There is a specific kind of corporate confession that only happens twice. Once when the IPO prospectus is being drafted and the underwriters circle the language about social impact, and once when the founder gets in front of a friendly audience and says it out loud so the press cycle clears before the roadshow starts.
On May 26, 2026, in a Commonwealth Bank of Australia CEO conversation in Sydney, OpenAI’s Sam Altman told CBA’s Matt Comyn that he was “pretty wrong” about AI’s economic impact. The specific reversal: “I’m delighted to be wrong about this. I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened.” Earlier this month, Anthropic’s Dario Amodei reframed his “AI could eliminate 50% of white-collar jobs” thesis from last year into a Jevons-paradox argument — “if you automate 90% of the job, then everyone does the 10% of the job,” and that 10% “expands to be 100% of what people do” while productivity goes up tenfold.
Both companies are reportedly preparing IPOs this year at roughly $1 trillion valuations each.
What Altman actually said
The phrasing matters. Altman did not say AI is not eliminating entry-level work. He said he expected more of it to have happened by now. The implied claim — that the rate is slower than predicted, not that the direction is reversed — is the one a roadshow lawyer can defend. He paired it with a personal anecdote: he tried delegating his Slack and email responses to AI, then started doing them manually again. “We really do care about our interactions with people. This thing is not something that I can imagine myself outsourcing to an AI anytime soon.”
This is a different sentence than “AI is not going to replace knowledge workers.” It is “AI is not going to replace the CEO of OpenAI.” The class of work the CEO of OpenAI does — relationship maintenance, judgment calls, public-trust performance — is exactly the class of work that an Anthropic-style “10% expands to 100%” reframe describes. The class of work he was originally talking about — entry-level white-collar tasks that look procedural — is exactly the class of work where Meta cut 8,000 and Intuit cut 3,000 and Cisco cut roughly 4,000 and Wix cut 1,000 this month alone. None of those press releases used the word “apocalypse.” They all used the word “AI.”
The Yale Budget Lab line that Altman is implicitly leaning on — that there has been no significant change in occupational mix or unemployment duration in high-AI-exposure jobs since late 2022 — is a measurement of stocks, not flows. A job category does not have to shrink for the entry pipeline into that category to be closing. A 22-year-old who would have been a junior copywriter at an agency in 2023 may now be working in food service while the agency’s headcount line stays flat because the senior copywriter who would have retired in 2026 is still there. That is consistent with the Yale data and also with the CBS News framing that layoffs are only part of the story — the hiring side is the rest of it.
What Amodei actually said
Amodei’s reframe is the more interesting one because it was more recent and because the underlying argument is sharper. In May 2025 he was the public face of “AI could wipe out half of entry-level white-collar work.” This month he is the public face of “Jevons paradox: cheaper computation leads to more computation, cheaper writing leads to more writing, lower cost per interaction does not mean fewer interactions.” He quoted Apollo’s Torsten Slok approvingly. He has the same view as Box’s Aaron Levie and Goldman Sachs’s David Solomon, who has been making the same argument since 2025 and just restated it in a NYT op-ed on May 22.
The Jevons frame is not wrong. It is also not what a pre-IPO Anthropic was saying twelve months ago. Twelve months ago, the company that built Claude was telling Congress that AI capability growth was an emergency. Today, the company that is about to ask the public market for a trillion dollars is telling Congress that the labor side will sort itself out through productivity gains. Both can be sincere. They cannot both be the message you want carried into the S-1.
What is being optimised
The pattern is now legible enough that anyone reading the IPO calendar can call the next move. When the regulatory risk is “Congress writes a Frontier AI Act because AI is going to wipe out the white-collar workforce,” the CEO most exposed to that risk is the AI lab CEO. The CEO whose business model is selling AI to the enterprise has the same exposure. Either the technology is dangerous enough to merit a moratorium, or it is productive enough to merit a $1T listing. The two narratives do not coexist in the same prospectus.
So the message gets recalibrated. The May 2025 version — “AI will eliminate 50% of entry-level white-collar jobs” — gets carbon-dated to a different regulatory moment. The May 2026 version — “I am delighted to be wrong, and the 10% expands to 100%” — gets timed to a roadshow.
The honest version, which neither CEO said in May, is the one Microsoft’s AI chief Mustafa Suleyman is still saying — AI will be able to automate most white-collar work within 18 months. Suleyman is not preparing for an IPO. Nvidia’s Jensen Huang is on the productivity side, but Nvidia already had its IPO. The companies that are about to list are the ones whose CEOs are now, suddenly, finding Solomon’s argument convincing.
What to watch
- OpenAI S-1 language. The “Risk Factors” section is where the position taken in Sydney has to either hold up or get re-disclosed. If the S-1 names AI-driven labor displacement as a material risk to revenue or regulation, Altman’s May 26 line is rhetorical. If it does not, the rhetoric is the position.
- Anthropic S-1 language. Same test, mirrored. The Jevons reframe either appears verbatim in the prospectus or it does not.
- The next AI-cited layoff. Some company will, in the next two weeks, announce a four-figure cut and credit AI in the press release. The interesting question is whether the press release cites Altman’s “delighted to be wrong” line back at him. Wix’s May 25 announcement was the latest before this; one of the next few will test whether the CFO class accepts the new vocabulary.
- The Yale Budget Lab Q3 update. The Yale tracker is the dataset both Altman and Solomon are leaning on. If the Q3 occupational-mix print starts to move, the May 26 framing has to move with it.
- Suleyman. Microsoft is not pre-IPO. If Suleyman holds his 18-month line through the rest of 2026, the gap between his statements and his peers’ will be the cleanest measurement of how much of this rhetoric is roadshow management.