The Missing First Rung: Young Workers and the Fifteen-Fold Disagreement

Part III of V. AI has not fired the young. It has quietly stopped hiring them.

I. Nineteen per cent

There is a number I keep returning to this autumn, and it is not a productivity figure. It is 19.

In August, Erik Brynjolfsson and his colleagues at the Stanford Digital Economy Lab updated their Canaries in the Coal Mine study, which is built on ADP payroll records. Employment of 22 to 25 year olds in the occupations most exposed to AI now sits about 19% below where it would be had it tracked same-age workers in less exposed jobs. A year earlier, the gap was 15%. (One outlet gives the earlier figure as 13%. The direction of travel is the same either way.)

Three details in the update matter more than the headline. The fall comes mainly from reduced hiring, not from more people being let go. It is concentrated in roles built on codified knowledge, the kind you can write down in a manual. And experienced workers, who lean on tacit knowledge, show no comparable decline.

Put plainly: AI has not fired the young. It has stopped inviting them in.

Part I of this series looked at Stanford HAI’s AI Hopes and Fears interviews, in which three of six leaders worried specifically about young people. This essay is about why they are right to, and why economists who agree on almost nothing else agree on this.

II. The fifteen-fold disagreement

Start with the disagreement, because it is enormous. Estimates of AI’s contribution to annual productivity growth now span roughly 0.07 to more than 1 percentage point. That is about a fifteen-fold range, among serious people looking at the same economy.

At the optimistic pole sits Brynjolfsson himself. In an FT op-ed, as reported by Fortune, he put US productivity growth at about 2.7% in 2025, nearly double the decade’s 1.4% average, and argued that the economy is leaving the investment-heavy part of the productivity J-curve for a “harvest phase”. He concedes that more quarters are needed to confirm it. Philippe Aghion and co-authors, reasoning from historical analogy, put the likely gain at 0.8 to 1.3 points a year.

At the sceptical pole sits Daron Acemoglu, whose Economic Policy paper caps the ten-year gain in total factor productivity at no more than 0.66%, and finds no evidence that AI reduces labour-income inequality. In a June interview he called the hyperscalers “extractive” and estimated that about 80% of AI discourse is “brainless”. Carl Benedikt Frey, in a secondary summary, notes US labour productivity of 1.8% annualised in the final quarter of 2025, and just 0.2% on the San Francisco Fed’s measure. And Goldman Sachs, in March: “We still do not find a meaningful relationship” between adoption and productivity across the economy, partly because so much AI equipment is imported.

I do not think we have to choose. My instinct, as ever, is Long-AND, not Short-OR. Brynjolfsson can be right that the long-run curve bends upward, and Acemoglu can be right that the gains, for now, are narrow, concentrated and unequal. The fifteen-fold spread is not noise. It is the economic signature of a transition whose winners and losers have not yet been sorted.

This is the terrain of Ghost GDP: The Long Night Before the Long Day: value that is unmistakable inside the task and hard to find in the national accounts. The question this essay asks is who is left standing outside during the long night.

III. Calm in aggregate, a squeeze at the bottom

Read only the aggregate data and you would conclude there is little to see. The Yale Budget Lab finds no broad shift in the labour market since ChatGPT, and AI-attributed layoffs were just 4.5% of US job cuts in 2025, which looks more like “AI-washing” than displacement. The Chicago Fed found that employment in the most AI-exposed occupations actually rose 4.05% between 2019 and 2024, and concluded that exposure “does not map mechanically onto job loss.”

Both pictures are true, and they reconcile cleanly once you look by age. ADP’s chief economist, Nela Richardson, makes exactly this point: the aggregate effect is modest, but the differences by career stage are large. Critics of the Canaries findings blame interest rates, tech over-hiring or remote work. Brynjolfsson’s reply is that construction, the most rate-sensitive sector, has low AI exposure, and that the pattern survives once the tech sector is removed.

I have a name for what this looks like from inside a firm: the Frozen Workforce. Nobody is marched out. Incumbents are kept, often retrained, sometimes promoted. The requisitions for graduates simply do not get approved. The organisation is frozen at its current shape, and the cost falls on people who never appear on its payroll, and therefore never appear in its layoff statistics.

Other signals point the same way, with caveats attached. A secondary digest of the AI Index 2026 reports employment of software developers aged 22 to 25 down nearly 20% from 2024. Anthropic’s January Economic Index reportedly found 45% of firms needing fewer entry-level hires, a figure that so far comes only from press coverage.

