If AI Does the Doing, How Do We Learn the Judgement?

I’ve been coding, badly and enthusiastically, for years.

I’m self-taught. That has traditionally meant learning by doing: trying something, getting it wrong, staring at an error message, searching for an answer, changing something, breaking something else and gradually developing a feel for how software works.

AI has changed that experience dramatically.

I can now build things that are beyond my unaided technical ability. The AI’s ability to produce software is increasingly greater than my ability to independently produce the same software.

That is extraordinary. It is also beginning to make me wonder about something much bigger than coding.

If AI increasingly does the doing, how do people learn the judgement?

Some of the drudgery was apprenticeship

We tend to describe repetitive or junior work as though its only purpose were the output.

A junior developer writes relatively simple code. A new analyst builds the basic model. A trainee drafts the first version. Someone early in their career does work that a more experienced colleague could probably complete faster and better.

Looked at purely through a productivity lens, much of that work is an obvious candidate for AI.

But the work was doing something else as well.

It was creating experience.

People learned what good looked like partly by producing things that weren’t very good. They learned where systems break by breaking them. They learned which questions mattered by initially asking the wrong ones. They developed judgement through repetition, feedback and exposure to progressively harder problems.

Some of what we’ve been calling low-value work was also, quietly, apprenticeship.

We may be asking people to jump further up the ladder

AI can allow someone to produce work that requires more expertise than they currently possess.

I experience this directly when I build software. I can describe what I want, interrogate the result, test it, challenge it and iterate. I can operate at a level that would have been inaccessible to me without AI.

That can be an incredibly powerful way to learn.

But there is an important difference between being able to produce something and independently possessing the expertise required to produce it.

The gap matters because somebody still has to judge the result.

As AI gets better at execution, the human role increasingly involves framing the problem, directing the work, spotting when the output is wrong, understanding consequences and deciding whether the answer should be trusted.

Those are often the things experience teaches.

We risk creating a strange development problem: the technology lets people skip some of the work through which previous generations acquired the judgement we now expect them to exercise.

Organisations don’t only produce outputs

This is easy to miss if we look at AI adoption only through productivity.

If a team can produce the same output with fewer people, that may be a genuine efficiency gain. If experienced staff can delegate routine work to AI and focus on harder problems, that can be valuable too.

But organisations don’t only produce outputs. They also develop capability.

We can’t just keep hiring experienced people forever. Somebody has to make them.

That means the most productive way to complete today’s task may not always be the best way to develop the person who needs to handle tomorrow’s much harder task.

The answer cannot be to preserve unnecessary work simply because previous generations had to do it. That would be like insisting people learn arithmetic without calculators because that is how we learned.

Instead, we need to understand what the old work was teaching and decide how that learning happens in the new environment.

We need to replace the learning, not just remove the work.

AI may help solve the problem it creates

There is an optimistic version of this story.

AI does not only remove learning opportunities. Used well, it can create entirely new ones.

A capable system can explain why something works, critique a decision, generate exercises, expose someone to situations they would rarely encounter, simulate different outcomes and adapt its explanation to the learner.

Someone early in their career could potentially receive forms of individual coaching that were previously impossible to provide at scale.

But that will not happen automatically.

If we optimise AI purely for output, the easiest path may often be to let the system do the difficult work and give the person the answer. If we care about developing capability, we may need to design some workflows differently: sometimes asking the person to reason first, sometimes requiring them to review and challenge an AI answer, sometimes making the system explain rather than simply complete.

The objective is not to make work artificially difficult.

It is to make sure productivity today does not accidentally consume the experience we need tomorrow.

This feels like a leadership problem

The apprenticeship question is ultimately a leadership and workforce-design problem, not a reason to slow down AI adoption.

AI gives organisations an obvious incentive to optimise for productivity, particularly when there are pressing problems to solve and a short horizon for delivering value. That is entirely rational.

But organisations don’t only produce outputs. They also develop capability.

The leadership challenge is holding both at once: using AI aggressively to improve today’s service and solve today’s problems, while being deliberate about how people build the judgement and experience the organisation will still need tomorrow.

That may mean changing career paths. It may mean treating review and challenge as development activities rather than overhead. It may mean deliberately exposing less experienced people to decisions that AI could technically make for them. It may mean using AI itself as a tutor, simulator and critic.

What matters is that we make the choice consciously.

We’ve spent a lot of time asking how AI can make experienced people more productive. We may need to spend the next few years asking how inexperienced people become experienced in the first place.