The Unemployment Number Is Lying to You
- Rich Washburn
- 2 days ago
- 5 min read
Andrew Yang went on camera recently and said what a lot of people in this industry have been saying quietly for months: hundreds of thousands of white-collar jobs are going away in the next twelve to twenty-four months. He cited Verizon's 13,000 cuts, Amazon's 16,000, Accenture's waves of reductions. He said he'd spoken to a senior Fortune 500 executive sitting on thousands of layoffs that haven't gone public yet.
His debate opponent pushed back with the standard rejoinder: unemployment is at 4.6%. The directly attributable AI layoff number is a tiny fraction of the workforce. The studies people cite started before ChatGPT even existed. Check back in a year or two. Both of them are working with real data. One of them is reading it correctly.
The Wrong Scoreboard
The unemployment rate measures one specific thing: the percentage of people actively looking for work who can't find it. It is not a measure of labor market health. It is not a measure of displacement. And it is particularly useless as a leading indicator of a structural transition in how work gets done.
Here's why. When AI eliminates the need for a role, companies don't fire the current occupant on Tuesday and post a press release on Wednesday. What actually happens is quieter and more durable. The person in the role retires, or takes a job somewhere else, and the position doesn't get backfilled. The open requisition that would have been posted eighteen months ago never gets approved. The hiring freeze that started as "cost discipline" becomes the new operating model. None of that shows up in the unemployment rate. All of it shows up in the hiring rate.
Yale Budget Lab put numbers to it earlier this year: the 2026 labor market features low layoffs and low hiring — especially of unemployed workers. That combination is the fingerprint of a structural contraction. It's not a recession. It's a reorganization. The machine isn't shedding workers. It's just quietly deciding it needs fewer of them going forward.
300,000 to 500,000 "missing jobs" by mid-2025 — roles that would have existed in a pre-AI labor market and simply weren't created. That's the number that doesn't make the headline. That's the number that matters.
What Meta Just Did
Let's make this concrete. In May 2026, Meta cut 8,000 employees. Not quietly — they announced it as an AI restructuring. First they reassigned 7,000 workers to AI-focused roles. Then they cut 8,000. The company was explicit about the causal chain: AI is absorbing work that humans used to do, and the org chart is being redrawn around that reality.
This is not a single data point. It's a pattern. Amazon cut 14,000 corporate roles this year, citing AI as a primary driver. Salesforce, Workday, Snap, Microsoft, Google — the list of companies reducing headcount while simultaneously announcing massive AI infrastructure investments is not a coincidence. It is a confession. When a company cuts people and pours billions into AI in the same quarter, they're not hiding the thesis. They're executing it.
The Anthropic Problem
One piece of evidence frequently cited by the skeptic side is Anthropic's own labor market research, which found no unemployment impact in AI-exposed occupations. There's something worth flagging about that finding.
Anthropic makes the AI. Their research measures how their AI is being used — Claude usage patterns, task substitution rates, occupational exposure indices. What it does not — and arguably cannot — measure is the downstream effect on hiring decisions made by companies that are integrating that AI and quietly concluding they need a smaller workforce to do the same work. It's not fraudulent. It's a narrow lens with a motivated frame. The same way a pharmaceutical company's internal research on drug efficacy isn't worthless, but you'd want to see it replicated by someone without a financial interest in the outcome.
The independent data tells a different story. Stanford's "Canaries in the Coal Mine" study found a 6% employment decline in entry-level occupations most exposed to AI, measured from late 2022 through mid-2025. The Congressional Foushee AI Jobs Report documented 54,694 jobs directly attributed to AI reductions in 2025 alone. Dario Amodei — the CEO of Anthropic, the same company whose research shows limited impact — publicly stated that AI could eliminate roughly 50% of entry-level white-collar jobs. The company's research division and its chief executive are not telling the same story.
Where Did the People Go
Here's the question that makes the unemployment number even less useful as a signal: where do displaced workers go when they don't file for unemployment? Three places, based on current data.
Some exit the labor force entirely. The Minneapolis Fed tracked over 6 million Americans who moved from "not in the labor force" back to active job seeking in Q3 2025 — and found the market wouldn't absorb them. They're not counted as unemployed. They're counted as discouraged, or as having left the workforce, or as not counted at all. Some go freelance. 70 million Americans are now freelancing — 36% of the entire workforce. That number has been climbing since 2023, and the pace is accelerating. A laid-off analyst who hangs a shingle and picks up three clients isn't unemployed in any statistical sense. She's also not receiving the salary and benefits she had before. Some start companies. Startup formation among recently displaced workers is up sharply — one recent estimate cited a 67% surge in new ventures following AI-related job displacement. Some of those will be successful. Most won't replace the income, benefits, or stability of the role they lost.
The unemployment number doesn't capture any of this. It sees a healthy number and reports a healthy market. The people living inside the data know better.
The Real Timeline
Yang's debate opponent said the studies people cite began before ChatGPT launched — implying the disruption they measured predates the technology doing the disrupting.
That's actually the most interesting part of the argument. The Stanford entry-level employment decline started in late 2022. ChatGPT launched in November 2022. The disruption and the tool arrived at essentially the same time, and the effects showed up in the data within months. We are now three and a half years into generative AI deployment at scale. The hiring freeze has had time to compound. The requisitions that weren't posted in 2023 weren't posted in 2024 either. The structural contraction is baked in at this point — and most of the companies doing it haven't announced it, because you don't announce a hiring freeze. You just stop hiring.
The argument that we should "check back in one to two years" was made in early 2026. The data from 2025 — 54,000 directly attributed AI job cuts, 300,000-500,000 missing jobs, Meta's 8,000-person restructuring — is the check-in. The check-in says: it started.
What This Actually Means
I'm not writing this to generate panic. The transition is real, the displacement is real, and the people experiencing it deserve analysis that doesn't gaslight them with a headline unemployment number that was designed to measure something else entirely. The honest read is this: we are in the early innings of a structural reorganization of white-collar work. The pace is accelerating — every quarter that AI models get more capable is another quarter of companies discovering they can hold headcount flat or shrink it while output stays the same or improves. The unemployment number will eventually reflect this. It always does, with a lag. The lag is not evidence that nothing is happening. It's evidence that the scoreboard was built for a different game.
Rich Washburn is a technologist and strategist working at the intersection of AI infrastructure, capital, and national security. He is Managing Partner and Chief AI Officer at Eliakim Capital, and CIO of Data Power Supply.
Comments