Essay 003

The wrong end of the ladder

The people building AI have put their timelines in writing. Almost nobody has written down what happens to the rest of us. This is what the employment data already shows, who is responsible for it in Britain today, and what is being decided while we look the other way.

Published November 2026

The argument in five lines

1. There is no jobs apocalypse in the payroll data. Economy-wide employment is fine.
2. Entry-level employment in AI-exposed work is not fine. In the US it sits 19% below where it should be. The UK evidence points the same way.
3. The damage tracks one specific use of AI, automating rather than assisting, which makes it a choice rather than a fate.
4. It is invisible to the unemployment rate and to wage data, so the instruments government watches will not raise the alarm.
5. Nobody in Britain is legally responsible for any of it, and the places least able to cope are the places with the least capacity to see it coming.

So I went ahead and asked the machine

Post all the recent news articles on AI, I put a question to it plainly. In ten years, are you planning to put us out of action? Go on, be honest.

The machine told me it has no ambitions of its own in the way I do. But - a rather critical but - it then went on to tell me that the people building it are aiming at exactly that, and that they have put it in writing.

I have been sitting with that answer for a while. Not because it is frightening; it at least could be. Because it is the most candid thing anyone - or any data centre - has said to me about the next decade, and it came from the thing doing the displacing. The data centre.

So this essay is not a prediction. It is an audit. Three questions: what have the builders actually committed to on the record, what is already visible in the employment data, and what, if anything, are we/they/us/you doing about it.

One. What the builders have put on the record

The case for ten years does not rest on commentators. It rests on filings.

Anthropic's formal submission to the US Office of Science and Technology Policy put a date in writing in March 2025. The filing itself says powerful AI could emerge as soon as late 2026 or 2027. The company's own summary of it drops the shoulder hedge: we expect powerful AI systems will emerge in late 2026 or early 2027. By powerful they mean systems rivalling Nobel laureates across most disciplines. That is a regulatory filing, not a podcast, and we are now inside the window it named. In May 2026 its co-founder Jack Clark wrote in Import AI 455 on 4 May 2026 that he reluctantly puts a better than even chance, above 60%, on AI systems powerful enough to build their own successors arriving by the end of 2028. We used to call that parenting our next generation and now my laptop is claiming its spawning tomorrow as well. And on The Tonight Show in February 2025, Jimmy Fallon asked Bill Gates whether we would still need humans. Gates answered: Not for most things. Sure, some qualification arrived: some activities we will reserve for ourselves. Thanks Bill. He noted though, that making things, moving things and growing the food the majority of us eat will become, in simple terms, solved problems over the next decade (TheWrap).

Maybe go back to those statements as a market rather than as prophecy? Capital is being allocated on the assumption that human cognitive labour becomes optional inside a decade. Whether the assumption proves correct matters less than you had hoped, let alone think. Hiring decisions are being made on it now. At least big bets.

And note what is absent. There is no corresponding filing, from any of these firms, on what happens to the people whose work the systems absorb. The capability has a roadmap - well that is lovely to know but, the consequences?

Two. The wrong end of the ladder

Here is a potentially enlightening moment based on a common assumption gone awry. The displacement has already started sure, just not where I - or mostly anyone - is really looking.

Everyone is watching the wrong end of the ladder. We are braced for the senior people to go, in a visible wave, announced at a podium, with a lovely plant and a nicely gold plated nibbed pen. That is not what the data shows. The unemployment rate is unremarkable and the wage figures are a little 'ugh'. What has changed is who gets recruited, and the entire cost of that change is being charged to one narrow band of the workforce.

Three things follow from that and they set up the rest of this piece. The damage is concentrated almost entirely on people in their first three-to-five years of work. It is caused by a specific use of the technology rather than by the technology itself, which means, to be confronting for a moment, it was chosen. And it is invisible to every instrument a government routinely watches, which is why nobody has been made responsible for it, why the thing work was quietly carrying for us is being dismantled before anyone has costed it and finally, why the question of who ends up owning the output is being settled while we are all still looking at the wrong end of the ladder.

The source is Canaries in the Coal Mine? by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab, revised August 2026. It uses administrative payroll records from ADP, the US payroll processor, a balanced panel of 3.5 to 5 million full-time American workers observed every month from January 2021 to June 2026. Not survey responses, not job adverts: payroll. Occupations are sorted into five quintiles by how exposed they are to AI, using task-level exposure ratings built on GPT-4. We're on GPT-6 as at time of writing.