IV. The missing first rung

Why the young, and why codified work? David Autor and Neil Thompson’s “Expertise” paper offers the cleanest mechanism I know. When automation removes the inexpert tasks in a job, wages for the work that remains rise and employment falls. When it removes the expert tasks, wages fall and employment rises, because less experienced people can now get in. The effect of AI depends on which end of the job it eats.

Generative AI, so far, eats the bottom: the first draft, the reconciliation, the literature search, the code scaffold, the slide nobody wants to build. Those were never just tasks. They were the apprenticeship. They were how a 23 year old became a 35 year old worth paying for judgement.

Every senior professional I have worked with learned the trade that way. You did the unglamorous work, someone better than you marked it up, and over a few hundred repetitions you absorbed what no manual contains. Kwame Anthony Appiah argues that expertise is shifting from producing work to evaluating it. I agree. But you cannot evaluate work you have never produced. The ladder still has its upper rungs. What it is losing is the first one.

The Bengio-chaired International AI Safety Report 2026 lists the decline of early-career employment in the most exposed occupations among its systemic risks, beside a quieter finding: clinicians’ tumour detection fell about 6 percentage points without AI after months of AI-assisted colonoscopy. Deskilling at the top, de-hiring at the bottom. They are the same phenomenon seen from opposite ends of a career.

So the question I would put to every board is not how many roles AI can remove. It is this: how does a profession reproduce itself when its apprenticeship layer disappears?

An empty brass birdcage with its door open, hanging in a dim mine tunnel lit by an amber lamp and cold teal light — The canary does not fall silent. It is simply never brought in.

V. What Stanford’s teachers see

The economists measure this through payroll data. Three of the six Stanford leaders in the series see it from the front of a lecture hall.

Chiara Sabatti, a statistician, is confident that society will eventually learn to use AI well. Her worry is speed: that the short-term pace causes mistakes and painful consequences that are hard to recover from. Her next project is institutional. “AI is shaking up the world of higher education,” she says, and she wants to help redesign what a university is for.

Emmanuel Candès gives the most vivid account of what AI has done to research: “What once took months to set up and iterate now takes hours.” He calls himself an optimist, yet he asks the question every graduate is asking: if AI can outperform a Stanford student, what makes us human? His answer is an analogy I like. A car is faster than a runner, but people should not stop running.

Angèle Christin is the series’ dissenter on framing. She is “less worried about superintelligence and more concerned about the concrete harms”, and among those harms she names layoffs justified by AI efficiency. One question keeps her awake: how institutions can support young people without dismissing their anxieties as irrational “AI fear”.

That last point deserves weight, because the young are telling us what they think. Gallup found in July that 39% of Americans say AI does more harm than good, up from 31%, rising to 47% among 18 to 29 year olds. Secondary coverage of a separate Gallup youth survey reports Gen Z excitement falling from 36% to 22%, and anger rising from 22% to 31%. This is not irrational fear. It is a reasonable reading of the hiring data.

VI. The money, and who holds it

Behind the jobs story sits a financial one. The IMF’s October outlook noted that the US and China account for more than three-quarters of net demand for AI-equipment imports, and Kristalina Georgieva called AI’s labour effect “a slow moving tsunami, but a tsunami nonetheless.” The Fund’s Annual Report flags debt-financed AI investment and circular financing among a few firms as contagion risks.

The sums are vast: hyperscaler AI capex of about $650bn in 2026, and $121bn of related bond issuance in 2025, according to The National. Joseph Stiglitz calls it a bubble and compares the coming displacement to Depression-era farm labour. The physical bill is arriving too: the IEA found data-centre electricity demand rose 17% in 2025.

Here is the convergence I find most striking. Across every camp, the shared fear is market concentration: Acemoglu; Aghion, who reportedly warns that cloud and chip dominance could blunt the gains; Stiglitz; and Mariana Mazzucato, with her warning of “digital feudalism” through cloud gatekeepers. Put that beside the Canaries data and you can see the Mad Max branch of The Fork: gains captured upstream by a handful of firms, costs pushed downstream onto people who have not yet started their careers.

Acemoglu has sketched where that road could end. Graduate unemployment of 30 to 40%, he said, could threaten democracy. He offered it as a hypothetical, not a forecast, and I read it the same way. But when a Nobel laureate is willing to say it aloud, it is a warning worth heeding.

That is why “We Must Act Now” matters. The July statement, organised by Brynjolfsson, Agrawal, Korinek and Cunningham and signed by more than 200 economists and AI researchers including sixteen Nobel laureates, argues that AI’s economic transformation may outpace the Industrial Revolution and that policy cannot wait for certainty. Rob Reich, the sixth voice in Stanford’s series, is a signatory.

