The Wrong AI Apocalypse

I saw three lines of a newsletter over coffee: a publisher trailing an essay arguing that governments should prepare for an AI jobs apocalypse. I had not yet read the article. But the premise was enough to expose the blind spot.

The public argument is still largely about displacement. How many jobs will AI destroy? How quickly? Which workers will be replaced first? Those are legitimate questions. But they are not the only ones, and they may not be the most important.

The visible panic is about jobs being cut. The quieter danger is about people never being formed.

A few decades in and around boardrooms leave you alert to a different kind of risk: not the obvious crisis, but the capability quietly being removed while everyone congratulates themselves on efficiency. The danger is not simply that AI may do work once done by people. It is that AI may remove the very work through which people acquire judgement.

That faculty — the ability to read an incomplete situation, sense what matters, detect what is missing, and know when a confident answer is wrong — is one of the most valuable assets in any organisation. It cannot be bought quickly. It cannot be downloaded. It is grown slowly, through exposure, repetition, error, correction and proximity to people who already have it.

And we may now be making it harder to acquire.

First, it is worth killing the easy panic, because the easy panic is not yet supported by the broad data. Look at the aggregate labour market and nothing is visibly bleeding. The Yale Budget Lab has tracked the American job market since the arrival of ChatGPT and found little evidence so far of economy-wide damage. Unemployment has remained low. Occupational change has not yet displayed the dramatic rupture many predicted.

The sceptics have a point: the headline numbers do not show an AI jobs apocalypse.

But that does not mean nothing is happening. It may mean we are measuring the wrong thing.

The story was never only who got cut. It is who never got hired.

Researchers at Stanford’s Digital Economy Lab have looked beneath the aggregate numbers and found a more troubling pattern. In occupations most exposed to AI, employment among the youngest workers — those aged roughly twenty-two to twenty-five — has fallen materially relative to older workers. Young software developers appear particularly exposed. The researchers are careful not to claim that AI alone explains the shift. They are right to be careful. Interest rates, hiring cycles and post- pandemic distortions all matter.

But that caution does not make the pattern irrelevant. It makes it precisely the kind of early-warning signal leaders should take seriously. Older workers in the same broad occupations have held up better. The junior layer is where the pressure shows first. Firms may not be announcing redundancies. They may simply be hiring fewer beginners.

That is not a small adjustment. It is a structural risk wearing the mask of a cost saving.

The entry-level job was never just a job. It was the first stage of apprenticeship. The junior analyst building the first version of the model, the trainee lawyer reading the dull documents, the young engineer debugging routine code, the assistant preparing the first draft — these people were not merely producing output. They were learning how the work works.

Much of that work was repetitive. Some of it was tedious. A great deal of it was inefficient. But inefficiency is not always waste. Sometimes it is the price of formation.

AI has made raw capability astonishingly cheap. Drafting, summarising, coding, comparing, modelling, translating and producing a first pass can now be done at speed and scale. That is powerful. Used well, it can make good people better. It can also help inexperienced people learn faster, if it is deliberately built into the apprenticeship process.

But that is the crucial distinction. AI can accelerate apprenticeship, or it can replace it. Those are not the same thing.

If the machine helps a junior employee compare drafts, interrogate assumptions, test alternative answers and understand why one answer is better than another, it can become a formidable tutor. If it simply does the work instead, the organisation gets the immediate output but loses the formation.

The spreadsheet arrives. The memo is produced. The code compiles. The board pack is assembled. Yet the person who would once have learned by doing has learned only to prompt, accept and move on.

The work disappears. Then the learning disappears. Finally, the judgement never arrives.

This is where many organisations will deceive themselves. They will say they are not eliminating apprenticeship; they are merely modernising it. They will tell junior people to use AI, then ask them to review whether the AI’s answer is correct, complete or good enough.

That sounds sensible. But it contains a trap.

If the junior’s response to AI output is simply to ask another AI whether the first AI was right, nothing has been learned. The machine has reviewed itself. The human has not exercised judgement; they have only managed a loop.

That is not apprenticeship. It is judgement laundering.

AI-assisted learning requires one non-negotiable discipline: the first act of judgement must be human. Before the machine is allowed into the loop, the junior person must form an unaided view. What is the issue? What matters? What is missing? What would they recommend? How confident are they, and why?

Only then should AI be introduced as a comparator, critic or tutor. The learning lies not in accepting the machine’s answer, but in explaining the difference between the human view and the machine’s view — and defending that explanation to someone with more experience.

That discipline matters because many young employees will not arrive fully formed as critical thinkers. Why would they? Judgement is not a credential. It is not conferred by a degree. It is developed by repeated exposure to real problems, real consequences and real correction. If employers assume it already exists, they will be disappointed. If they outsource its formation to AI, they will make the problem worse.

