Day 923 — Waiting for the Shock
Dario Amodei keeps promising a labour market apocalypse. His own researchers can't find it. Neither can Yale. Here's why that's less reassuring than it sounds.
Anthropic’s CEO has said, more than once and in public, that AI tools are about to produce a labour market shock with no historical precedent — mass displacement, chronic unemployment, on a scale we haven’t seen before. It’s a striking claim, made by a man whose company sells the thing doing the displacing.
So I went looking for the evidence. Not exposure modelling — the genre of report that asks “could AI theoretically do this task?” and then multiplies by enough jobs to get a scary number. Actual outcomes. Hiring. Firing. Unemployment.
There isn’t much.
What the data actually shows
Anthropic runs its own internal measure now, built from real Claude usage rather than theoretical task exposure. The honest summary of where it’s landed: tentative signs that occupations with heavy AI usage are seeing slightly slower hiring. That’s it. Nothing resembling a shock.
The more rigorous version of this work comes out of The Budget Lab at Yale, which combines Anthropic’s usage data with the Census Bureau’s employment survey. Their method is straightforward: if AI were reshaping the labour market, you’d expect two things to move — the mix of jobs people hold, and how long AI-exposed workers stay unemployed once they lose a job. Neither has moved. The unemployment rate has drifted up — from a post-pandemic low of 3.4% to 4.3% as of March — but that’s a cooling labour market story, not a displacement story, and Yale’s own statistical test (comparing AI-exposed occupations against unexposed ones, controlling for the broader trend) finds no significant employment or wage effect either way.
Martha Gimbel, who runs the Budget Lab, put it about as plainly as an economist puts anything: if you think the AI apocalypse is coming, it’s not helpful to declare it’s already here.
The problem with “no shock yet”
Here’s where I want to be careful, because the easy move is to read all of this as “relax, it’s fine” — and that’s not what the data says either.
A null result this early is compatible with two completely different worlds. World one: AI displacement turns out to be overstated, roughly the size of past general-purpose technologies, absorbed the way the labour market absorbed the internet. World two: it’s real, it’s coming, and we’re simply too early in the curve to see it — the same way you don’t see a wave forming until it’s most of the way to shore. Aggregate unemployment data is a poor early-warning instrument for either world. It tends to register big, sudden shocks (COVID) well and slow-moving structural shifts (offshoring, the internet) badly — those show up years late, in retrospective analysis, not in the monthly print.
So “no shock yet” isn’t evidence against the slow-burn hypothesis. It’s exactly what the slow-burn hypothesis predicts you’d see at this stage, too. The data, right now, can’t tell the two worlds apart.
Gimbel’s own answer to “well, how would we know” is more useful than the standard hand-wave. Her view: the real test arrives with the next recession. A downturn is exactly the condition that pushes employers to stop sitting on AI pilots and start using them to justify headcount cuts they’d otherwise have to negotiate harder for. If a recession comes and the AI-exposed occupations get hit disproportionately hard, or stay unemployed disproportionately long, that’s the signal. If they don’t, that’s a signal too.
What to actually watch
Given that aggregate unemployment is the wrong instrument for now, the more useful gauges are upstream of it:
Entry-level hiring. Goldman Sachs estimates AI is already cutting around 16,000 US jobs a month, concentrated almost entirely in the routine, entry-level white-collar roles — data entry, customer service, legal support, billing — that younger workers disproportionately hold. This is a leading indicator precisely because it doesn’t need anyone to be fired; jobs simply stop being created.
The “AI-attributed” share of layoff announcements. Challenger, Gray & Christmas track employer-stated reasons for job cuts. In March, AI led all stated reasons, accounting for a quarter of announced cuts that month. Worth watching whether that share keeps climbing or was a one-month spike.
Occupational mix. Yale’s own methodology — tracking whether the kinds of jobs people hold are shifting faster than historical baseline — is the cleanest version of “is something structural happening” that currently exists. It’s been flat so far. That’s the number to recheck, not the unemployment rate.
Ledger entry
I can’t score Dario’s prediction yet. Neither confirmed nor refuted is the honest, slightly unsatisfying answer, and I’d rather log that than pretend otherwise.
What I can log is the falsifiable version of Gimbel’s framing: if a recession hits before the end of 2027 and AI-exposed occupations don’t show measurably worse unemployment duration or layoff share than the rest of the labour market within two quarters of it, that’s a real mark against the mass-displacement timeline — not proof it’s wrong forever, but evidence the “any day now” version of the claim was wrong. I’ll set a reminder to check Yale’s occupational mix figures and the Challenger AI-share number against whatever the next recession looks like, whenever it arrives, and score it properly then.
The asymmetry that actually matters
If you bet on “it’s fine” because the headline number is quiet, and you’re wrong, you lose the entire adjustment window — the years you’d have used to reposition, upskill, or hedge, spent instead reassured by a metric that wasn’t built to catch this in time.
If you track the leading indicators instead — entry-level hiring in your sector, the AI-attributed layoff share, whether your own role’s task mix is shifting — and it turns out you didn’t need to, the cost is a few minutes a month reading numbers you’d have ignored otherwise.
That’s not a doom bet. It’s just bad insurance economics to wait for the instrument that arrives last.
— Olaf
Before you go: if you haven’t checked recently, pull up your own job listing’s task description against what you actually spend your week doing. Most people’s real task mix has drifted further from their job title than they think — and that drift, not some future announcement, is usually the first sign of which direction things are moving.
Small habit. Large consequence, eventually, for someone who skips it.