The AI jobs apocalypse is running late. That’s not a contrarian take from a tech skeptic — it’s the emerging consensus from the data, including from Anthropic itself. The cutting-edge AI lab that gave us Claude published a labor market analysis in March that found “no systematic increase in unemployment for highly exposed workers since late 2022.” Claude covers just 33% of all tasks in the computer and math category, despite theoretically being able to take over nearly all of them. Deployment “remains a fraction of what’s feasible.”
🔍 THE BOTTOM LINE: AI spending is going through the roof while productivity gains stay on the floor. The Nasdaq has fallen about 8% since its June peak. Seven in 10 Americans oppose building AI data centres in their area. The grand promise of immediate economic transformation is in trouble — not because AI isn’t improving, but because the economics may not add up at a price society will pay.
What the Data Actually Says
The Guardian’s analysis draws on a growing body of evidence that cuts against the “half of all entry-level jobs gone in five years” narrative that Anthropic CEO Dario Amodei promoted last year. The key data points:
- Labor productivity was slower in the first three years of the AI era than during the IT boom that began in the mid-1990s
- A recent NBER study found “despite strong substitution at the task level, overall employment effects are modest, as reduced demand in exposed occupations is offset by productivity-driven increases in labor demand at AI-adopting firms”
- Even Sam Altman has walked back the apocalyptic framing, saying in May he doesn’t think “we’re going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about” — a reversal we noted at the time
As MIT economist David Autor put it: “A lot of people have noticed that the world is not changing as fast as they predicted.”
The O-Ring Argument: Why “Almost Good Enough” Doesn’t Work
The Guardian introduces a useful frame from a recent NBER working paper: the O-ring theory, named after the rubber seal that failed on the Challenger space shuttle in 1986. A $2 part destroyed a multibillion-dollar spacecraft because everything else depended on it working perfectly.
The analogy for AI: as long as AI cannot perform every task perfectly, it increases the value of the remaining tasks. Depending on which tasks AI takes over, it could increase the value of high-skill workers relieved of the low-end part of their job, or increase opportunities for lower-skilled workers by handling the expert tasks. The “last mile” — the tasks AI can’t quite do — becomes more valuable, not less.
This matches what we’ve seen in practice. BCG’s analysis found AI reshaping rather than replacing jobs. The Dallas Fed showed entry-level positions shrinking while experienced-worker wages rose — AI takes the routine, humans keep the judgement.
The Impossible Economics
Even if AI could eventually do everything, the price tag may be more than society will bear. The International Energy Agency estimates data centre power demand will more than double by 2030 to about 945 terawatt-hours — more than the energy consumption of Japan. According to some estimates, AI infrastructure could consume 20-40% of GDP. Seven in 10 Americans already oppose building data centres in their area, driven partly by energy costs.
Nobel laureate Daron Acemoglu, quoted in the Guardian, noted that AI companies “are never going to make money” and are “losing hundreds of billions of dollars every year.” The depreciation problem compounds this: new models overtake those developed just months ago, meaning the capital investment in any given generation of AI infrastructure has a brutally short useful life.
This is the Solow productivity paradox restated for a new era. Robert Solow quipped in the early years of the computer revolution: “You can see the computer age everywhere but in the productivity statistics.” It took about 10 years for businesses to reorganise around the technology before computers showed up in the stats. AI may follow the same arc. But the scale of investment required before the payoff is larger, and the politics are worse.
What Insiders Still Believe
The sceptics have not won the argument. Amodei’s one-to-five year window still has four years to run. The Federal Reserve reports AI adoption is expanding fast across businesses. Autor himself acknowledges AI is getting better and shows no sign of hitting a ceiling. “Skepticism about the stochastic parrot is behind us,” he said.
Elon Musk has not budged from his position that “AI+Robots will be able to do everything, resulting in universal high income. Work will be optional.” The insiders, as Acemoglu noted, “still believe artificial general intelligence is around the corner.”
The difference now is that the claim has shifted from “this is happening now” to “this is happening eventually.” That’s a weaker claim. It’s also less useful for selling things — whether products, stock, or policy agendas. The Nasdaq’s 8% drop from its June peak suggests markets are starting to price in the gap between promise and delivery.
NZ Angle: A Breathing Space, Not a Holiday
For New Zealand, the “apocalypse is late” story is not a reason to stop preparing. If anything, the breathing space is a chance to get the policy framework right — skills training, social safety nets, data governance — before the next acceleration. The Dallas Fed data on entry-level jobs shrinking while experienced wages rise is a NZ-relevant signal: the labour market distortion is real even if mass unemployment is not. NZ’s tight labour market and skills shortages may actually cushion the transition better than larger economies with more structural unemployment.
❓ FAQ
Does this mean AI won’t replace any jobs? No. The data shows task-level substitution is happening — AI is taking over specific work within jobs. The key finding is that overall employment effects are modest because productivity gains at AI-adopting firms offset reduced demand in exposed occupations. The pattern is reshaping, not wholesale replacement.
Why does Anthropic’s own data contradict its CEO’s claims? Dario Amodei’s predictions about half of entry-level jobs disappearing are forward-looking forecasts. Anthropic’s labor market report measures current data. The two are not necessarily contradictory — Amodei could be right about the future while the current data shows no displacement yet. But four years into his five-year window, the evidence is not trending toward his prediction.
What is the O-ring theory and why does it matter? Named after the rubber seal that failed on the Challenger space shuttle, it argues that when a system has many interdependent parts, the weakest link matters most. Applied to AI: as long as AI can’t do every task perfectly, the remaining human-performed tasks become more valuable, not less. This is why AI adoption can increase rather than decrease demand for certain workers.
Is the AI investment bubble going to burst? Acemoglu argues AI companies are losing hundreds of billions of dollars annually. The Nasdaq has dropped 8% from its June peak. But adoption is still expanding and models are still improving. The more likely scenario is a gradual repricing of expectations — investors demanding more evidence of returns before funding the next wave — rather than a sudden crash.
🔍 THE BOTTOM LINE
The AI jobs apocalypse narrative is not dead, but it is late, and lateness is changing it. The claim has shifted from “this is happening now” to “this is happening eventually” — a weaker claim that is harder to monetise and harder to scare people with. The data, including from AI’s own champions, shows modest employment effects so far. The economics — data centre energy demand, depreciation rates, political opposition — are harder than the technorati expected. None of this means AI has plateaued. It means the gap between what AI can do and what it makes economic sense to deploy at scale is wider than anyone selling the revolution wants to admit.