You cannot hire judgement in a hurry. You grow it, slowly, on the job, beside people who already have it. Or you do not grow it, and one morning you discover that the organisation is full of people who can operate powerful tools but cannot reliably tell whether the tool’s answer is brilliant, irrelevant or catastrophically wrong.

That is the famine we are planting. We are eating the seed corn and calling it efficiency.

This should sound familiar. Across the closing decades of the last century, Britain and America made a series of decisions about skilled work that looked rational at the time. Training was expensive. Apprenticeships were slow. Skilled labour could be obtained elsewhere. Manufacturing could be outsourced. Practical expertise could be allowed to wither because, in the short term, the arithmetic worked.

Every annual decision looked defensible. The long-term result was a loss of capability.

A generation later, both countries struggle to find enough of the skilled hands needed to build houses, power grids, factories and infrastructure. Ministers can announce programmes. Boards can approve capital expenditure. Consultants can draw charts. But you cannot conjure a master craftsperson inside a budget cycle. Once a capability chain is broken, rebuilding it takes years.

And it was not inevitable. Germany faced many of the same pressures — globalisation, lower-cost labour abroad, changing industry — but preserved a deeper commitment to vocational formation through its dual apprenticeship system. It did not solve every problem, and no country’s system is perfect. But it retained a national habit that Britain and America weakened: the deliberate reproduction of skill through work.

That is the precedent. We have already seen what happens when societies stop reproducing skilled people. The damage is invisible for years, then suddenly it is everywhere and cannot be repaired quickly.

The only difference this time is speed and silence.

No minister will announce that the country has stopped growing judgement. No chief executive will tell investors that the company has saved money by weakening its future leadership bench. No board paper will say: “We have automated the formative layer through which our future senior people used to learn.”

It will simply appear later as a shortage no one can explain. Not a shortage of clever people. There will be plenty of those. Not a shortage of people able to use AI tools. There will be millions of them. The shortage will be of people with enough accumulated judgement to govern the work, challenge the machine, carry responsibility and make decisions when the data is incomplete.

This is the boardroom version of the problem. The risk is not simply that the organisation becomes thinner at the bottom. It is that, over time, it loses the capacity to govern itself.

Organisations do not fail only because they lack intelligence. They fail because they misread reality, suppress dissent, mistake fluency for truth, underweight second-order consequences, and allow systems to drift beyond human understanding. AI does not remove those dangers. In some cases, it magnifies them.

A company full of brilliant machines still has to be an organisation capable of governing them.

That requires human judgement. It requires people who understand not only the answer, but the work behind the answer. It requires enough lived experience to know when something feels wrong before the dashboard proves it. It requires people who have been formed by the discipline of doing, checking, failing, being corrected and trying again.

Yet the economic temptation points the other way. Every executive who quietly stops backfilling junior roles can explain the decision. The work is being done faster. The costs are lower. The output is acceptable. The team is leaner. The technology is improving. The business case is obvious.

At the level of the individual firm, the decision may be rational. At the level of the system, it may be insane.

If every firm stops training and simply tries to hire experienced people trained elsewhere, the arithmetic works beautifully until the pool runs dry. Each company believes it is being efficient. Collectively, they are dismantling the supply chain of experience on which they all depend.

There is, of course, a strong objection. The champions of the technology will say this argument underestimates AI. Why assume that human judgement remains scarce? Perhaps AI will soon supply judgement of its own — not merely pattern recognition, but practical wisdom, strategic sense and sound decision-making.

Perhaps it will. But look at what that argument asks leaders to do. It asks them to dismantle the only proven supply of judgement they have today on the promise that a better one may arrive later. That is not a strategy. It is a bet on an unshipped feature.

If machine judgement comes late, comes unevenly, or does not come at all, the organisation is left with neither the judgement it failed to grow nor the judgement it hoped to buy.

And even if the optimists are right, the problem does not disappear. If AI becomes more capable, the need for human governance does not fall. It rises. Someone must still decide what the system is for. Someone must notice when it is optimising the wrong objective. Someone must detect drift, bias, hallucination, manipulation or overreach. Someone must carry responsibility when the machine’s recommendation fails.

That someone cannot be hollow. They need judgement of their own.

So the serious question for leaders is not whether to use AI. That argument is over. The serious question is how to use AI without destroying the human formation on which the organisation’s future depends.

The trap is to treat AI as something to buy and bolt onto existing work. That is how the judgement layer gets stripped out by accident. The harder and more important task is to redesign work around the technology: to decide deliberately where AI replaces work, where it accelerates work, and where humans must still do the formative tasks because those tasks are how they learn.

The board-level question is simple:

Which junior tasks have we automated, and which of those tasks used to build the judgement we will need five years from now?

