The Great Walk-Back

Why AI's loudest doomsayers softened their job-loss warnings — and whether we should believe the new story

I. Two Prophets Change Their Story

In May 2025, Dario Amodei, the chief executive of Anthropic, told Axios that AI could eliminate half of all entry-level white-collar jobs within five years and push unemployment to between 10 and 20 per cent.^[1] A month later, Sam Altman of OpenAI was warning that entry-level roles faced serious risk of elimination. The two most influential voices in artificial intelligence agreed: the disruption would be fast, deep, and brutal.

Twelve months later, both men have changed their story.

On 26 May 2026, on stage in Sydney with the chief executive of the Commonwealth Bank of Australia, Altman said he had been “pretty wrong” about AI's economic impact: “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened. I'm delighted to be wrong about this.”^[3]

The same month, at a financial-services briefing in Manhattan, Amodei reached for a new intellectual framework — the Jevons Paradox: “If you automate 90% of the job, then everyone does the 10% of the job. And the 10% kind of expands to be 100% of what people do.”^[4] Automation, in the new telling, is a productivity multiplier, not a job destroyer.

Days after that briefing, on 1 June 2026, Anthropic filed confidentially for a stock-market listing. OpenAI filed the same month, targeting a valuation approaching $1 trillion.^[2]

You do not need to be a conspiracy theorist to notice the sequence. You just need a calendar.

II. The Timeline Is the Tell

It is worth being precise here, because the defence of both men will be that views change as evidence changes — gradually, honestly, in public. That is not quite what happened.

As late as 27 January 2026, Amodei was still warning that AI could cause “unusually painful” disruption to jobs.^[5] In February, Forbes was reporting that he had doubled down on the warning.^[6] The Jevons conversion arrived in May. The IPO paperwork arrived in June.

Three months from “unusually painful” to “productivity multiplier” is not the pace of scientific revision. It looks a lot more like the pace of a roadshow. But — and this is where honesty demands we slow down — it is also true that by May, the data really had refused to cooperate with the apocalypse.

III. The Case That the Doomsayers Were Simply Wrong

The Budget Lab at Yale has been tracking the American labour market since ChatGPT launched in November 2022. Its finding, updated through March 2026: no meaningful change in unemployment for workers in AI-exposed occupations, and no significant shift in the occupational mix of the economy — the metric that moved sharply in previous technological transitions.^[7]

Anthropic's own labour-market research, published in March 2026, reached the same headline conclusion by a different route: measuring what Claude is actually used for rather than what it could theoretically do, its economists found no systematic increase in unemployment among highly exposed workers since late 2022.^[8] PwC's global Jobs Barometer, built on nearly a billion job advertisements, went further: AI appears to be making workers more valuable, not less, even in the most exposed sectors.

So when Altman says he is delighted to have been wrong, that is a defensible reading of the evidence. If the walk-back stopped there, the generous interpretation would win.

But the same data carries a darker footnote. Anthropic's researchers found suggestive evidence that job-entry rates for 22-to-25-year-olds in exposed occupations have fallen by around 14 per cent — the entry-level rung creaking first, exactly where Amodei originally pointed.^[8] Tech layoffs passed 115,000 by mid-May 2026 — nearly matching the whole of 2025 in under five months.^[3] Bloomberg reported in July that technology and finance are shedding roughly 28,000 jobs a month, with AI repeatedly cited as a driver.^[9] TechCrunch now maintains a running list of employers who name-check AI in their layoff announcements.^[10]

And then there is the fog machine. A Resume.org survey of 1,000 US hiring managers found that 59 per cent admit to emphasising AI in layoff announcements because it plays better with stakeholders than admitting financial strain — while only 9 per cent say AI has actually replaced roles outright.^[11] Corporate narrative, in other words, is unreliable in both directions: ordinary companies over-attribute layoffs to AI when it flatters them, and AI companies under-attribute when it flatters them. Believe the payroll data, not the press releases.

IV. Follow the Money

Now for the other argument — the one currently circulating as a “conspiracy theory”, and which deserves a fairer hearing than that label suggests.

