Every month a single number is released, and every month it says the labour market is fine. It is the most trusted figure in economic life and, for the thing we most need to watch, it has gone blind. The interesting question is not whether unemployment is low. It is what a low unemployment rate is now capable of concealing.
I. The Number That Sees Nothing
On 5 March 2026, Anthropic’s economists published a study of how artificial intelligence is moving through the labour market, and buried in it is the most quietly damning sentence of the year. Using a new exposure measure built from actual usage rather than guesswork, they report that there is, as they put it, ‘no impact on unemployment rates’ for the occupations most exposed to AI. Reassuring, until you read the next finding in the same paper: job-finding for workers aged twenty-two to twenty-five entering those same exposed occupations has fallen by roughly fourteen per cent since the arrival of general chat models. The authors are scrupulous — they call the estimate barely significant, and I will hold them to that scruple later. But hold the shape of it now. The headline is calm. The entrance is closing.
Begin here, on a frontier lab’s own economists reporting the number that undercuts the reassurance everyone else takes from the headline, because it is the cleanest possible demonstration of the problem. This is not a critic alleging that the unemployment rate is broken. It is the people best placed to measure the effect telling you that the aggregate shows nothing while the young are turned away — and that both things are true at once.
The most trusted number in economics is not lying. It is looking in the wrong place, calmly, on schedule.
II. Stock and Flow
Why can one number hold both facts without flinching? Because of what the unemployment rate actually measures, which is not what most people who cite it believe. It is a stock: the share of people who want work, are looking, and do not have it, counted at a moment. It was engineered, in the middle of the last century, to detect a particular event — the dismissal, the lay-off, the factory gate closing on people who had been inside it. It counts the newly ejected.
AI displacement in this cycle is not, mostly, an ejection. It is a non-admission. The junior analyst role that is simply not opened, the graduate cohort a firm decides it no longer needs because the work the cohort used to do is now done by a model with a senior reviewing — these people are not dismissed, because they were never hired. They do not appear in the stock of the unemployed as a lay-off; they appear, if at all, as a slow thinning of the flow into work, spread across a category the headline number never disaggregates. A rate built to register the closing of a gate is structurally blind to a gate that is quietly never opened in the first place. The event it was designed to catch is not the event that is happening.
This is not a flaw the number can be scolded out of. It is doing precisely what it was built to do. We are simply pointing an instrument calibrated for dismissal at an economy practising non-admission, and then treating the silence as good news.
A rate built to count the dismissed will never see the door that was quietly never opened.
III. Same Firm, Opposite Directions
If this were merely the business cycle, the young and the old would move together. They do not. The sharpest evidence comes from Brynjolfsson, Chandar and Chen, whose 2025 study — they called it, aptly, canaries in the coal mine — used payroll records rather than surveys to track who is actually on the books. Their finding, updated with a further year of data in the Stanford AI Index for 2026, is stark: employment of software developers aged twenty-two to twenty-five has fallen by close to twenty per cent since its 2024 peak. Over the same period, at the same firms, developers past thirty saw their numbers rise by something between six and twelve per cent.
Read that pair again, because a single figure can be explained away and a divergence cannot. Same occupation, same employers, same window — the young cohort contracting sharply while the experienced cohort expands. A downturn does not do that; a downturn takes the juniors and the seniors together. What does that is substitution at the bottom of the ladder: the model now does the work the twenty-three-year-old was hired to learn on, supervised by the thirty-eight-year-old whose judgement the model cannot yet replace. The firm is not shrinking. It is changing its shape, and the headline unemployment rate, which sees neither age nor occupation nor the shape of a firm, records the whole manoeuvre as nothing at all.
The economy did not lose the young. It changed shape around them, and the aggregate called it calm.
IV. The Instrument We Actually Need
The remedy is not to distrust numbers; it is to measure the thing that is actually moving. What the two studies have in common is the variable that makes them legible: not the stock of the unemployed but the flow into work, cut by age and by exposure. That is the instrument a finance ministry steering this transition actually needs — a monthly job-finding rate, disaggregated by age band and by how exposed the destination occupation is to AI. It would have shown the fourteen per cent and the twenty per cent in real time, on the same chart, while the headline rate sat flat.
None of the data required is exotic. Payroll systems already hold it; the exposure measures already exist; the statistical agencies already run the surveys. What is missing is the decision to publish the cut that would embarrass the reassuring number. That is a choice, not a technical limit, and choosing not to make it is itself a policy — the policy of preferring a calm dashboard to an accurate one.
We have the data to see the door closing. What we lack is the decision to point a light at it.