But the next question is just as important:

Where AI is being used in development roles, do we still require a human first pass before the machine is allowed to critique, complete or correct the work?

Most organisations will not have a good answer. They may have an AI strategy. They may have a technology roadmap. They may have policies, pilots, vendors and productivity targets. But few will have mapped the connection between work removed today and judgement missing tomorrow. Fewer still will have designed the discipline that prevents AI-assisted learning from becoming AI-assisted dependency.

That is where serious governance should begin.

What serious leaders should do now

First, separate output from formation.

Some work is merely output and should be automated without sentimentality. No organisation has a duty to preserve pointless drudgery. But some work is formative even when it looks inefficient. It teaches pattern recognition, discipline, context, error detection, client sense, commercial instinct and professional confidence.

The question is not simply: “Can AI do this task?”

The better question is: “What did this task used to teach?”

If the task taught nothing, automate it. If it taught judgement, redesign the apprenticeship before removing the work.

Second, require a human first pass.

This is the rule most organisations will miss. AI should not be the junior employee’s first resort in every formative task. There must be moments where the person is required to think before the machine assists.

That does not mean banning AI. It means sequencing it properly.

Before using AI, the junior should write down their own view: the issue, the options, the recommendation, the risks, the missing information and their confidence level. Then AI can be used to

challenge that view, expose gaps, test alternatives or suggest improvements. After that, the junior must explain what changed and why.

The discipline is simple: human judgement first, machine critique second, human reconciliation third, senior review last.

Without that sequence, the person may become fluent in using AI without becoming competent in the underlying work.

Third, rebuild junior roles rather than eliminate them.

The junior employee of the AI age should not be left doing obsolete manual work for the sake of nostalgia. Nor should they be reduced to a passive consumer of machine output. The better model is supervised acceleration: AI as tutor, critic, simulator, comparator and second pair of eyes, with the human still required to reason, decide, defend and learn.

A junior analyst should not merely ask the machine for the answer. They should be required to compare the machine’s answer with their own, explain the difference, identify missing assumptions and defend the recommendation. A trainee lawyer should not merely accept the draft. They should be asked what the draft misses, what it overstates and where the legal risk sits. A young engineer should not merely paste generated code. They should be tested on why it works, where it might fail and how it should be improved.

That is not anti-AI. It is the serious use of AI.

Fourth, make apprenticeship explicit again.

For years, many organisations have pretended that development happens naturally if talented people are hired and kept busy. That was always lazy. In an AI environment, it becomes dangerous. If the routine work no longer teaches by default, leaders must design the teaching deliberately.

That means named mentors, deliberate review, structured exposure to messy problems, and real accountability for managers who are supposed to be growing people rather than merely extracting output from them. It means treating the development of judgement as a leadership responsibility, not a human resources slogan.

Fifth, measure erosion of judgement as a real risk.

Boards already ask about cyber risk, compliance risk, succession risk and operational resilience. They should now ask about capability reproduction.

Are we still growing people who understand the work deeply enough to govern it?

Can our people reason without the machine?

Where have we automated formative work?

Where do we require a human first pass?

Which parts of the business are becoming dependent on tools no one truly understands?

Where have we removed junior work without replacing the learning mechanism?

Are we rewarding managers for developing judgement in others, or only for cutting cost this quarter?

These questions belong in the boardroom because the consequences will land there. A company that loses the capacity to form judgement is not merely creating a future talent problem. It is weakening its own governance system.

Some of this is larger than any single company. A trained workforce is a shared inheritance. Every firm draws on it. If no firm helps replenish it, the pool runs dry. That is the case for renewed apprenticeship systems, public early-career schemes and incentives that reward the creation of capability rather than only its consumption.

But the collective nature of the problem is no excuse for individual passivity. The company that keeps forming judgement while its rivals strip theirs out will have an advantage. It will have people who can govern the work, challenge the technology, train others and make decisions under uncertainty. It will be able to hire from within when others are searching a market they have helped empty.

Protecting the judgement pipeline is not civic charity. It is competitive strategy.

AI will not do this to us by itself. We will do it, if we choose to — one reasonable-looking saving at a time, each decision defensible, each spreadsheet persuasive, each quarter improved, all the way to a future in which the organisation no longer contains enough people capable of governing what it has built.

That is the wrong apocalypse.

The danger is not simply that machines take the jobs. The deeper danger is that we remove the work through which humans learn to exercise judgement, and then discover too late that no machine can absolve us of the need to govern.

The good news is that a disaster we are inflicting on ourselves is one we can still call off. The technology is not destiny. The choice is in the design of the work.

So the question I would put to any board is not whether your people will be clever enough ten years from now. They probably will be. The question is whether the organisation you are building will still contain the judgement to govern itself.

At the moment, too many companies are heading the wrong way. There is still time to prove otherwise.