مقالات
The Unemployment Rate Is Fine. That Is the Trap.
By Gagan Malik
Three years of AI headlines and unemployment is near record lows. The office is still full. The doomers were wrong again. The chatbot took nobody's job. Pour yourself a flat white and open your portfolio dashboard. Here is the problem with that. Anthropic's economists Maxim Massenkoff and Peter McCrory, writing in their February 2026 paper "Labor Market Impacts of AI," measured actual AI usage against theoretical capability across every major occupational category. The headline finding everyone repeated: no unemployment spike. The finding nobody mentioned: hiring of workers aged 22 to 25 in high-exposure roles dropped 14 percent from late 2022 onwards. The jobs did not disappear. They stopped being created. Those are two different problems, and you are currently holding only one instrument. anthropic
You Checked the Unemployment Stats and Felt Relieved. Put the Dashboard Down.
Harvard Business Review, in a February 2026 survey of 1,006 global companies, found that firms are cutting headcount projections because of AI's anticipated capability, not its demonstrated performance today. Pre-emptive structural reduction. Not a response to evidence. A bet placed quietly, in advance, on a future that has not fully arrived yet. hbr
The unemployment rate measures the stock of existing jobs. It does not measure the flow of new ones that were never opened. When a firm hires fifteen graduates instead of twenty-five, the rate does not move. Ten people who would have been hired become unrecorded, uncompensated, and invisible to every dashboard the optimists are currently pointing at. The instrument is not broken. It is measuring the wrong thing. That is worse.
Where Did All the Juniors Go?
I know this move because I made it. In Q3 2025, I cancelled two hires: Aisha, a junior UX designer, and Callum, a full-stack engineer fresh out of his bootcamp. I redirected $20,000 (£15,000) of combined headcount budget into AI tooling instead. In December 2025, I finished a design sprint and a working prototype alone in five days and sent the final files. The surprise was not the speed. It was what I noticed when I hit send: no Slack thread, no handoff call, no one else in either folder. I had built the whole thing in a room with no one in it and had not registered that until the door closed.
I kept the full margin. Aisha, twenty-two, first-class design degree from Northumbria University, graduated July 2024, sent eleven applications between August and December and received zero offers. Not because her portfolio was weak. Two of those companies had quietly removed the junior design tier from their hiring plan before she graduated. Callum, twenty-four, three years of self-taught development and a GitHub full of side projects, reached final rounds at two firms and watched both roles evaporate in the same quarter his offer was supposed to come. Neither of them appears in any unemployment statistic. Both are, on paper, fine. I did not think about either of them once until I sat down to write this sentence.
The Hire That Never Happened Is Still a Career That Was Taken.
The Atlantic, in a February 2026 feature on AI and the labour market, reported that Dario Amodei, the chief executive of Anthropic, publicly expects 10 to 20 percent unemployment as AI capabilities reach full deployment. That is not a critic's prediction from the outside. That is the founder of the product doing the displacing, describing what he expects his own product to eventually do at scale. theatlantic
The distance between 'no unemployment spike yet' and '10 to 20 percent at completion' is not a reason to exhale. It is a timeline with most of its travel still ahead. Entry-level roles that are not being posted today are the leading edge of a structural adjustment that headline statistics will not register for another three or four years. By then the counterfactual is invisible, the policy window has closed, and the debate is purely academic for everyone who needed the pipeline to be open and found it shut before they arrived.
The Alarm Is Not Going Off. That Is the Problem.
Carbon monoxide is colourless, odourless, and undetectable without the right instrument. You feel fine. Slightly tired, maybe. The room looks completely normal. The first signal that something is wrong is frequently the last one you are conscious enough to act on. The problem was never invisible. You were checking for smoke. Carbon monoxide does not trigger a smoke detector. That is not what kills you.
Forbes, in a March 2026 methodological critique of the Anthropic paper by analyst Hamilton Mann, argued that the study cannot measure what it does not track: wage compression, stalled promotion velocity, and the systematic disappearance of junior pipeline roles. None of these appear in unemployment data. The unemployment rate is the smoke detector. It is not going off. The junior hiring pipeline has been running 14 percent below its pre-2022 level for three years, per the Anthropic paper's own figures, and every government alarm is perfectly, uselessly silent. The gap between the instrument and the reality it fails to capture is not a technical footnote. It is where Aisha and Callum live. forbes
The Case for Patience Is Genuinely Strong. Here Is Why It Is Not Enough.
Let me hold the optimist position at full strength, because it has earned it. The Peterson Institute for International Economics, in a March 2026 review of the AI labour research base, argued that the evidence is genuinely thin, that the Anthropic paper measures one company's product usage rather than economy-wide displacement, and that every major automation wave in recorded history produced net job creation rather than net destruction. The historical null hypothesis is that new categories of work emerge to absorb displaced workers. History has a strong track record on this. I am not going to pretend that argument is weak, because it is not. piie
The counter is specific, not rhetorical. Daron Acemoglu and Pascual Restrepo, in a peer-reviewed paper published in Econometrica in 2022, found that automation accounted for 50 to 70 percent of the rise in US wage inequality between 1980 and 2016, and demonstrated that net job creation requires a sufficient supply of new complementary tasks for displaced workers to absorb. When new task creation slows, automation produces sustained wage depression and reduced employment share even without touching the headline unemployment number. The historical optimism is not wrong in principle. It is conditional on new work arriving fast enough, and for the right people. Neither condition is confirmed by the current evidence.
Three Sentences Before You Close This Tab.
The Financial Times noted in March 2026 that legal, finance, and management sectors sit at 80 to 90 percent theoretical AI exposure and under 15 percent observed usage, with most of the adoption curve still ahead. When that gap closes, it will close fast, because the deployment barrier is a software procurement approval and a six-week onboarding cycle, not a factory retooling. If you are Aisha or Callum, portfolio current, waiting for the pipeline to reopen, the smoke detector will not tell you when the room has already changed. ft
