AI Cut Junior Hiring 22%. The Fed Says Nothing Happened.
Three large studies measured artificial intelligence against entry-level hiring in the last twelve months and produced three different answers. Harvard found junior hiring at AI-adopting firms down 22%. Stanford found 22-to-25-year-olds in AI-exposed occupations sitting 19% below their peers. The Federal Reserve found no effect at all. None of the three is wrong. They measure different quantities, and that difference is the entire story.
This matters beyond economics, because these are the numbers that will be quoted at you in a board meeting within the next quarter. Whoever wins the argument sets hiring policy, training budgets and AI regulation for the next decade. Right now each side quotes only the study that agrees with it, which is a reliable sign that nobody has read all three.
The three findings, side by side
| Study | What it measured | Sample | Finding |
|---|---|---|---|
| Harvard, Aug 2025 | Hiring and headcount by seniority, inside firms that adopted generative AI | 62m workers, 285,000 firms, 245m job postings, 2015 to 2025 | Junior hiring down 22% in 2023; junior headcount 7.7% below non-adopters after six quarters; senior headcount unchanged |
| Stanford, Aug 2026 | Employment of young workers in AI-exposed versus less-exposed occupations | US payroll data, updated from the 2025 edition | Ages 22 to 25 in the most exposed occupations 19% below less-exposed peers, up from 13% a year earlier |
| Federal Reserve, 2026 | Job postings, layoffs and headcount across the whole economy | Job-postings data joined to the Census Business Trends and Outlook Survey; regional business surveys | No evidence AI adoption is reducing hiring; very few AI-driven layoffs; aggregate employment effect under 0.4% for 2026 |
Harvard and Stanford agree on the mechanism, not just the direction
The Harvard paper is the more forensic of the two. It tracks employment by seniority within each firm, so it can separate a company that shrank from a company that simply stopped hiring. The answer is unambiguous: the junior decline is driven by slower hiring, not by more separations. Nobody was fired. The door was quietly closed.
The Stanford update, published this month, cuts the data a different way. It compares young workers in AI-exposed occupations against young workers in less-exposed ones, and the gap widened from 13% to 19% in a single year. Older workers in those same exposed occupations are largely untouched.
Two independent teams, two datasets, two methods, and both land on the same explanation: the effect concentrates where the work is codified and spares the work that is tacit. Roles built on formal, documented, written-down procedure lost ground. Roles built on judgement acquired by practice grew.
The codified part of the job was the documented part
Nobody in this debate says the next part out loud. Exposure was not handed down by the model. It was manufactured in advance by the organisation itself. Codified knowledge is the knowledge somebody wrote down: the onboarding guide, the runbook, the ticket templates, the decision tree taped inside the service desk. Every one of those artefacts exists because a manager wanted junior work to be repeatable.
Repeatable is the same word as automatable. The firms that documented their junior work best are the firms that automated it first, and the documentation became the specification for its own replacement. That is the inversion at the centre of this data: the document pipeline that made the work legible to a new starter is the pipeline that made it legible to a language model.
Why the Federal Reserve is not wrong either
It would be convenient to dismiss the Fed as slow. It is not. Its March 2026 analysis joined job-postings data to the Census Business Trends and Outlook Survey and found no evidence that AI adoption reduces hiring. Its New York bank reported in August 2026 that firms describe very few AI-driven layoffs, and called the effect evolutionary rather than disruptive. The Atlanta bank put the expected aggregate employment effect for 2026 below 0.4%.
Those same regional surveys explain most of the gap. AI use among service firms went from 25% to 40% in a year, with 44% expected within six months. Among manufacturers it went from 16% to 26%. So when the Fed averages across the economy, well over half the denominator has adopted nothing yet. Harvard and Stanford look deliberately inside the adopters. Averaging a concentrated effect across a mostly unaffected population is exactly how a real 22% becomes a rounding error.
Hiring is a flow. Headcount is a stock.
This is the reconciliation, and it is not a compromise between the camps. It is a measurement fact. Harvard and Stanford measure a flow: who gets hired, per quarter, into which seniority band. The Fed measures a stock: how many people currently hold a job, and how many lost one.
A job that is never posted generates no layoff, no separation, no unemployment claim and no plant-closure notice. Every instrument built to detect job loss is a stock instrument. Job non-creation is invisible to all of them. The harm never lands on anyone the labour statistics can see, because it lands on people who do not have the job yet.
That is why better aggregate data will not settle this. Aggregate employment is the wrong instrument for a distributional event, in the same way an average temperature tells you nothing about one cold room. The identical pattern showed up in security operations, where two surveys measured the same room and returned opposite verdicts because one asked leadership and the other asked the floor.
The economist behind the scary number does not forecast an apocalypse
The Stanford team that produced the 19% figure also concluded that an AI job apocalypse is unlikely. That is not a contradiction and it is not a retreat. It is the correct reading of their own data: a sharp, real, measurable hit to one age cohort in one set of occupations, with no aggregate catastrophe attached. Both halves are true at once, and any summary that keeps one and drops the other is doing politics rather than economics.
Nobody has costed the pipeline
All three studies report the same secondary finding, and it gets almost no attention: senior employment is fine. Tacit-knowledge roles are growing. Experienced people are worth more than before.
Seniors are made out of juniors. Tacit knowledge is what you get from doing the codified work badly for two years while somebody more experienced corrects you. Remove the codified work and you have removed the apprenticeship, and the bill arrives in about eight years, long after the executive who approved the freeze has moved on. It is a decision with a long lag and no feedback signal, the same structural trap that sinks most AI projects that die after the pilot, where the cost lands in a budget nobody connects back to the original call.
What to do with three answers
- Planning macro exposure: use the Fed number. It is the right instrument for a whole economy, and the honest answer there is that nothing dramatic has happened yet.
- Planning your own firm: use Harvard and Stanford, but only if you have actually adopted. Their finding is about adopters, and if that is you, the 22% is your number and the 0.4% is not.
- Deciding what to learn: exposure tracks codified work. The defensible ground is judgement that was never written down, which is also the hardest thing to acquire now that the entry route to it is narrowing.
- Quoting any of it:say which quantity the number measures. “Hiring fell 22% at adopters” and “employment is flat” are both true, and the gap between them is the entire argument.
Why it matters
The instruments we use to detect economic harm were built for an era when harm arrived as job loss. This one is arriving as job non-creation, which no statistical series was designed to catch. By the time an aggregate number moves far enough to end the argument, the cohort that absorbed the change will be five years into a career that started later, paid less, and skipped the part where somebody taught them. Waiting for the aggregate to agree is itself the decision.