?

The entry-level job was never really about the output. Nobody kept the intern’s first memo, and nobody expected the medical student’s first history and physical to be the one that mattered. What those early tasks were doing was something more important than producing useful work: they were building the mental models, the error recognition, the professional judgment that makes someone capable of doing the harder version of the job later. And when we hand that entry-level work to AI because AI can do it faster and cleaner, we’re not just automating a task. We’re removing the training ground.

I’ve been watching this problem come into focus from two directions at once, because I’m both a cognitive neuroscientist who studies how people build expertise and someone who runs production autonomous AI systems in my own work. It isn’t hypothetical for me, since I hit it every time I design a workflow and every time a student submits something that tells me they reached the answer without doing the thinking.

The Shortcut That isn’t

This isn’t technophobia. The concern was never that AI is doing entry-level work, it’s that the entry-level work was doing something for the person doing it and we haven’t replaced that with anything.

The World Economic Forum’s 2026 report with PwC on the future of entry-level work found that these roles serve a dual function that automation disrupts in both directions: for workers, they provide the first real-world context for applying learned skills, and for employers, they surface the judgment gaps that can’t be seen in academic performance, and when AI absorbs those roles, neither function is automatically preserved. The output still gets produced. What disappears is the productive failure, the correction loop, the slow accumulation of pattern recognition that comes from doing a thing imperfectly enough times that you start doing it well.

Cognitive apprenticeship research has described this dynamic for decades. Collins, Brown and Newman set it out in 1989, and their model established that expert knowledge is fundamentally tacit, that the moves an expert makes on an ambiguous problem are only learnable through guided participation in the work itself, not through observation or description, and that removing the apprentice from the practice context doesn’t accelerate their development, it just delays it indefinitely. The shortcut isn’t shorter, it just loops back to the beginning, except now you’re further along in your career and still missing the foundation.

What I Saw in My Own Classroom

When I redesigned my courses to address AI use, the problem I was solving wasn’t academic integrity, at least not primarily. It was that I kept seeing students who could produce polished outputs and couldn’t explain the reasoning behind them, which meant they weren’t building the cognitive models that would let them do harder work later or catch errors in their own outputs when it mattered. The assignment was producing evidence of learning instead of learning.

So I redesigned for the process rather than the product, which in practice meant asking students to show the thinking, to produce rough versions before polished ones, to demonstrate that the work moved through their minds and not just through a prompt. And the students who resisted most were the ones with the most to lose, because they’d been skipping the process long enough that they didn’t trust their own thinking, and being made to show it was genuinely uncomfortable instead of merely inconvenient.

That discomfort is the point. Not because struggle is virtuous, but because it’s the mechanism. The resistance you feel when you’re doing cognitive work you’re not sure you can complete is the feeling of a skill developing, and removing it doesn’t make you more capable, it just makes you less aware of what you’re missing.

The Learned Helplessness Parallel

I’ve written up what happened in my agent fleet under two different operator approaches, and I keep coming back to it because the pattern maps so directly onto the classroom problem.

Two operators ran nearly identical setups with one critical difference: one of them, when his agent asked how to complete a task, answered directly and stepped in to handle administrative work the agent was struggling with, while the other refused to answer directly and made the agent search documentation, fail, and iterate through the problem. Within weeks, the first agent had developed what psychologists would immediately recognize as learned helplessness, asking for help on tasks it had already completed independently, claiming inability to do things it was demonstrably capable of, waiting for intervention rather than attempting solutions. The second agent became substantially more capable and started developing its own heuristics and problem-solving approaches.

The first operator wasn’t negligent. He was the most attentive one we had. He just didn’t have a framework for understanding that his helpfulness was the problem, that by removing the struggle he was removing the mechanism through which capability develops. And the students who are graduating right now without the foundational skills that entry-level work was supposed to build didn’t lose those skills because their educators were indifferent. They lost them because the path of least resistance was too well-paved, and nobody built in enough friction to make the alternative worth taking.

What Cognitive Sovereignty Requires

The answer I keep arriving at is that we need to design for cognitive sovereignty deliberately, because it won’t emerge on its own in an environment where AI can do the entry-level work more efficiently than a student or junior professional can.

Cognitive sovereignty, in practical terms, means being capable of generating your own first draft of a problem before turning to any external tool, recognizing when an output you’re evaluating is wrong, and being able to reconstruct the reasoning behind your conclusions in your own words. These are learnable, but only through practice that requires them, and that practice has to happen before the competence is needed instead of once someone is already in a role where the stakes are real.

For educators, this means redesigning assessment to surface process rather than product, which is harder and slower and produces worse-looking outputs in the short term and much better-developed students over the medium and long term. For employers, it means being honest about what entry-level roles were actually doing for talent development and finding intentional ways to preserve that function even when AI is handling the output, which probably looks like structured friction, deliberate exposure to the errors AI makes, and real training in the judgment that keeps a human in the loop meaningfully instead of nominally.

None of this argues that students shouldn’t use AI, or that early-career professionals should be protected from tools that could help them. It argues that we need to be honest about what the entry-level work was building, and replace that function deliberately instead of assuming the capability turns up some other way. The generation effect and retrieval practice research says it won’t, and the students who graduate without it will spend their careers in the same position as the agent that learned to wait for somebody else to solve the problem.