The layoffs are not a misdiagnosis

115,430 tech workers have been laid off in the first five months of 2026, many of them under an AI-productivity rationale that, by the industry's own internal research, describes a capability that does not yet reliably exist. TechCrunch this week called this "AI psychosis" on the part of CEOs. I want to argue the opposite. The layoffs are not a cognitive failure. They are working exactly as designed — for the people the system is designed to work for.

The numbers come from Layoffs.fyi, which tracks announced cuts across 152 companies in the first five months of the year — nearly matching the 124,636 layoffs recorded across the whole of 2025.¹ AI efficiency is the most frequently cited justification. The TechCrunch piece, which went to #1 on Slashdot and #6 on Hacker News the day it ran, frames this as executives over-extrapolating from polished demos into mass workforce decisions.¹²³

The capability gap is real and documented. Box CEO Aaron Levie, quoted in the same TechCrunch piece, describes what he calls AI's "last-mile distance" problem: agents can complete most of a task and then collapse at the final step, the one that needs judgment or sign-off.¹ MIT research referenced in the article models an 80 to 95 percent task success rate for AI agents arriving by 2029 — three years from now, not today.¹ A Harvard Business Review study cited in the same piece found that when AI generates output inside organisations, the bottleneck moves to executive approval; the chokepoint shifts upward, it does not disappear.¹

So the situation is this. 115,430 people have already lost their jobs on the basis of a productivity claim that, by the industry's own preferred research institutions, will not be validated for another three years, and that, even when it is, will leave a human approval step that someone still has to staff.

Why "psychosis" is the wrong frame

I understand the appeal of the diagnosis. It is a way of saying that the emperor has no clothes without having to name who profits from the parade. If CEOs are simply deluded, over-impressed by demos, over-confident about timelines, then the problem is cognitive, the solution is better information, and nobody has to be held accountable for anything except poor judgment.

But layoffs justified by AI work for the executives who order them whether or not the AI works. They reduce headcount costs immediately. They produce a story for investors that aligns with where capital wants the sector to go. They consolidate decision authority in the smaller group of people who remain. None of this requires the productivity claim to be true. It only requires the claim to be sayable.

That is not psychosis. That is incentive-aligned behaviour, accurately read.

Who is in the room when the decision gets made. A CEO announcing AI-driven layoffs is accountable to a board and to public markets. Both reward the AI narrative independent of whether the deployment delivers. The workers being laid off are not in either room. Neither, in most cases, are the engineers and operators closest to what the tools can and cannot actually do — which is why the practitioner audiences on Hacker News and Slashdot read this story as a labour-relations problem rather than a productivity one.²³ The people who actually deploy these systems do not believe the productivity narrative being used to justify the cuts. They have said so, in public, on the forums they read.

The last mile is also a power story

Levie's "last-mile distance" thesis is more useful than he probably intended. If AI stalls at the final step and humans must still approve output, the question is which humans.

It is not the 115,430 who have been laid off. Their last mile has already been removed from the chain. The approval step that the HBR research identified sits with whoever is left — narrower teams, more senior staff, and, ultimately, executives. When a workforce of a thousand people held a thousand small approval rights distributed across the org chart, no single person held very much leverage. When the same volume of output flows through a much smaller group of approvers, the approvers accumulate the decision authority that used to be distributed.

This is not a technology transfer. It is a power transfer that the technology is being used to legitimise. The labour cost falls. The decision rights concentrate. Both of these are useful outcomes for the people making the decisions, and neither of them depends on the AI doing what the press release claims.

The optimistic reading does not survive the structural one. You could read the HBR finding the other way and say: well, if executives remain the approval bottleneck, then total labour displacement is structurally limited, because output without approval is worthless. That is true in a narrow sense, and it is also why so much of the current layoff wave is going to land badly in eighteen months when companies discover they cut the people who were quietly running the approval chain in the first place. But "the system will limit how badly it can go" is not a defence of the layoffs. It is a description of the wreckage that comes after.

What to watch

Two things are worth watching over the next year. The first is whether the 2029 MIT capability timeline shifts forward or back as agent deployments mature in production rather than in demo conditions; the existing productivity research that TechCrunch cites finds no robust relationship between AI adoption and aggregate productivity gains so far, which is not encouraging.¹ The second is whether the firms now running lean re-hire quietly into the approval layer once the output bottleneck becomes visible internally, which is the pattern the HBR finding predicts.

If both happen — capability arrives more slowly than promised, and quiet re-hiring follows the cuts — the right name for what we are watching now is not psychosis. It is a one-way transfer: of jobs, of income, of standing, of voice, justified by a capability claim that the people making it did not need to be true.

The cost has already landed. The capability has not. That gap is the story.

Footnotes

¶