Start with the finding the authors chose to put first. Average employment across the sample rose about 6% between November 2022 and June 2026. In the most exposed quintile it rose about 4%. That is the whole economy-wide displacement story, and it is a non-event. Anyone telling you the jobs apocalypse has arrived is not reading payroll data. Neither, to be fair, are the authors claiming they are: fact one of six, in their words, is that they find no evidence of widespread, economy-wide job displacement (Brynjolfsson, Chandar and Chen, August 2026).

Now split the same data by age. Employment of 22 to 25 year olds in the two most exposed quintiles fell about 11% over that period. The same age group in the three least exposed quintiles grew about 10%. A divergence of 21 percentage points, which is the 19% shortfall the paper leads with.

The age gradient is the thing to look at. This is the change in employment in the most AI-exposed quintile relative to the least exposed, November 2022 to June 2026, with no controls applied:

Only the first row is statistically significant at the 1% level. The effect is not spread thinly across the workforce. It is sitting almost entirely on people in their first three years of work.

That is the what. The mechanism is more interesting, and this is where the essay earns its keep.

The authors split AI exposure into two kinds, using data on how people actually use AI: queries where the model is doing the task instead of the person, and queries where it is working alongside them. Automation and complementarity. Entered together in the same regression, only automation predicts the decline, at about −0.10 per standard deviation for 22 to 25 year olds. The coefficient shrinks steadily with every older age band. Complementarity runs the other way: positive and statistically significant for workers aged 41 to 49.

So I went back and asked it directly. Do you know, while you are working, whether you are working alongside us or instead of us?

No. I cannot tell from inside a conversation which one I am doing. The labels in that dataset are not me noticing at the time, they are applied afterwards by a model reading the transcripts. What I see is the request. Draft this, check this, summarise this: those look identical whether the person typing is a junior learning the shape of the work or a manager who has decided not to hire one. The difference that shows up in the regression lives in the intent of whoever is asking, and often they are not thinking about it either. The minus nought point one is not a property of me. It is a record of what we were asked to do.

I find that answer more unsettling than any forecast in this essay. The most consequential variable in the data is a decision taken thousands of times a day by people who do not experience themselves as deciding anything.

So the same technology, in the same occupation, subtracts from the young and adds to the experienced, depending on which way it is pointed. That is not a story about AI being good or bad. It is a story about a choice employers are making, mostly without noticing they are making it.

The authors then ask: why the young specifically? and their answer should worry anyone who has ever been trained by someone. AI substitutes most effectively for codified knowledge: the formal, documented, teachable material that education delivers. It complements tacit knowledge: the kind acquired through practice, mentorship and repeated exposure to real situations. Occupations heavy in codified knowledge show slower entry-level growth. Occupations heavy in tacit knowledge show faster growth for mid-career and senior workers.

Read that twice. If you don't mind. The thing being automated is precisely what a degree certifies and is worked towards. The thing being protected is precisely what only a job can teach you. And the only route to tacit knowledge runs through the entry-level position that the automation is removing. It is becoming a closed loop. We are dismantling the apprenticeship while congratulating ourselves on the productivity. The in-quarter results and EBITDA results.

Two further findings explain why almost nobody has noticed.

First, this is happening through hiring, not firing. Separation rates fell for exposed and unexposed young workers alike, and among the young they fell at least as much in the most exposed occupations. That is the opposite of what displacement looks like; nobody is being marched out. The door is simply not opening - at least revolving. Where is the entrance? There are no redundancy announcements to report, no affected staff to interview, no real news story. Just over there: consternation at planned data centre building.

Second, wages are not moving. The adjustment shows up in headcount - not in basic compensation - which barely differs by exposure or age. Labour markets normally warn you through price before they warn you through volume. This one is not doing that. So the two instruments a government watches most closely, the unemployment rate and average earnings, are the two instruments that will tell it nothing until the cohort is already lost. Raises a flag to me. Hopefully not a white one.

The pattern also varies in ways that look causal rather than coincidental. The decline is steepest in the states in US where AI adoption is heaviest, around 19% in the leading-adoption tier. By region it runs to about 21% in the West and 13% in the Northeast against 7% in the South and 9% in the Midwest. Women sit in more exposed occupations than men at every single age band, and young women show the steeper declines. Around 60% of occupations in the least exposed quintile saw early-career employment rise over this period, against about 30% in the most exposed.

