AI Is Doing the Grunt Work That Used to Turn Beginners Into Experts

Key points

• A Harvard working paper covering 65 million workers at more than 280,000 U.S. firms found junior employment fell roughly 9 percent within six quarters after companies adopted generative AI.
• A Stanford analysis of ADP payroll records found the youngest workers in the most AI-exposed jobs lost ground after late 2022, while more experienced colleagues held steady.
• Stanford researchers found losses concentrated where AI automated tasks outright; where AI only assisted, junior employment held or grew.
• Aviation's response to the same problem, pilots growing rusty because autopilots fly so well, was to build deliberate manual practice back in, not to remove the automation.
• The proposed fix for AI-augmented workplaces is a "manual gate": a designed-in moment where a human does the work without AI help, specifically to keep a skill alive.

There is a version of this story that reads like a simple economic headline: AI is taking entry-level jobs. That is true, but it is the smaller part of the problem.

The bigger issue, as IEEE Spectrum AI has explored, is what happens to an entire profession when the apprenticeship vanishes. You cannot become a senior engineer without first being a junior one. You cannot skip the frustrating, unglamorous, formative work and still develop the instincts that make an expert worth listening to in a crisis.

What does the data actually show?

Junior workers are losing ground; senior workers are not. That gap points directly at which tasks AI is absorbing.

The Harvard paper tracked 65 million workers across more than 280,000 U.S. firms. After a company adopted generative AI (the technology behind tools like ChatGPT that can write, code and reason in plain language), junior employment dropped around 9 percent within a year and a half compared with firms that had not adopted it. Senior roles kept growing. A separate Stanford analysis found the same pattern in ADP payroll data: young workers in AI-heavy occupations fell behind after late 2022, while experienced colleagues were largely unaffected.

The Stanford team found one particularly telling detail. Losses appeared where AI automated tasks completely. Where AI merely assisted, junior employment held steady or even rose. The tool is not the problem. The problem is when the tool removes the training.

New York Fed researchers offer a different explanation, pointing at remote work rather than AI. Firms are reluctant to hire inexperienced people they cannot train in person. But notice both explanations describe the same broken thing: the path through which expertise passes from one generation to the next.

Has any industry solved this before?

Yes. Aviation spent decades learning this lesson, at enormous cost.

Modern autopilots fly aircraft better than humans do in normal conditions. That is exactly the problem. The more reliably the automation performs, the less practice pilots get at hand-flying. Then, on the rare day the automation fails, you have a highly trained professional who has not truly flown the plane in years.

The clearest example is Air France flight 447 in 2009. Iced-over sensors fed the autopilot bad data. The autopilot did what it was designed to do: it handed control back to the crew. What followed was not a mechanical failure. The pilots could not read the aerodynamic situation and recover it manually. A recoverable problem became fatal.

The industry's answer was not to remove the autopilot. In 2017 the FAA issued Safety Alert for Operators 17007, formally recognising skill decay as a safety hazard and encouraging airlines to build deliberate manual flying back into routine operations. Some airlines now require pilots to hand-fly the initial climb and descent on normal flights, trading a small amount of fuel efficiency to keep raw skills warm.

That trade is the whole point. A system that produces incompetent operators is not optimised. It has just hidden its failure mode somewhere the spreadsheet cannot see.

What is a "manual gate" and how would it work?

A manual gate is a designed pause in a workflow where a human does the work without AI help, not because it is faster, but because the skill needs the exercise.

Imagine a software team that uses AI for most of its coding. The team decides that debugging, tracing a defect to its root cause, is the skill it cannot afford to lose. When a fault appears in a critical part of the codebase, a junior engineer (deliberately chosen) must reproduce the failure, trace it, and write a test capturing the bug, all with the AI assistant switched off. Only after the engineer commits to a diagnosis does the AI come back on, to propose fixes and check for similar problems elsewhere.

When the engineer's diagnosis and the AI's disagree, that disagreement is the point. It surfaces before the bad day, not during it.

This reframes the junior role entirely. Some entry-level work is not overhead to be cut. It is the training apparatus for your future senior staff, the same way a pilot's manual-flying hours are not wasted fuel but a deposit against the day the autopilot quits.

Common questions

Does this mean companies should use AI less?

Not exactly. The argument is about where you use it, not how much. Automating a task completely is different from using AI as an assistant; only the former removes the practice that builds expertise.

Is this just a problem for engineers and pilots?

No. Any profession where you learn by doing, from medicine to accounting to journalism, faces the same dynamic if AI absorbs enough of the entry-level work before practitioners have built their instincts.