The Junior Problem: Where Does Judgement Come From Now?

Ask any senior professional where their judgement came from and the answer is usually some version of: years of doing work that nobody enjoyed.

Reading hundreds of contracts to develop a feel for which clause is unusual. Building models by hand until a wrong number looks wrong. Writing the first draft of everything until you know what a good one reads like. Sitting in on calls with nothing to contribute.

Almost none of it was efficient, and nobody misses it. But it was doing something beyond producing output — it was building the pattern library that judgement runs on. And it's the exact category of work AI now does in seconds.

Judgement is compressed exposure

The reason a senior person can glance at something and know it's wrong isn't reasoning speed. It's that they've seen several hundred of the thing, and the anomalous one stands out against a background they don't consciously have access to.

That background is built by volume and by consequence. You have to see many instances, and it has to matter whether you got them right — which is why watching someone else do it, or reading about it, doesn't produce the same result.

Both of those are exactly what production work provided, incidentally, while producing the output that was the ostensible point.

Reviewing isn't the same as doing

The obvious response is that juniors now review AI output instead, and reviewing is a higher-order skill.

It's a real skill, and it's not a substitute — because reviewing only works if you already have the pattern library. Reviewing something in a domain where you have no background produces a check on whether it reads well, which is precisely the check that AI output is guaranteed to pass.

There's also a subtler cost. Doing produces failure, and failure is where the strongest learning lives. Reviewing an output that's 90% right teaches much less than making the same mistake yourself and having to fix it, because you never engage with why the other 10% is wrong.

What this means if you're early in your career

Do some things the slow way, deliberately. Not everything, and not out of principle — pick the two or three skills your role is actually built on and build them with your hands. Someone in a numbers job should be able to build the model. Someone in a writing job should be able to write the thing.

Predict before you look. The cheapest exposure-building habit available. Before reading an AI output, write two lines on what you expect it to say. Then compare. The gap between prediction and result is the learning, and it takes thirty seconds.

Ask for the reasoning, not the answer. Making AI show its working turns each use into a partial apprenticeship, if you actually read it. It's not equivalent to doing it, and it is much better than reading only the conclusion.

Seek the consequential work. Judgement needs stakes. Volunteer for the things where being wrong matters and someone senior will tell you, because that combination is now rarer and it's where the development happens.

What this means if you manage juniors

Assign the thinking, not the typing. The valuable part of the old work was the pattern exposure, not the labour. You can keep the first without the second — have someone review fifty examples and report what varies, rather than produce fifty by hand.

Make them predict, then compare. Ask what they expect before they look at the AI output. Their prediction is a much better read on their development than their edit of the output, and it's the only visibility you get into whether the pattern library is forming.

Say what you noticed. Senior judgement is largely tacit, and it used to transfer by proximity — the junior who watched you reject something learned from the rejection. Less of that happens now, so say it aloud: what you noticed, what made you uneasy, why the third option was wrong. That narration is the training material.

Protect some slow work. If everything a junior touches is AI-first, they will be fast and shallow at three years, and the shallowness won't be visible until something unusual arrives.

The uncomfortable part

This is a real problem without a clean solution. The old apprenticeship was expensive and inefficient, and organisations aren't going to reinstate it out of concern for professional development. Nobody is going to hand-write fifty contracts to build intuition when the alternative takes an afternoon.

So the honest position is that judgement now has to be built deliberately rather than as a by-product — which means someone has to decide to do it, and pay for it, and it's easy to defer for a very long time before the cost shows up.

Nia — the model she built by hand

Nia joined an investment firm two years ago into a team using AI heavily. She could produce a valuation model in an hour that would have taken her predecessors a week.

What she noticed was that she couldn't tell when one was wrong. The outputs looked like models, and she had no basis for a reaction to any of them.

She now builds one from scratch each month, deliberately, on her own time budget — a full afternoon for something she can produce in an hour. Her description is that she's paying for the reps her predecessors got for free.

Graham — the two lines before looking

Graham leads a policy team in central government. His analysts were producing good briefings and, in his assessment, not developing.

His change was small: before reading any AI output, write two lines on what you expect it to say. Then compare and note the gap.

The prediction gap turned out to be the thing worth discussing in supervision — much more informative than the finished briefing, which had been telling him about the model. His summary: he'd been reviewing output and calling it development, and this is the first thing that's actually shown him someone's judgement forming.

The one thing

Judgement was a by-product of production work, and the production work has gone. Reviewing doesn't replace it, because reviewing requires the pattern library that production was building.

If you're early: do some things slowly, predict before you look, and seek work where being wrong matters. If you manage: assign the thinking rather than the typing, and say out loud what you noticed — that used to transfer by proximity and now it doesn't.