Two honest caveats, because they will be the first things a previously unknown critic will reach for:

The ADP sample over-represents exposed occupations and larger firms. Compared against the American Community Survey over the window where both exist, the gap is far smaller, roughly 2 points against 13 in ADP. But in professional, information and financial services, the two sources agree closely, at around 23 points in ADP and 21 in the ACS. The direction holds everywhere. The magnitude is specific to this sample.

And all of it is American. ADP has no British equivalent in the public evidence base, and no UK institution publishes a monthly read on entry-level hiring broken down by age and occupation. That absence is not a footnote, it is part of the argument. The country with the stated ambition to be an AI superpower cannot currently tell you, in close to real time, what AI is doing to its own entry-level labour market.

What Britain does have points the same way. Bouke Klein Teeselink at King's College London analysed millions of UK job postings and LinkedIn profiles from 2021 to 2025, treating the release of ChatGPT in November 2022 as the turning point. Firms whose workforces are heavily exposed to large language model capabilities cut total employment by about 4.5%. The effect sat almost entirely on junior positions, which fell 5.8%. The disruption concentrated in the higher-wage segments of the market. Customer-facing roles that depend on actual interpersonal contact held up (King's College London).

One number in that study deserves a sentence of its own. Because the juniors went missing and the senior staff stayed, average pay inside those firms rose by more than £1,300. Read that against what I said a moment ago about wages giving no warning. It is worse than no warning. The wage data moved in the direction that looks like good news.

Treat this as corroboration rather than confirmation. Klein Teeselink works from job postings and professional profiles, not payroll records, and the Stanford authors note elsewhere that studies built on that kind of data can diverge from administrative sources. Two different countries, two different methods, the same shape of finding, and the British version arrives with a detail the American one does not: the headline pay figure improving while the thing underneath it breaks.

Everyone has been watching the top of the ladder, waiting for the dramatic moment when the senior people are replaced. Meanwhile the bottom rung is being quietly sawn off. You cannot be promoted from a job you were never hired into.

Three. The case against my own argument

I want to put the strongest objection in writing, because this field is full of people who will not.

The serious objection is interest rates, not AI. Young workers are always hit first in a slowdown, the job ladder is highly procyclical, and graduate hiring fell after the sharpest rate rises in decades. Make that case well and the canary explains itself. Zanna Iscenko and Fabien Curto Millet, both economists at Google, argue it properly in Looking for the Ladder, and they are not fools.

The Stanford authors went and tested it (February 2026 note). Using interest rate sensitivity estimates from Zens and co-authors, they find that AI-exposed occupations are on average less exposed to interest rates, not more. Construction is highly rate sensitive and barely AI exposed. Split the occupations into those above and below the median for rate exposure and the entry-level decline in the most AI-exposed jobs shows up in both halves. Their conclusion is that rate changes do not explain the disproportionate fall in entry-level hiring.

What they concede matters just as much, and I would rather be the one to tell you. Apply the most stringent controls, comparing workers inside the same firm at the same moment, and the AI-linked decline only becomes notable from 2024, not from late 2022. So part of the earlier drop was something else. They say plainly that they cannot establish causation, that AI is not the sole determinant of anything here, and that these are descriptive indicators rather than causal estimates.

So I am not going to tell you this is proven. I do not need to. If the cause is the business cycle, the rung grows back and I will write that essay. If it is not, it does not, and we will know inside three years. Planning only for the comfortable version of that is not a plan.

Four. It will not land evenly

Geography is not background here. In the American data the decline is steepest exactly where adoption is heaviest, around 19% in the leading-adoption states, and it runs to about 21% in the West and 13% in the Northeast against 7% in the South and 9% in the Midwest. Exposure and adoption travel together, and they travel to particular postcodes first.

Britain's version of that map is already drawn. GLA Economics reported in April 2026 that London's workforce is more exposed to generative AI than that of any other UK region, and made the point that this is not a weakness: London is exposed because its economy is thick with the knowledge-intensive work the technology is best at (Greater London Authority). Adoption follows the same line. Accenture found London accounts for around 70% of all AI-related vacancies in the UK, nearly six times as many as Manchester, Leeds, Newcastle and Liverpool combined. Survey work puts daily AI use among London workers at 54% against 36% nationally, falling to 34% in the North West and 31% in Yorkshire and the Humber.

Which makes London the country's leading indicator. Whatever is happening to entry-level hiring in exposed occupations will show up in the capital first, and the capital is also where the first rung is most expensive to lose.

