Sam Altman told a Sydney audience on May 26 that he was “pretty wrong” about one of his own most-quoted predictions. “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened,” the OpenAI chief executive said, speaking alongside Commonwealth Bank of Australia CEO Matt Comyn. “I’m delighted to be wrong about this.”
Dario Amodei has made a similar turn. The Anthropic chief executive spent 2025 warning that AI could wipe out up to half of entry-level white-collar jobs within five years and push unemployment toward 10 to 20 percent, and calling on the industry to stop “sugar-coating” that outcome. In recent remarks he has shifted to describing AI as a productivity multiplier rather than a replacement engine, framing the work that survives automation as work that gets done by fewer people doing more.
What is actually new
Both reversals landed within days of each other in late May, and both leaders offered a version of the same explanation: adoption has been real but slower and messier than either predicted, and the jobs that disappear are not disappearing in the clean, sudden way the earlier warnings implied. Altman has since argued that AI functions as a “convenient” excuse companies reach for when announcing layoffs that were coming for other reasons anyway, a distinction worth sitting with the next time a earnings call blames headcount cuts on AI efficiency.
Amodei’s reframing leans on what he calls a Jevons paradox effect: automate 90 percent of a job and the person left holding the remaining 10 percent does not become redundant, they become roughly ten times more productive at the part of the job that still needs a human. It is a more measured claim than a 2030 unemployment forecast, and notably, both men made their walk-backs public in the run-up to periods when their companies are widely expected to prepare for public offerings, a timing coincidence outlets including Fortune and Forbes have pointed out explicitly.
What this means for hiring and workforce planning
For business leaders, the useful signal is not that the risk is gone, it is that the near-term forecast changed and the far-term one did not. Amodei has continued, in separate policy writing, to argue for real structural risk to employment over a longer horizon even while softening the immediate timeline. That is a more defensible planning posture than either extreme: budget for AI to change roles gradually rather than eliminate categories overnight, and be skeptical of any layoff announcement that reaches for AI as the explanation without showing its work. We saw exactly that ambiguity play out in Microsoft’s 4,800 job cuts, where the AI-efficiency narrative and the ordinary cost-cutting narrative were both plausible and neither was fully confirmed.
It also lines up with a pattern we have tracked on the revenue side: AI-native companies are scaling headcount-light and revenue-heavy, the dynamic behind AI startups now hitting $100 million in revenue in eighteen months, not seven years. Fewer people producing more revenue faster is the same phenomenon Amodei is describing, just measured from the company side instead of the labor market side. The services giants are testing the same thesis from the other direction: TCS plans up to 8,900 engineers embedded inside clients rather than cut, a bet that AI reshapes outsourcing instead of ending it.
What to watch
Watch how this plays out through the next earnings season, when any company citing AI for layoffs will now face more scrutiny given Altman’s own admission that the excuse is often convenient rather than accurate. And watch whether Amodei’s longer-horizon warnings resurface with more force once Anthropic’s own IPO timeline, whenever it lands, is no longer a live consideration shaping what its CEO says in public. Both companies have since confirmed confidential S-1 submissions; our IPO watch separates what is confirmed from what is speculation.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