Consider who actually buys a trillion-dollar IPO. Index funds. Pension funds. Sovereign wealth funds. The handful of asset managers who already own a slice of everything else — the banks, insurers, consultancies, staffing firms and professional-services giants whose payrolls AI is supposed to hollow out. Pitching them means, in effect, asking the owners of the old economy to fund the thing you have spent a year promising will burn it down.

Vijay Vijayasankar, an analyst at Genpact, made the observation bluntly: “Labor-based industries all took a serious hit in the public markets after Sam and Dario predicted the apocalypse of white-collar jobs” — and both companies, he noted, now sell services themselves.^[11] MIT economist David Autor put it more dryly: the industry's leaders “may have realized it was simply bad business to say that your great new product will destroy the economy.”^[2]

It is not only investors. Enterprise buyers would rather sign off on “augmentation” than on a machine for making their colleagues redundant. Regulators are watching. The political cost of the bloodbath framing rises every month it stays in the headlines.

Is there a smoking gun — an investor on record demanding a softer message? No. Are there motives beyond the IPO? Certainly: regulatory heat, political exposure, the social cost of being the face of the apocalypse. And Amodei, to his credit, hedged in the very same speech: “AI is moving faster than all these previous technologies… it's possible you get these weird behaviors and this big disruption.”^[4]

But here is the thing about this particular “conspiracy”: it does not need to be one. Nobody has to pick up a phone. When the people you need money from are also the people your product threatens, the softer story gets rewarded automatically — in valuation, in coverage, in regulatory goodwill. Founders respond to that gradient the way water responds to a slope. Incentive alignment does everything a conspiracy would, without a single meeting.

V. The Honest Verdict

Both things are true.

The 2025 apocalypse predictions were wrong — at least on timing. The independent data (Yale's) and the interested data (Anthropic's own) agree: mass displacement has not shown up in the unemployment figures. Anyone who repeated the “half of all white-collar jobs” line owes their audience the correction.

And the walk-back is commercially convenient in ways its authors have not acknowledged. The narrative turned precisely when the incentives turned, faster than the data alone would justify, delivered by men about to ask the public markets for more money than any companies in history.

The practical conclusion is to stop outsourcing your view of the future to either version of these men. They were doing narrative work in 2025 — urgency recruits talent, alarms regulators into engagement, and commands attention. They are doing narrative work now. Underneath both performances, the evidence has been quietly consistent: significant task-level disruption, concentrated in screen-based cognitive work, arriving more slowly than the hype — with the entry-level rungs creaking first.

VI. What This Means for the Belonging Economy

For operators in fitness, leisure and wellbeing, three things follow.

First, the exposure maps never changed. Every study cited above — Anthropic's, Yale's, PwC's — draws the same shape: the highest AI exposure sits in screen-based cognitive work, and the lowest in embodied, face-to-face, trust-dependent work. Coaching, instructing, community operations and floor-level service sit at the resilient end of every chart. That was true during the bloodbath narrative and it is true during the Jevons narrative. Your structural position has not moved; only the marketing around it has.

Second, the entry-level signal is the one to watch. The 14 per cent decline in job entry for 22-to-25-year-olds in exposed occupations is the most robust early warning in the data. If the office stops hiring young people, a generation still needs somewhere to start. A sector that can offer genuine career paths beginning on the gym floor — not in the back office — is positioned to absorb ambitious talent the spreadsheet economy no longer trains.

Third, build strategy on structure, not soundbites. If you tore up your plan over the 2025 doom, or if you relax because of the 2026 walk-back, you have made the same mistake twice: mistaking a chief executive's fundraising narrative for a forecast. Watch entry-level hiring, watch payroll data in exposed sectors, watch your own local labour market — and see our companion piece, The Jobs That Aren't Coming Back, for the full data picture.

The men who told you the apocalypse was coming and the men now telling you it has been cancelled are the same men, and both statements moved in step with their interests. The data has been steadier than its narrators. Read it directly — and build for the decade it actually describes.

Sources

Data and statistics cited are sourced from third-party reports and correct at time of publication. Figures may have been updated since. This is not financial or legal advice.