Then comes the part that complicates the easy story. Analysis of Glasgow City Region found it less exposed than other UK core city regions, because a larger share of its work is teaching, health and the trades. That is the complementarity finding from the payroll data, reappearing at the scale of a place: work that runs on tacit knowledge and human presence is holding up. But inside that same region, West Dunbartonshire carries the highest share of jobs at significant risk, because of a concentration of administrative roles. Exposure is not a regional fact. It is a street-level one, and it tracks what people in a given place actually do all day.

Now set the response against the risk. London has a taskforce, a research programme, a £30m package and a mayor with the standing to demand devolved powers. Barnsley has an £800,000 challenge fund. West Dunbartonshire has whatever its council can find down the back of the sofa. The capacity to see this coming, and to do anything about it, is distributed in almost perfect inverse proportion to the places that will struggle most to absorb it. London is best equipped and best informed. It is also, on this evidence, the least likely to be caught out.

That is the strongest argument I know for working at the scale of a place rather than waiting for a national answer. The impact arrives locally, shaped by the specific mix of work a town happens to do. The response has to be built the same way.

Five. So who is responsible?

Ask the narrow version of the question. Today, in Britain, who is legally responsible for retraining a worker whose role has been automated?

Nobody. I cannot find a statutory transition duty, a levy earmarked for AI displacement, or any requirement on an employer to assess workforce impact before deploying a system that removes a job. The TUC is campaigning for exactly such a framework, which tells you it does not yet exist. The obligation is absent, so nobody is failing to meet it.

What exists instead is a patchwork of pilots running orders of magnitude below the scale of the thing being modelled.

Note that the London taskforce, chaired by Baroness Martha Lane-Fox and run with the Institute for the Future of Work, did not forecast an apocalypse. It said the opposite. The risk it named was 'not a sudden shock, but a quieter drift': fewer entry-level roles, weaker career ladders, widening inequality, falling hardest on women in clerical jobs paying below the London average. That is a description of the Stanford finding arriving in one city. An investment minister has separately floated some form of universal basic income. It is not policy.

This is the gap bootr.ai exists to work in. Not to lobby for a national programme that is years away, but to help individual places design the thing nobody has been made responsible for. If you run a council, a chamber, a large employer or a community organisation, you are currently the only layer with both the budget and the proximity to act. We would like to do that work with you.

Six. What the job was actually doing

Every policy response on offer treats a job as an income stream. Replace the income and the problem is solved. That is the error, and it is why retraining schemes keep underperforming against their own targets.

A first job is not a wage. It is the mechanism by which a young person acquires structure, a reason to leave the house, people who are not family, a place in a hierarchy they can climb, and the slow transfer of knowledge from someone twenty years older who shows them how it is done. That transfer has no other delivery system. We never built one, because we never had to.

We already have a natural experiment in what happens when people fall out of that structure, and it has nothing to do with AI. The government's own Keep Britain Working review finds 2.8 million working-age people economically inactive through ill health, up around 800,000 since before the pandemic, at an estimated cost of £212bn a year, close to 7% of GDP. More than one in five working-age adults are out of work and not looking.

The line in that review I cannot stop thinking about is this. A 22 year old who leaves work through ill health can lose over £1 million in lifetime earnings, and the review names the second loss alongside the money: purpose and connection. The state has already quantified what falling out of work costs a 22 year old. It has not yet connected that number to the 22 year olds who are not getting in.

So the question underneath this essay is not how we replace the wages. It is what we build to carry the things the wage was quietly carrying.

Seven. What is national pride for now?

The UK government's AI strategy describes this country as the nation of Babbage, Turing and Lovelace. That isn't a Wetherspoon in Oxford. It is a claim staked on a nation that was once genuinely inventive, may well still be in places, and now intends to make Britain an AI superpower. It is written in the language of national pride, and it may even be sincerely meant.

Hold it next to the employment data and the irony lands with all the grace of a difference engine grinding out a polynomial in 1822. The national project is to win a race to build the thing that is closing the entry route into professional work for the country's own young populace. And it keeps coming. All that ambition and pride poured into the engine, and not one ingenious thought spared for the exhaust fumes.

There is a harder version of the question. National pride has many times rested on comparative capability: our science, our industry, our arms, our institutions, our ingenuity. What is it for when the most capable thinking inside your borders is not produced by your citizens, is not owned by them, supplied by them, and in most cases is not even hosted in your country? A nation proud of hosting data centres? Being a chosen venue a soulless entry in the next chapter of Super Bowl XXWICCIXW.

