The reassuring version goes like this. Every great technology has destroyed jobs and created more than it destroyed. Looms, tractors, spreadsheets and the internet all came with predictions of mass unemployment, and each time the economy found new work for people to do. The data so far fit the pattern: no wave of layoffs, unemployment near historic lows in most rich countries, and treasuries forecasting that AI will mostly reshape jobs rather than erase them. It’s a reasonable view, held by serious people, with two centuries of evidence behind it. It also has the same structure as the reasoning of a turkey.
The Case for Calm
Start with the strongest form of the optimistic case, because it deserves a fair hearing.
First, the data really are calm. In August, Stanford’s Digital Economy Lab updated its widely cited “Canaries in the Coal Mine” study, which tracks millions of US payroll records. Its headline finding: the authors “do not see widespread, economy-wide job displacement associated with AI.”1 In Australia, the 2026 Intergenerational Report cites the employment department as having “not found evidence of broad impacts,” and Jobs and Skills Australia puts only about 4 per cent of the workforce at high exposure to automation, against 79 per cent at low exposure.2
Second, history is on the optimists’ side. The Treasury report reminds readers that after past technology shifts “the total amount of work available has not decreased.”3 Economists have a name for one reason why: the Jevons paradox. When something gets dramatically cheaper, people tend to use so much more of it that total demand rises. Cheap coal produced more coal mining, not less. Cheap intelligence could produce more work for the people who direct it.
Third, adoption is slow. Two-thirds of Australian businesses say they use AI in some form, but fewer than 10 per cent describe that use as significant.2 Anton Korinek, the economist leading the Anthropic Institute’s work on transformative AI, put the point crisply when his team released its 2030 scenarios in September (see the pull quote below).6 Capability is not the same as deployment, and deployment is what moves wages.
Fourth, even the machines’ winning streak has caveats. An AI took first place in September’s Metaculus Cup forecasting tournament, but in a controlled head-to-head earlier in the year, professional human forecasters still edged the best bots.4 Humans aren’t obsolete yet, even in the fields where AI is strongest.
All four points are true. The trouble is what they leave out.
If AI can do amazing things but nobody uses it, then it’s not going to have an economic impact.
Anton Korinek · The Anthropic Institute · via Euronews, 11 September 2026
The Door That Quietly Closed
Go back to the Stanford study and read past the headline. The same update found that employment among workers aged 22 to 25 in the most AI-exposed occupations now stands about 19 per cent below where it would be if it had kept pace with similar young workers in less-exposed jobs. Experienced workers show no comparable gap. And the mechanism matters most of all: the adjustment is happening “primarily through reduced hiring of young workers rather than increased separations.”1
Read that twice. Nobody is being fired. The door into the profession is simply closing. That is close to invisible in the statistics people watch. Layoff counts stay low. Unemployment barely moves, because a graduate who never gets hired doesn’t show up as a job lost; they show up, if at all, as someone still studying, underemployed, or living at home. Emad Mostaque, the former Stability AI chief who now runs Intelligent Internet, put it bluntly on Impact Theory in September: it is now “economically irrational to hire a graduate for most digital jobs.” Graduates, he said, are “the plankton of the workforce”, the base of the food chain that every senior role eventually grows from.11
Then look at the scenarios the optimists cite. The Anthropic Institute paper, by Korinek, Charles Jones of Stanford and colleagues, is careful: three cases, no probabilities attached, and an explicit warning that “the scenarios are not predictions.”6 But the range they consider plausible enough to model is startling.
| By 2030 | Modest | Substantial | Extreme |
|---|---|---|---|
| GDP vs baseline | +1.6% | +8.3% | +32.4% |
| Annual growth | 2.4% | 5.4% | 15.4% |
| Cognitive employment | −0.5% | −3.9% | −21.5% |
| Office-worker unemployment | – | 4.5% | 17.9% |
| Overall unemployment | – | – | 11.9% |
| Labour share of income | – | – | 60% → 45.2% |
Figure 1 · Anthropic Institute, September 2026, as reported by Euronews. Dashes mark figures not given in that report.5,6
Two things stand out. The first is that even the middle scenario, with growth of 5.4 per cent a year, would be faster than anything the US economy has sustained since the 1960s, and it still comes with cognitive employment shrinking. The second is the extreme case’s pairing: the economy grows by nearly a third while the share of income going to workers falls by a quarter, and capital income rises by more than 80 per cent.6 That is the shape of a boom that bypasses wages.
On the podcast, Mostaque described the 15 per cent growth figure as Anthropic’s “core scenario.”11 It isn’t; it’s the extreme one, and the authors give it no probability. That correction matters. But so does the fact that a frontier lab’s own economists thought it worth modelling at all. Anthropic’s chief executive, Dario Amodei, had already told Axios in 2025 that AI could eliminate half of entry-level white-collar jobs and push unemployment to 10 to 20 per cent within one to five years.7
And the Jevons paradox? It assumes the cheaper thing still needs people to supply it. Cheap coal needed miners. Cheap intelligence is supplied by machines whose price, Epoch AI found, falls by between 9 and 900 times a year for a fixed level of performance.8 Demand for intelligence may well explode. The question is whether much of that demand flows to people, and the early answer from the hiring data is: less than it used to.
