The training gap, part one: AI has to be in the curriculum

Part one of two. Production used to do the teaching. It has been absorbed, and nothing replaced it — so here is an attempt at what should. Part two is the three things about this proposal I can’t solve.

I ended an earlier piece by asking who is designing the replacement for the training that routine production used to provide.

My answer was nobody. It happened by accident for long enough that it never acquired an owner, a budget or a line on anybody’s objectives.

I’ve written that worry four or five times now without attempting a solution, which is starting to look less like intellectual honesty and more like a habit. So this is an attempt, offered with low confidence.

The short version: AI use has to be inside the training rather than beside it — in journalism and editing courses, in graduate schemes, in the first two years of a career. Not as a separate module. Inside the craft teaching itself.

Where this keeps arriving from

Worth setting out, because I didn’t go looking for this problem and it isn’t a hobby-horse. It keeps turning up at the end of arguments about something else, which is usually a sign that something structural is underneath.

Raising the floor is not building a ladder is the origin. My own project evidence showed AI helping the least experienced person most, which cuts against the worry — and the uncomfortable conclusion was that both can be true. Output improves while the thing that used to produce judgement disappears.

The best AI course I’ve taken is the thing I built for myself hit it from the other side. Building your own system works because fifteen years of judgement gave me a domain to build it around. As advice to somebody starting out it’s close to useless, and I had to say so in the piece.

You can give someone my AI tools. You can’t give them my triggers. ran into it while conceding that triggers are probably trainable rather than innate. If noticing is a craft skill built by years of practice, then the question is immediately where somebody builds it now.

Being an ageing content writer may have made me comfortable with AI reached it as a limit on the whole argument. The comfort depends on having already accumulated the twenty years. It is not transferable advice.

The agency model doesn’t have to die found it in the supplier economics. The agency pyramid was the training institution for an industry, and the model everybody now proposes removes it while requiring its output.

Five arguments, five different starting points, same unresolved ending. That’s the reason for attempting something here rather than writing it a sixth time.

The comparison everybody reaches for, and where it fails

The obvious precedent is print to digital.

An entire profession had to retrain. People who had spent careers thinking in column inches learned to think in pages, then in feeds. The skills that mattered shifted, some people didn’t make the transition, and the industry came out the other side still recognisably itself.

It’s a comforting story and I’ve caught myself using it. It has a problem.

Digital created entry-level work while it destroyed other work. Content farms, SEO agencies, social teams, community management, the whole apparatus of the 2010s web. Much of it was poor work — I’ve been rude about that era at length — but it employed enormous numbers of people at the bottom, and they learned things while doing it.

The transition worked partly because the disruption came with a hiring boom attached.

There is no equivalent here. AI absorbs entry-level work without generating a new category of it. So the comparison tells you that industries can retrain, which is true and worth knowing. It doesn’t tell you where people stand while they do it, which is the actual question.

Treat it as evidence that the thing is possible, not as evidence that it happens on its own.

Not a module about prompting

The version of this that will get built is a module about prompting, bolted onto an otherwise unchanged course. It will be popular, it will be easy to timetable, and it will be worthless within eighteen months.

Prompting is interface knowledge. It dates at the speed of the interface, which is roughly the speed at which teaching a CMS in 2005 dated. Anybody currently building a curriculum around prompt technique is teaching the part that will be automated next.

What doesn’t date is judgement about output — and that has to be taught against the machine rather than in isolation from it, because the failure modes are specific.

Can you tell when a paragraph is fluent and says nothing?

Can you spot a claim that has been asserted confidently and has no evidence behind it?

Can you notice when an argument is nearly right in a way that will embarrass you in public?

Can you tell the difference between a piece that made a decision and a piece that averaged several?

Those are teachable. They’re also testable, which matters more than it sounds — you can hand somebody five pieces of machine output, three of them subtly wrong, and find out what they can see. That is a marking scheme, and marking schemes are what turn a good intention into a course.

The failure mode this creates

The proposal that follows naturally — juniors learn by editing AI output instead of by producing first drafts — has a problem I want to name before somebody else does.

Editing machine output teaches judgement only if somebody senior is checking the editing.

Without that, you don’t produce editors. You produce approvers. I’ve written about that distinction as a personal risk; as a training model it’s considerably worse, because an approver who has never been corrected doesn’t know that’s what they are. The output looks identical from outside. Something plausible got published and nobody objected.

