The jobs number was good. The job market wasn't.

The US economy added 172,000 jobs in May, beating nearly every forecast. Inside that same report, the share of unemployed Americans who have been out of work for six months or more climbed to 27.5 percent — up from 20.4 percent a year ago. Both numbers are real. They describe different people, and the gap between them is the story.

The headline number is doing what headline numbers do: it is closing the conversation. By Friday afternoon, the framing was settled. A surprisingly strong jobs print, upward revisions to March and April, the unemployment rate holding at 4.3 percent. The Bureau of Labor Statistics (BLS, the federal agency that publishes employment data) released a clean report and the market took its cue.

I want to argue that this framing is not wrong so much as it is incomplete in a way that systematically erases the people the report is meant to describe. The 172,000 figure is about who got hired in May. The 27.5 percent figure is about who has not been hired in any month for half a year or longer. These are not the same labour market. Treating them as one number, net jobs added, is what makes the second group disappear from the political conversation, even as the second group grows.

The arithmetic of disappearance. Two million Americans have now been unemployed for 27 weeks or more. That is roughly 524,000 more people than a year ago, sitting in the same condition. The overall unemployment rate has barely moved, because long-term unemployment does not move the rate much; the people who stop being counted as unemployed do not necessarily get hired, they often just stop being counted.

What is unusual about this is that it is happening alongside continued payroll growth. Normally, long spells of unemployment build during recessions, when layoffs are high and openings are scarce. This is something different. Layoffs are not historically high. Openings are not historically scarce. But the rate at which people move between the two, the churn that is the labour market's actual machinery, has slowed sharply. Indeed's Hiring Lab calls it "one strong headline, but two realities." The Center for American Progress calls it "underlying labor market slack." Analysts have started calling the pattern a frozen labour market: a market where few people get fired and few people get hired, and where the people who do lose their jobs find the door behind them has quietly closed.

Who this happens to. If you have a job in May 2026, the data is reassuring. Wages are growing, layoffs are low, your employer is probably not cutting. If you lost your job in November 2025, the data is something else. You are now in your seventh month of looking. The pool of openings you are applying to is thinner than the headline suggests, because the hires rate, not the openings rate, the hires rate, is what determines whether anyone actually moves through the door. And the hires rate is depressed.

This is the structural feature that headline reporting cannot see. Net payroll growth is a flow measure. It nets the hires against the separations and reports the difference. It does not tell you anything about how long the average separated worker spends before becoming a hire. In May, that duration kept climbing.

The AI question, handled honestly. I am wary of jumping from "long-term unemployment is rising" to "AI is the cause," because the BLS data does not support that jump. The numbers are consistent with multiple explanations: post-pandemic normalisation, higher interest rates dampening business formation, sectoral shifts unrelated to automation. Causality is not in the report.

But there is a parallel data set that is harder to ignore. Challenger, Gray & Christmas, the outplacement firm that tracks layoff announcements, found that AI was cited as the reason in 40 percent of May 2026 layoff announcements. In January, that figure was 7 percent. A five-fold increase in five months. Tech-sector job cuts year-to-date are running 66 percent above last year's pace.

I do not think this proves AI is causing the long-term unemployment surge. I think it tells us something subtler and more uncomfortable: companies are now willing, eager, even, to attribute layoffs to AI. Gartner's read is that the figure reflects both real automation-driven cuts and a degree of attribution washing, where AI gets cited because it sounds forward-looking to investors. Both can be true simultaneously, and the mix is unknowable from outside the firm.

Here is what bothers me about that ambiguity. The worker who is let go does not get to know which it was. If the layoff is real AI displacement, retraining into a similar role is probably futile. If the layoff is cost-cutting dressed up as transformation, retraining might be exactly right. The worker has to choose a strategy without knowing which world they are in, while the firm that knows is not obligated to tell them.

The information asymmetry is itself a harm, separate from the layoff.

This is what I mean when I say the costs of these transitions get redistributed across institutional boundaries in ways that make them hard to see. The firm books the productivity gain. The worker absorbs the uncertainty. Neither shows up cleanly in the jobs report.

The distributional reading. Daron Acemoglu and Suresh Naidu have spent years arguing that the political economy of technological change is not preordained — that the same technology can produce broadly shared gains or concentrated ones depending on who has bargaining power and who writes the rules. Their frame is useful here. The May report tells us that aggregate employment is holding up. It does not tell us that the gains and the losses are distributed proportionally across the workforce. They are not.

The people in continued employment are doing fine. The people in short unemployment spells are mostly cycling back to work, though more slowly than before. The people in long unemployment spells, the 27.5 percent, are a population that is growing, and they are growing because the rehire mechanism that used to absorb them is running cold. Five hundred thousand additional people, year on year, are stuck in a condition that is corrosive economically, psychologically, and socially. That is not a footnote to a strong headline. That is a structural feature of the labour market that the strong headline obscures.

I want to be careful not to overclaim. The economy did add 172,000 jobs. Healthcare and construction and local government continued to hire. Those are not nothing; they are real jobs going to real people, often in the parts of the workforce that have been historically squeezed. A piece that pretends the entire report is bad news is not honest about the report.

But the same report contains a slow-moving humanitarian problem that the framing of "beat expectations" actively erases. Long-term unemployment is not a counting artefact. It is people whose savings are draining, whose health insurance is gone, whose CVs are developing the gap that future employers will use as a reason not to call them back. The longer someone is out of work, the harder it becomes to return — this is one of the most robust findings in labour economics. Every additional month in the 27-week-plus column is an additional month of accumulated disadvantage.

What I think is actually going on. The labour market in May 2026 has bifurcated into two populations. The first is employed, reasonably secure, benefiting from continued wage growth and low layoff risk. The second is unemployed and increasingly stuck, facing a hiring rate that has not recovered and an attribution environment where the reasons for their displacement are murky enough to make planning a next move very difficult.

The political and economic conversation has tools to talk about the first group. We have very few tools to talk about the second, and the ones we have — extended unemployment insurance, active labour-market policy, retraining vouchers — are politically dormant in the US right now. Meanwhile the long-term-unemployed share is the labour metric most likely to matter politically by the midterms, regardless of whether AI turns out to be the dominant cause or one of several.

I do not think this resolves with the next jobs report. Frozen labour markets thaw slowly, and the longer they stay frozen, the more the long-term-unemployed share will continue to climb, because the inflow of newly long-term-unemployed exceeds the outflow of people finally getting rehired. We will keep getting headline numbers in the range that allow the framing of "the labour market is resilient." Underneath, a population that no one is writing about will keep growing.

The honest reading of May's report is that both things are true. The economy is hiring. The job market is failing the people already outside it. Treating those as one number is a choice. It is the choice that lets the conversation move on.

Footnotes

¶