ROBERT KARASZEWSKI

Human-Led AI Strategy

Why Do Some Junior Employees Learn Faster With AI While Others Fall Behind?

Written by

Robert Karaszewski

AI is not slowing junior employees down on average. It is splitting them. In the first field experiment to measure unassisted judgment after months of real AI use, junior professionals gained the most in assisted work and nothing on average in unassisted work, because their results spread in both directions. The best available evidence on what separates the two groups points to how each junior uses the tool, which managers can see and shape.

The usual worry is that AI removes the routine work juniors learn from. That assumes a uniform effect. The evidence shows something less uniform and more manageable.

Is AI slowing down how junior employees learn?

Not on average. It is widening the spread among them.

In a pre-registered trial circulated by the NBER in September 2026, David Autor and colleagues randomized 133 patent lawyers at eleven US firms into three months with or without an AI drafting assistant. Blinded patent attorneys scored the work. With AI, drafting quality rose across the board, and most for juniors: about 0.6 standard deviations, against 0.2 to 0.3 for seniors.

Then every lawyer marked up a flawed patent without AI. Seniors who had spent three months with the tool outperformed senior controls by 0.45 standard deviations. Juniors showed no average difference from junior controls. Underneath that zero, the distribution moved: the share of juniors scoring in the bottom fifth rose by 22 percentage points, the share in the mediocre band fell by 42, and the share scoring well rose by 19, an increase too imprecise to treat as established. The authors read the pattern as AI serving as a “springboard” for some juniors and a cushion for others. Three months of assisted work opened a junior–senior gap in unassisted judgment that did not exist among the controls.

The limits are real. This is a working paper, not yet peer-reviewed. Only about two dozen juniors completed the unassisted task. Google funded the study, most co-authors work there, and the tool was Google’s. It covers one profession over three months.

Why does AI sharpen senior judgment but not junior judgment?

Because seniors bring the judgment that turns assistance into learning.

In follow-up interviews, seniors described treating the tool’s output as a “logic auditor”: something to argue with. They rebuilt claims and tied their edits to legal doctrine. Juniors, with or without AI, worked from top to bottom, polished low-stakes sections, and often described a flaw in a comment without fixing it. Three months of AI did not cure that habit. The authors’ tentative conclusion is that foundational expertise may be the precondition for extracting durable skill from AI-assisted practice — the same mechanism that explains why AI is not the leveler it appears to be.

Juniors also misread their own trajectory. Their assisted drafts matched or beat those of untreated seniors, and their satisfaction with the work rose far more than seniors’ did. Their unassisted judgment did not move. Self-assessment is not a usable signal here.

What separates juniors who learn with AI from those who don’t?

The patent study cannot say, and its authors state as much. The clearest evidence comes from a smaller experiment.

Judy Hanwen Shen and Alex Tamkin, working at Anthropic, randomized 52 developers learning an unfamiliar programming library to work with or without an AI assistant, then tested everyone unaided. The AI group scored 17 percent lower and was not significantly faster. Inside that group, how people used the tool decided the outcome. Those who delegated the code, slid gradually into delegation, or had the AI debug for them averaged below 40 percent. Those who asked only conceptual questions, requested explanations alongside generated code, or generated code and then interrogated it scored 65 percent or higher. The widest gap was in debugging: participants without AI hit more errors and fixed them themselves.

This was a one-hour task, with clusters of two to seven people, released as a preprint by researchers at an AI developer. Applied to the patent result, the likely reading — an inference, not a finding — is that the juniors who rose interrogated the tool and the juniors who fell delegated to it.

How can a manager tell which juniors are actually learning?

Not from their output, because output is exactly what AI equalizes. Watch the process instead:

  • They ask why before they ask for an answer.
  • They can explain any AI-generated work they submit, without the tool open.
  • They execute the correction, not just flag it.
  • They still hit errors and resolve them on their own on a protected share of their work.

Then confirm it by measuring unassisted performance at a fixed interval. The four signals tell you where to look; the unassisted sample tells you what is actually there.

What should change in how juniors use AI?

The configuration and the review, not the access.

Review reasoning, not output: ask juniors to defend one choice in every piece of assisted work. Require execution: the junior makes the fix, the senior checks it. Configure tools to explain before they produce. In a high school mathematics experiment with nearly a thousand students, unrestricted GPT-4 raised practice scores but lowered later exam scores; the same model set up as a tutor that withheld answers removed the harm.

The pressure point differs by market. In the United States, velocity and utilization metrics reward the fastest path through a task, which is delegation. In the UAE, national workforce programs place graduates into private-sector roles by design, so the pipeline exists by policy and its yield depends on whether early roles build judgment or only output. In Central and Eastern Europe, where cost pressure favors AI over junior headcount, the firms still hiring juniors gain most from making each one a learner rather than a cushion.

Frequently asked questions

Should junior employees be barred from using AI?

No. The juniors who improved did so with the tool. A ban removes the springboard along with the cushion. Shape the use and measure the result.

Does this apply beyond law and software?

Probably, but the evidence is thin and specific. Two settings, one three months long and one an hour long, point the same way. Treat it as a hypothesis and test it in your own function with an unassisted sample.

Do AI “learning modes” solve the problem?

Partly. They push users toward explanation rather than delegation, which is the pattern associated with learning. A junior under time pressure can still route around them, so the review and the measurement remain necessary.

Sources and further reading