Which means the expensive part of the old apprenticeship — a more experienced person spending real time on somebody else’s mediocre work — doesn’t go away. It moves. It stops being fixing the writing and becomes interrogating the judgement, and it is not cheaper. It might be worse value in the short term, because the output was already publishable before the senior person got involved.

That is the actual cost of this, and any version of the proposal that doesn’t budget for it is decoration.

The evidence, once I went looking for it

I wrote the paragraphs above from personal observation, and I should be honest that until this month they were an assertion rather than a finding.

A study from KPMG and the University of Texas at Austin, published in Harvard Business Review in July, put 523 early-career professionals through real business tasks with an AI agent and measured who added value beyond what the AI produced on its own. It sorted them into three groups. Amplifiers — just over half — directed, challenged and refined the AI’s work and beat it. Delegators — about a quarter — largely accepted what the AI gave them and matched it, no better and no worse. Apprentices — the remaining quarter — critiqued the AI’s output as readily as the amplifiers did, and made it worse anyway, chasing the wrong issues or steering it in the wrong direction.

The finding that matters most here: amplifiers and apprentices tested at similar levels of critical thinking, domain knowledge and AI literacy going in. Resume and aptitude didn’t predict which group somebody landed in. What separated them was closer to judgement-in-use than judgement-in-principle — not knowing that the AI should be challenged, but knowing what, specifically, to challenge.

That’s a sharper version of the approver-versus-editor distinction two sections up, and it isn’t mine. It’s an empirical answer to a question I could previously only assert.

There’s a second, cruder data point worth adding beside it, because it answers a different question. The KPMG/UT Austin study is about how junior people perform once they’re already in a role and using AI. Separately, UK graduate vacancies fell 45% year on year and SignalFire found a 50% drop in post-graduate role starts across functions — the first evidence that the roles themselves may be getting scarcer, not just that what happens inside them is changing. The fuller account is here, which is where that argument actually lives.

Nobody has the incentive

Universities can put AI into a curriculum but can’t produce the second half, which needs real work and experienced people correcting it.

Employers can produce the second half and have every reason not to. The output is already acceptable. Slowing down to develop somebody costs money now and pays back to whoever employs them in four years, which is a textbook case of a benefit landing outside the organisation that funded it.

This isn’t a moral failing on anybody’s part. It’s a structural gap, and structural gaps don’t close because people write essays about them.

The honest position is that I can describe what should be taught and I can’t tell you who pays for it. The one thing I’d say to anybody running a team: the cost of not doing it is invisible for about five years and then permanent, which is exactly the profile of a decision that never gets made deliberately.

What the deliberate version could actually look like

The same research points at something more useful than a diagnosis: a few concrete ways to make the second half of the old apprenticeship visible again, rather than only asking somebody to fund it in the abstract.

Junior staff could be asked to show the decision, not just the deliverable — which assumptions they questioned, what they checked, what they changed and why. Managers could compare a task’s AI-only output against the final human version and discuss where the difference came from, which turns “was this good” into “what did the person actually add.” Some work could still be done without AI occasionally, not because manual production is virtuous but because supervising a process is easier once you’ve done it yourself. And responsibility could scale up gradually — bounded decisions with fast feedback, before anyone is trusted to direct an AI system on something consequential — rather than jumping straight from execution to supervision.

Since I wrote the proposal, an MIT committee has reached the same measurement problem from higher education. It found that AI could already produce credible responses to almost any written assignment in the undergraduate curriculum, and recommended oral exams, portfolios and in-person conversations tied to work completed elsewhere. Those are different formats, but they make the same move: assess the reasoning around the answer rather than treating the finished answer as proof that learning happened.

That is significant evidence that the old proxy is breaking. It does not solve the harder part of this proposal — how people acquire judgement, who spends time correcting it and who pays — but it shows a major institution accepting that improving the assignment will not be enough. The assessment itself has to expose more of the process.

None of that answers who pays for it. But it turns “somebody senior needs to spend real time on this” from a vague cost into a specific list of things that time buys, which is at least a start on the budget I said I couldn’t produce.

What this proposal doesn’t survive

That’s the argument. Put the tool inside the training, teach judgement against its output, accept that somebody senior still has to spend real time on work that already looks finished, and find whoever is going to pay for it.

I’d leave it there if I thought it held up under pressure. It doesn’t, in three specific ways — one about the people who refuse these tools on grounds that aren’t stupid, one about the much larger group who tried them and couldn’t see the point, and one about whether any of this sorts itself out regardless.

Part two is those three problems. None of them is rhetorical, and I don’t have answers to any of them.