Artificial intelligence has spent two years firing almost nobody — and that is exactly what should scare you. The quiet catastrophe in the American labor market is not the mass layoff the doomsayers promised; it is the entry-level job that simply stops being posted, the first rung pulled off the ladder before a single human is handed a pink slip. Erik Brynjolfsson and his team at Stanford have now put a number on the silence: workers aged 22 to 25 in the jobs AI can do are running 19 percent behind their peers, and the gap has widened every year since they started measuring. The people eating that shortfall are the youngest adults in the room — the ones with a fresh diploma, a student-loan balance, and no offer letter to show for either.
The researchers titled their paper “Canaries in the Coal Mine”, and the metaphor is not decoration. A canary dies first because it is the smallest and the most sensitive, and the canary in this mine is a 23-year-old. Brynjolfsson — the Stanford economist who has spent years arguing AI would reshape work rather than merely speed it up — published the revised numbers in August alongside Bharat Chandar and Ruyu Chen. The finding is brutal precisely because it is boring: no crash, no collapse, no factory going dark. Just a door that has been quietly closing since the day ChatGPT went public.
Erik Brynjolfsson — The Canary Is a Twenty-Three-Year-Old
The number everyone should be arguing about is 19 percent. Employment for workers aged 22 to 25 in the occupations most exposed to AI now stands about 19 percent lower than it would have had it kept pace with their peers in less-exposed jobs, according to the revised paper from Stanford’s Digital Economy Lab. In raw terms, employment for young adults in the two most-exposed buckets fell roughly 11 percent between November 2022 and June 2026 — while the same age group in the three least-exposed buckets grew about 10 percent. Same age, same economy, opposite directions. The difference between them is whether a machine could do the job.
The gap has not stalled; it has widened. When Brynjolfsson and his colleagues first published in August 2025, the shortfall stood at 15 percent. By June 2026 it was 19. Experienced workers show no comparable gap at all. Strip away the caveats and the message is this: the people in their forties are fine, and the people trying to start are not.
The canary didn’t stop singing. It just got harder to hear over the quarterly earnings calls.
The Hiring Door — Nobody Got Fired, the Job Just Stopped Existing
Here is the detail that separates this from a routine recession story: it is not about people losing jobs. The adjustment operates almost entirely through reduced hiring of young workers, not through increased firings. Nobody is being marched out of the building. The building is simply no longer putting out the help-wanted sign for anyone under 26. The labor market is not shedding the young; it is refusing to admit them in the first place.
That distinction matters because it is invisible in the numbers we normally watch. The unemployment rate can stay flat, and total employment can keep climbing, while a specific cohort of 22-year-olds quietly learns that the jobs they trained for now belong to a subscription. The pain is concentrated in a place aggregate statistics are built to smooth over.
A layoff has a date and a name. A hiring freeze has neither, which is why it took payroll data to catch it.
Bharat Chandar — Codified Knowledge Is the Thing the Machine Ate First
The Stanford team also found a sharp dividing line in what kind of work is vanishing. Employment has fallen for young workers in occupations built on codified knowledge — the formal, standardized, textbook stuff that can be written down and taught. Employment has actually risen for experienced workers in roles that depend on tacit knowledge — the judgment you only get from doing the thing a thousand times, under a mentor, in a real situation. Generative AI is brilliant at the first and helpless at the second, and the data now show the labor market splitting along exactly that fault line.
Bharat Chandar, the labor economist on the paper, has flagged another gradient that should keep policymakers awake at night: women face greater AI exposure on average. The occupations the machine is best at eating are the ones where women are disproportionately employed. So the first rung of the ladder is being pulled up, and it is being pulled up a little faster for the people who already had the shakier grip on it.
The machine didn’t take the senior job. It ate the entry-level job, which was the only job a young person was ever going to get.
Samuel Dodini — Texas Is Counting the Job Postings That Vanished
The Stanford payroll data are not the only place the pattern shows up. Researchers at the Federal Reserve Bank of Dallas — Samuel Dodini and Tucker Smith — combed through millions of online job postings and found the same quiet retreat. After ChatGPT arrived in late 2022, job openings fell for occupations whose tasks AI can automate, and the decline was not confined to startups or failing firms. Surviving, established companies posted fewer openings and shifted their ads away from AI-exposed roles. By the first quarter of 2025, postings for the most-exposed positions were down about 8 percent relative to less-exposed ones, and the Dallas economists estimate the shift cut total job postings in Texas by roughly 2.6 percent in 2025, according to their analysis.
The authors are blunt about who absorbs the hit. Because the entry-level roles posted online rarely require more than a couple of years of experience, the decline lands on new labor-market entrants — recent graduates, first-job seekers, people switching careers. The New York Fed’s own tracking shows the unemployment rate for recent college graduates has climbed to unusually high levels during this same stretch of AI adoption, a shift that is easy to miss and hard to reverse.
You don’t need a crystal ball when the Dallas Fed and the New York Fed are pointing at the same kid.
The First Rung — A Generation Loses the Bottom of Its Ladder
Strip away the methodology and what remains is a question with a very long fuse. If the entry-level job is the way a worker learns the tacit knowledge that AI cannot copy, then cutting off that first job is not just a short-term earnings hit — it is how a generation fails to become experienced in the first place. The Dallas Fed found the shift has already changed behavior: some current college students are adjusting their studies to dodge the occupations the machine is eating. That is a rational response from an individual 19-year-old, and a slow-burning structural problem for everyone else.
Brynjolfsson is careful not to overclaim. The paper is descriptive, not a causal proof, and he insists no single study is definitive. But the pattern survives every alternative explanation his team throws at it — technology firms, interest rates, remote work, firm churn. The gap keeps widening well after interest rates peaked, and it is concentrated exactly where observed AI usage is doing the automating rather than the assisting.
The people who built the machine keep promising it will create jobs. It already did. It created the one nobody will hire you for.
Sources: Stanford Digital Economy Lab, Federal Reserve Bank of Dallas.