VII. Singapore’s version of the pattern

I will save Singapore for Part V, but the resemblance is too close to ignore. MOM told Parliament in May that PMETs in ICT, financial services and professional services face relatively higher retrenchment rates, which it reads as restructuring rather than falling demand. Retrenchments rose to 3,830 in the first quarter of 2026, falling on degree holders and PMETs. Those are precisely the codified-knowledge roles Autor’s framework predicts AI will thin first.

The IMF’s Asia-Pacific outlook found that employer demand fell 10% in AI-exposed jobs before stabilising, and concluded that AI is “reshaping work more than eliminating it”. I hope so. Singapore has the instruments: GRIT traineeships for fresh graduates, Career Conversion Programmes and redesigned SkillsFuture AI pathways. The open question is whether a traineeship can rebuild a rung that employers have decided they no longer need.

VIII. Where the published record lands

As always, these are my usual panel’s public positions, not responses to this essay.

Dario Amodei has been the bluntest builder on this. In The Adolescence of Technology, his January essay, he warned that about half of entry-level white-collar jobs could go within one to five years, and proposed progressive taxation of AI profits and redeploying rather than firing workers. AI, he wrote, will “test who we are as a species”. I think his timeline is aggressive. His remedy is right but incomplete: redeployment protects incumbents. It does nothing for the graduate who is never hired.

Yoshua Bengio‘s International AI Safety Report gives the issue institutional weight, listing early-career decline as a systemic risk alongside deskilling. Bengio is usually heard on loss of control. It is worth noticing that the scientific consensus document puts the entry-level squeeze in the same frame.

Erik Brynjolfsson is the optimist who built the alarm. He believes the harvest phase has begun, and his own data show who is not invited to the harvest. Hope about the aggregate, candour about the distribution: that is the posture I would recommend to any leadership team.

The sharpest pushback comes from Arvind Narayanan and Sayash Kapoor, who treat AI as “normal technology” whose diffusion takes decades. Kapoor criticises a debate that “expects societal impacts to occur at the speed of technological development”. Gary Marcus adds that AI is often a cover for layoffs, which fits the AI-washing data. Both critiques land. But neither explains a gap that widened from 15% to 19% in a year, and that survives once you strip out tech and rate-sensitive sectors. Slow diffusion across the economy is entirely compatible with a fast shock to the one cohort with no seniority to protect it.

IX. A prediction

By the end of 2028, the most sought-after graduate programmes in professional services, law and software will no longer be marketed as jobs. They will be marketed as apprenticeships in judgement: smaller cohorts, longer supervision, and work organised around reviewing, correcting and signing off on what AI produces. The firms that build them will grow the next generation of partners. The firms that keep the Frozen Workforce will discover, early in the 2030s, that they have nobody ready to promote.

I also expect the Canaries gap to widen past 20% before it narrows. Short-term turbulence first. The long-term abundance arrives only if we rebuild the rung.

X. The first-rung audit

One exercise this week. Pull your organisation’s graduate and entry-level hiring numbers for 2023, 2024, 2025 and 2026. Lay them beside your AI deployment roadmap. Then put three questions to your leadership team:

The task question. Which tasks did our juniors do in 2023 that AI now does?
The learning question. Where, specifically, will today’s 23 year olds learn the judgement our managers learned from those tasks?
The owner question. Who in our organisation is accountable for that answer?

If the third question produces silence, you have found your Frozen Workforce.

Next, in Part IV, Who Holds Custody, I turn from the economists to the Pope, the philosophers and the politicians, and to the question beneath all of this: who decides.

The Hopes and Fears series

Part I. Six Voices, One Turn: What Stanford’s AI Leaders Now Fear
Part II. The Composition Gap: Brilliant on Parts, Unreliable on Wholes
Part III. The Missing First Rung: Young Workers and the Fifteen-Fold Disagreement (you are here)
Part IV. Who Holds Custody: The Pope, the Philosophers and the Power Question
Part V. Fear Cannot Be Our Response: Singapore’s Hope-Led Bet

Sources

2 responses

Who Holds Custody: The Pope, the Philosophers and the Power Question

October 30, 2026

[…] Part III. The Missing First Rung: Young Workers and the Fifteen-Fold Disagreement […]

LikeLike

Fear Cannot Be Our Response: Singapore's Hope-Led Bet

October 30, 2026

[…] Part III. The Missing First Rung: Young Workers and the Fifteen-Fold Disagreement […]

LikeLike