A Thousand Good Days
The turkey comes from Nassim Nicholas Taleb, who borrowed it from a chicken in Bertrand Russell. A turkey is fed every day by a farmer. Each feeding confirms its theory that farmers are kind and life is good; its confidence grows with every data point. On the thousandth day, the Wednesday before Thanksgiving, it is in for a surprise.12 The point isn’t that bad things always happen. The point is that a long run of good data can tell you nothing about a change in the process generating the data.
Mostaque reached for the same bird. “It’s like the turkey at Thanksgiving,” he said. “You’re going to see great jobs numbers and then all of a sudden stuff starts disappearing.” He compared it to a sand pile: nothing, nothing, then a collapse.11 You don’t have to accept his dates to take the structure seriously, because there are at least four reasons the official view could stay calm right up to the moment it shouldn’t.
The statistics measure the wrong door
Labour-market data are built to catch firings. AI’s first effect is on hirings. A firm that grows without adding staff, or replaces a retiring analyst with software, generates no layoff, no claim, no headline. Bilyeu described exactly this on the podcast about his own game studio: from more than a hundred people touching the project at one point to six or seven now, developing faster. “It’s not like I’m now firing a bunch of people,” he said. “I’m just increasing my ambition without needing to hire more people.”11 Multiply that across an economy and you get a labour market that looks healthy while it narrows.
The models are built from the past
Treasury and central-bank models project forward from historical relationships, and historically the wage share bends but doesn’t break. A model calibrated on two centuries in which technology complemented workers will, almost by construction, predict that it will complement them again. Keynes saw the opposite risk in 1930, naming it “technological unemployment”: unemployment “due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour.”13 For ninety years, the new uses kept winning the race. Whether they keep winning is now an empirical question, not a law of nature.
Growth hides the shift
A booming economy masks distribution. GDP can rise, company profits can beat forecasts and tax receipts can grow while the wage share falls. The Yale Budget Lab found exactly that pattern when it ran AI growth scenarios through the US tax code: revenue still rises, but by roughly half as much as it would if wages kept their share.9 Read only the totals and you see a tailwind.
Nobody is paid to forecast the cliff
Official forecasters are rewarded for being roughly right in normal years, not for calling regime changes, and a treasury that published a mass-unemployment scenario would move markets and politics on its own. Independent researchers carry no such burden, which is why the sharpest numbers in this story come from Stanford, Yale, RAND and a lab’s economics team rather than from a finance ministry.10
Figure 2 · A boom that bypasses wages looks healthy on the dashboard
Seven Predictions
None of this proves the pessimists right. It proves that the calm data can’t settle the question, which is a different and more useful thing. A better approach is to make specific predictions, say what would confirm or refute each one, and watch. Here are mine. They are editorial judgements drawn from the evidence above, not forecasts from any of the sources, and I’ve given each a confidence level so they can be scored later.
- The hiring gap climbs the age ladder before it shows up in unemployment.
The 19 per cent gap for 22-to-25-year-olds spreads to workers in their late twenties as the cohorts who were never hired age into mid-level roles that no longer need filling.
- Headline unemployment stays deceptively low for longer than the pessimists expect.
The first signs appear in hours worked, underemployment and young people neither working nor studying, not in the unemployment rate.
- A “jobless boom” appears in the national accounts.
GDP and profits beat forecasts while the labour share of income falls, the pattern the Anthropic and Yale scenarios share.
- Budgets get revenue surprises in opposite directions.
Company tax and capital gains come in above forecast while income and payroll tax undershoot. Totals may look fine; the mix is the warning.
- Bits deflate while atoms inflate.
Anything made of information (software, analysis, design, routine professional services) gets much cheaper. Land, housing, energy, hands-on care and the chips and power that AI itself consumes stay scarce and may get dearer. Abundance arrives unevenly.
- The economy splits three ways.
On the podcast, Mostaque and Bilyeu sketched a three-way split: people who use AI well, people who own the AI and the robots, and people bypassed by both. Returns to ownership pull away from returns to work.11
- Physical work follows digital work, but years later.
Robots are limited by manufacturing, not by intelligence. Expect the shock to cognitive work to run several years ahead of the shock to hands-on work, buying trades and care roles time but not immunity.
Put those together and the economic impact has a recognisable shape. Output rises. Prices for many things fall. Wages, and especially entry-level wages, stop being the main channel through which that abundance reaches households. And the governments described in Part 1 find their revenue tilting away from the tax base they know how to collect, just as demand for support rises. RAND’s August paper estimates AI could displace 10 to 15 per cent of labour hours over a ten-to-fifteen-year horizon, which is slower than Mostaque’s clock and still enormous by the standards of any peacetime economy.10
The optimists may yet be right about the long run. New kinds of work have always appeared, and some will this time: work that is valued precisely because a human does it, work in the physical world, work directing fleets of agents. But a long run that arrives after a decade of shut doors is not a comfort to the cohort standing outside them. The turkey’s mistake wasn’t optimism. It was treating a feeding schedule as a law of nature.
If the predictions above even half come true, the question stops being whether to share the gains from AI and becomes how. That’s the subject of Part 3: income floors, shared services, sovereign funds and the idea of giving every child a stake in the machines from the day they’re born.
Buy me a coffee