V. What This Essay Cannot Yet Prove
Let me mark the limits of my own argument before my critics do, because an essay that overstates its evidence deserves to lose. The fourteen per cent is, in the authors’ own word, barely significant. The twenty per cent is real and on payroll data, but a single occupation and a short window are not a civilisation. Attribution specifically to AI, as opposed to a post-pandemic hiring correction or a higher-rate environment squeezing junior budgets first, is not settled, and anyone who tells you it is settled is selling something.
So I am not claiming the catastrophe is proven. I am claiming something narrower and, I think, harder to escape: that if the effect is real, the headline unemployment rate is the one instrument guaranteed not to show it until it is far too late to respond, and that we are currently reassuring ourselves with exactly that instrument. The argument rests on the mechanism, not the coefficient.
And I should declare my own stake, because I hold others to it. I sell no labour statistic and this essay favours no product of mine. But I have argued for years that our economic instruments are measuring the wrong quantities — I have given that argument a name and staked reputation on it — so I am not a neutral party to a thesis that says the instruments are blind. Discount accordingly. Then weigh the discount against the fact that the numbers here are not mine; they are Anthropic’s and Stanford’s, and they were produced by people with no stake in my framing at all.
I would rather be early on a real thing than exact about a comfortable one.
VI. The Case Against This Essay
Four objections, and the third forces the title itself to give ground.
The instrument-already-exists objection. Statistical agencies do not publish only one number; the labour-force surveys already carry youth unemployment, job-finding rates, and age breakdowns. The essay stages the aggregate as blind when the sighted numbers are a click away. Answer, partial. Granted: the data exists and I overstated the blindness. But existence is not attention, and no finance ministry steers by the youth-by-exposure job-finding cut, because it is not the number that moves markets or headlines. The argument is about which instrument governs, not which is technically available — and I should have said so rather than implying the number cannot be computed.
The attribution objection. The essay concedes AI-specific causation is unsettled, then demands a new official series on the strength of it. If the junior contraction is a rate-driven budget squeeze, publishing an alarm-coloured job-finding series would cry wolf. Answer, partial. The reply is that the proposed instrument measures entry-path suppression whatever its cause — that is a feature, an early-warning gauge that does not need to know the culprit to be worth watching. But concede the honest cost: it would sometimes fire on non-AI causes, and a gauge that cannot attribute can be politicised.
The ‘retire’ objection. The title says retire; the evidence supports, at most, demote. One occupation and a barely-significant fourteen per cent cannot justify decommissioning the most battle-tested statistic in macroeconomics. Conceded. The title is deliberately provocative and the provocation outran the argument. The defensible claim is not that the unemployment rate should be retired — it should not; it remains the right instrument for the event it measures — but that it should be dethroned as the primary indicator of this transition and paired with an exposure-graded job-finding series that carries equal official weight. I reduce the thesis to that, and the earlier sections should be read through this reduction.
The interest objection. The author has a standing intellectual stake in ‘the instruments are measuring the wrong thing’. Answer. Declared in §V; the evidence is third-party.
A provocation in a title is a debt the argument has to pay, and this one paid at a discount.
VII. Coda: The Dial
Return to the number that sees nothing. Next month it will be released again, and it will again say the labour market is fine, and it may even be right about the thing it measures — the dismissed are not, yet, being ejected in unusual numbers. But a dial can be accurate and useless at the same time, if it is reading the wrong quantity, and the quantity that matters now is not how many were pushed out but how many were never let in. We are steering a civilisational transition by an instrument engineered, in a different century, to stay calm through exactly this manoeuvre. Keep the dial; it still reads the event it was built for. But take it off the top of the dashboard, and bolt beside it the one gauge that reads the door — job-finding, by age, by exposure, every month, with equal official weight. The reassurance the old number offers is real. Left alone at the centre of the instrument panel, it is also the most dangerous thing on it.
We are steering by the one dial built to read steady while the road quietly runs out.
Sources
- Anthropic, Labor market impacts of AI: A new measure and early evidence, 5 March 2026.
- Brynjolfsson, E., Chandar, B. & Chen, R., Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, August 2025.
- Stanford Institute for Human-Centered AI, AI Index Report 2026.
Provenance note: the widely repeated new-graduate unemployment figure of 5.7 per cent, the ‘entry-level postings down 35 per cent’ and UK ‘−38 per cent’ claims circulating alongside these studies are reported at second hand and are held out of the argument above until their BLS, ONS and Budget Lab primaries are located. They shape the reading; they do not support a sentence.
Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol 
Leave a comment