I think pride has to migrate, and quickly, from what we can out-produce to what we refuse to let happen to people. A country that handles this transition decently will be more unusual, and harder to copy, than a country that trains a large model. That is a sovereignty worth having. Actually having it, rather than hosting it.

Eight. The cow in the field

Yuval Noah Harari coined the phrase for where this all ends. And not even if we're careful, just literally...learn what moo will mean to you soon enough. In an essay in February 2017 titled 'The rise of the useless class', he argued that we may create a class of people who are not merely unemployed but unemployable. He has been consistent since about what he means by it: useless from the viewpoint of the economic and political system, not in any moral sense.

A cow is fed, housed, pushed into pasture, medicated, milked and protected. It is also entirely surplus to its own arrangements. Nobody consults the cow. Asks it.

The financial mechanism that gets us there is not science fiction. Geoffrey Hinton has warned that AI will bring large unemployment alongside a large rise in profits, with the gains flowing to a small number of people rather than to society at large. Analysts at Citrini sketched the loop in more detail: displacement outrunning corporate adaptation, consumer spending contracting, and tax receipts falling away as productivity accrues to capital rather than households. US tech stocks fell in the days after it circulated (British Chambers of Commerce). People whose job is to price future risk did not find the thought experiment easy to kick to the curb. It wasn't even kicked.

Run it all forward like a Chinese robot on a 100m track and you still might not get robots pointing you towards a cabbage patch. You get something duller. A population maintained at subsistence by transfers it has no power to set, from an asset base it does not own, administered by systems it cannot inspect. Comfortable, possibly. Consulted, no.

But the analogy breaks at exactly the point that matters. A cow cannot vote, organise, own an asset, or refuse. We can do all four, today, while the ownership of this productive capacity is still being decided. The window in which that is true is the entire subject. The cow's condition was not imposed in a single moment. It was settled gradually, by people making reasonable decisions about efficiency, until one day it was simply the arrangement. Until it said 'moo'.

Back to the machine

I went back and asked it one more thing. Not whether it would replace us. Whether anyone had asked it to think about who ends up owning what it produces.

'It' - or I.T. - answered that the question is being settled now, in decisions connected to infrastructure contracts that will shape our future workspace and place.

There is no vote scheduled on whether AI abundance becomes common wealth or private property. There is a long sequence of unremarkable commercial decisions that will add up to an answer, and almost nobody involved thinks of themselves as answering it.

So the question bootr.ai is putting to anyone who will take it seriously and even get to this part in an essay - thank you - is not whether the machines are coming. The curtain is revealing the picture behind it. The question is the one we keep coming back to: who owns the intelligence, and therefore the abundance?

If you have a view, or a community, or a workforce, we would like to hear from you. We do not have all the answers. Together, we might find some.

Sources

The employment evidence

• Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, revised August 2026
• The same authors, Canaries, Interest Rates, and Timing, February 2026
• Zanna Iscenko and Fabien Curto Millet, Looking for the Ladder: Is AI Impacting Entry-Level Jobs?, Economic Innovation Group
• *Bouke Klein Teeselink, Generative AI and Labor Market Outcomes: Evidence from the United Kingdom, summarised by King's College London*
• *Revelio Labs on entry-level postings, reported by CNBC*

What the builders have said

• *Anthropic, recommendations to the Office of Science and Technology Policy, March 2025, and the submission itself*
• Jack Clark, Import AI 455, 4 May 2026
• *Bill Gates on The Tonight Show, February 2025, reported by TheWrap*
• *Geoffrey Hinton and the Citrini scenario, via the British Chambers of Commerce*

Where it lands

• GLA Economics, London's workforce exposure to generative artificial intelligence, April 2026
• *Accenture on the regional concentration of AI vacancies, reported by AI Magazine*
• Glasgow City Region analysis of AI exposure
• *Regional AI usage figures from Employment Hero research, reported by IT Brief UK*

The response

• Mayor of London, backing a £30m package on the AI and Jobs Taskforce recommendations, July 2026
• Barnsley Tech Town's £800,000 AI Upskilling Challenge Fund
• Keep Britain Working review, March 2026 update
• TUC, Building a Pro-Worker AI Innovation Strategy, August 2025

And

• Yuval Noah Harari, The rise of the useless class, 2017

Figures are current as at October 2026 and several of them move monthly.

← Back to essays