The layoffs are keeping their jobs

For about eighteen months, the CEOs of the largest AI labs told the public that their technology would eliminate a great many white-collar jobs. In roughly the last month, several of them have stopped saying that. The workers displaced under the first story are still displaced. That is the piece.

The Wall Street Journal's front page on Monday collected the reversal in one place.¹ Sam Altman, on stage at a conference, allowed that OpenAI had been "roughly right on technological predictions and pretty wrong on the social and economic implications." Dario Amodei, who last year floated the possibility that AI could wipe out half of entry-level white-collar work and push unemployment to somewhere between ten and twenty percent, has swapped the wipeout register for a "human flourishing" one.² Andy Jassy and Jeff Bezos, who spent 2025 telling shareholders that AI would let Amazon do more with fewer people, are now emphasising productivity and augmentation. The WSJ's framing is that the industry has "flipped."

I want to take the most generous reading of this seriously before I say what I actually think.

The steelman is real, and I will not flatten it. The aggregate US labour market has not, so far, produced the mass-displacement numbers the earlier framing implied. Headline unemployment sits inside historical norms. Some of the more careful economic work suggests the AI effect on employment is showing up as hiring freezes, thinner entry-level rungs, and sectoral churn, not as a headline unemployment shock.³ On that reading, Altman is doing something honourable: he made a public prediction, the evidence came in softer than the prediction, and he is updating in public. If that is what happened, it should be praised, not sneered at. Public figures who correct themselves when the data moves are rarer than public figures who do not.

I want to say clearly: some of the walk-back probably is that. Belief-updating on eighteen months of labour data is a legitimate move, and I do not want to write a piece that treats every executive revision as bad faith.

But the steelman does not carry the whole story, and here is where I part company with it. Three things are true at the same time, and the reassuring frame only picks up one of them.

The first is that aggregate labour-market data hides sectoral concentration. "Unemployment is fine" is a claim about the average. The people who lost customer-service roles at firms that publicly credited AI for the restructuring are not consoled by the average. Copywriters, junior paralegals, first-line support staff, coding bootcamp graduates staring at a shrunken entry-level market — these groups are absorbing something real, and it does not appear in the headline number because the headline number is not designed to see it. When Amodei's original prediction is judged against BLS aggregates, it looks alarmist. When it is judged against the specific occupational categories he named, the picture is less flattering to the walk-back.

The second is that CEO messaging and enterprise deployment are two different systems. Sam Altman does not sign the procurement contracts. The Fortune 500 procurement officer choosing between a Copilot enterprise seat and a headcount does not consult the transcript of Altman's conference remarks. She looks at the cost line. The Communications Workers of America and various enterprise-tech analysts have been pointing out for months that the buyer-side substitution logic, humans versus tools, at a large cost delta, has not changed because lab CEOs changed their tone.⁴ If anything, the softer public framing gives deploying firms cover: the labs are saying AI will augment workers, so we can say that too, while continuing to do what we were already doing.

The third is the timing. Fortune reported in May that both OpenAI and Anthropic are preparing roadshows this quarter.² I am generally cautious about IPO-cynicism arguments, because they are cheap and they flatter the writer. Executives update their views for many reasons, over long horizons, and pattern-matching every softer statement to the next S-1 is lazy. But the coordination of the pivot across at least four executives, in roughly a one-month window, at a moment when two of the four are about to ask public markets to price them, is not nothing. It is at least a fact that deserves to be named rather than politely ignored.

Read Altman's own sentence carefully. "Roughly right on technological predictions and pretty wrong on the social and economic implications." That is not a small admission. It says: the people who ran the technology forecast, the labs themselves, were more accurate than the people who ran the social forecast, who were also the labs themselves. Which means the labs' internal model of what their products would do to people was worse than their model of what the products would do technically. There are people who have been saying exactly this for several years. Some of them work at CWA. Some of them are labour economists whose warnings were characterised as alarmist in 2024. Some of them are the workers who lost jobs under the first framing and are now told, in a different tone, that everything is fine.

None of those people have been invited to help draft the new framing either.

That number is the ground the pivot is happening on. It is not a technical finding; it is a political one. Public consent for the current pace of AI deployment is thinning, and pending IPOs need public consent, or at least public tolerance. A "human flourishing" pitch deck reads better into that headwind than a "we automate white-collar work" one. Whether or not any individual executive is consciously calibrating to Pew's number, the industry as a system is.

The distributional question is the one that keeps getting skipped. When the first framing was set, the people it described, the workers who would allegedly be displaced, were not consulted. When the framing was revised, the people who had actually been displaced under the first story were not consulted either. In both rounds, the framing was set by the firms doing the deploying, spoken to the audiences whose confidence the firms needed (investors, regulators, media), and the population whose lives were the subject was addressed as a topic rather than as a party.

This is what I mean when I say the walk-back protects the firms, not the workers. If Amodei's original prediction had been correct and half of entry-level white-collar jobs had vanished by 2028, the labs would carry political and regulatory exposure that the walk-back is designed to reduce. If the original prediction was overstated and the softer reality is closer to what materialised, the workers who were let go in the intervening period on the strength of the original story are still let go. Either way, the firms come out lighter than they went in. The workers do not.

I do not think it is churlish to notice this. I think it is the whole point.

What actually happens next depends on the enterprise buyers, not the lab CEOs. If procurement continues to substitute AI tools for headcount at anything like the cost delta reported by CWA and enterprise-tech analysts, the softer public framing will run in parallel with a hardening deployment reality, and the gap between the two will widen. If enterprise buyers themselves are updating, if the rehire-regret stories are large enough to slow substitution, then the softer framing will match the ground truth, and the walk-back will look, in retrospect, like calibration. I do not yet know which of those two is the dominant force. I am watching enterprise deployment data more closely than I am watching CEO speeches.

What I would not do is take the speeches as the story. The story is what enterprises are buying, what workers are experiencing, and whether the workers have any voice in either. On the last of those, the answer in July 2026 is the same as the answer in January 2025.

The most honest thing an AI lab CEO could say right now is not "we were pretty wrong on the social and economic implications." It is "we said things that shaped how our customers thought about their workforces, and we said them without consulting the workforces, and we are now saying different things, also without consulting them, and we would like credit for the update." I have not heard anyone say that. I do not expect to.

Footnotes

¶