The Canaries gap hit 19 percent, and it only widened for people who never got the first job

On August 12, the Stanford Digital Economy Lab published an update to the study that has done more than any other to shape how people talk about AI and jobs. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen re-ran “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” on ADP payroll records covering more than 26 million American workers, with the data now running to June 2026. The headline number moved in the direction everyone feared. Employment of 22-to-25-year-olds in the most AI-exposed occupations now sits about 19% below where it would be had it tracked their less-exposed peers, up from a 15% gap in July 2025.

The number that matters more is the one that did not move. Experienced workers in the same occupations show no comparable gap. Between November 2022 and June 2026, employment for 22-to-25-year-olds in the two most-exposed occupation quintiles fell roughly 11%, while it rose about 10% in the three least-exposed quintiles. The declines concentrate where AI substitutes for tasks and employment is flat or rising where it complements them. And the authors are explicit about the mechanism: the adjustment is coming through reduced hiring rather than separations. In their words, it “is showing up primarily in employment rather than base pay,” and “We do not see widespread, economy-wide job displacement associated with AI.”

I have read a lot of commentary on this update over the past month and most of it treats the 19% as a bigger version of last year’s story. I think that misreads the paper. The update is a different story, because the widening happened in only one place. Nobody who already had one of these jobs is measurably worse off. The whole effect is borne by people who have never held the job at all, and that changes what the sensible response looks like.

The gap is a shut door

Think about what a hiring freeze at the bottom of a ladder produces if you measure it a year later. The people on the ladder are still there. Their pay, according to the same data, has not moved much. The people who would have joined the bottom rung in the past twelve months simply were not added, and the longer the freeze runs, the larger the missing cohort looks against the trend line. A gap of 15% becomes a gap of 19% without a single additional person losing anything, because the entire increase is made of jobs that were never posted.

This is exactly the pattern the Stanford numbers show, and it is why “AI took jobs” is the wrong sentence. The more accurate sentence is that AI stopped the first job from being posted. The distinction sounds academic until you notice that it points the two groups of workers in different directions. A displaced worker needs a new job. A never-hired worker needs a first job, and the job that would have been the first one is the specific thing employers have decided they can skip.

Why employers can skip it is not mysterious. The tasks that a first-year analyst or a junior developer used to do were the tasks a firm could hand to someone with no context, because the tasks themselves carried the context: format this, summarise that, reconcile these, draft a first pass of the other. Those are the tasks that a model does now for a fraction of a salary, and a firm that used to hire six graduates to get two good seniors in five years has quietly decided to buy the seniors later and skip the six. Whether that decision is wise for the firm in 2031 is a separate question, and I suspect the answer is no. The decision has been made either way, and it is the graduate who bears the bill.

The macro backdrop makes the freeze easy to hide

The September 4 jobs report from the Bureau of Labor Statistics is the reason this freeze does not look like a crisis from the outside. Nonfarm payrolls rose 162,000 in August against a 53,000 consensus, and a prior twelve-month average of just 31,000. Unemployment held at 4.1%. Food services and drinking places added 59,000 jobs, local government education added 42,000. On a national scale the labour market looks fine, and a politician can say so with a straight face.

Inside that headline, the information sector shed 23,000 jobs in August, which put its twelve-month average at minus 8,000 per CNBC’s count. Professional and business services were little changed. So the picture is two speeds: restaurants and schools hiring, software and media not. A 22-year-old with a computer science degree does not compete for the 59,000 restaurant jobs, and the 4.1% unemployment rate tells them nothing about their own odds. Teen unemployment, incidentally, edged up to 14.1% in the same report, which is a reminder that the bottom of every ladder is where slack shows first.

There is a second reason the freeze is easy to hide, and it is about who counts as unemployed. A graduate who cannot get the analyst job they trained for and takes a retail or hospitality job while they keep applying is employed, in the statistics, in a sector that is hiring. The BLS sees a filled restaurant job. The Stanford payroll data sees a missing analyst. Both are correct, and the person in question would describe their situation using the second description, not the first. When the information sector’s twelve-month average is negative and food services adds 59,000 in a month, some portion of that 59,000 is the canary cohort, wearing an apron, still refreshing the job boards.

What I find most useful about putting the Stanford paper next to the BLS release is the way it dissolves the argument that has been running in circles for a year. One side says the AI jobs panic is overblown because unemployment is low. The other side says the panic is justified because graduates cannot get hired. Both are reading the same economy accurately. The aggregate is healthy, and one narrow cohort inside it is in real trouble, and the reason the two facts coexist is that a hiring freeze at the entry level of a few occupations is small enough to vanish inside 162,000 net new jobs while being large enough to define the working life of everyone it touches.

What a first job is actually for

If the job that is missing is the first one, then the question for a graduate is how to acquire the thing a first job used to provide by some other route. I want to be precise about what that thing is, because it is not the tasks. The tasks are gone and are not coming back. What the first job provided was context: watching how decisions get made, learning what a good version of the work looks like, being trusted with something small and then something larger, and acquiring a reference who will vouch for your judgement rather than your degree. Every one of those is still needed by every employer, and none of them is being supplied by a model. The employers have simply stopped paying for the delivery mechanism.

The strategy that follows is to get past the first job by any door that is open, and to stop treating the front door as the only legitimate entrance. The front door, the graduate scheme at a large firm in an exposed occupation, is the door the Stanford data says is shut. The other doors are still open to varying degrees.

The first is the adjacent role. The Stanford finding is that employment rose about 10% for young workers in the three least-exposed quintiles. That is a large number of jobs, and some of them sit one desk away from the exposed occupation the graduate trained for. A person who wanted to be a junior analyst at a consultancy can be an operations coordinator at a client, a data person inside a hospital network, the one technical hire at a firm that has never had one, or the person a logistics company puts in charge of the spreadsheet nobody else understands. The job title is worse. The context is the same, and context is what the first job was for.

The second is the smaller firm. Large firms can skip the junior tier because they have a deep bench of seniors and can buy more. A forty-person company has no bench, and a cheap, willing, fast-learning graduate who arrives with a working knowledge of the tools is closer to a senior in that environment than they would be anywhere else. The pay is lower and the brand is weaker. The exposure is wider, and exposure is what compounds.

The third is the apprenticeship, in the plain sense of a structured entry route where the firm has committed in advance to training people it has not yet met. Deloitte UK announced on September 10 that it will take on 1,353 people into early-career roles this autumn, from graduates to industrial placements, a level it says is consistent with previous years, in a sector where the talk is about cutting juniors to bank AI savings. I would not build a career plan around one firm’s press release. I would notice that the firms which have kept an intake are now the scarce asset in the market, and that scarcity is a signal about where to apply.

And there is a fourth door, which is to make something that shows judgement without needing anyone’s permission to start. A public project, a shipped tool, a piece of analysis on real data, a written explanation of a hard decision that a hiring manager can read in ten minutes. This has always been advice, and in the past it was slightly precious advice, because a good degree and a good interview would get you in anyway. Now the interview may not be offered, and the project is the only artefact that says what the first job used to say on your behalf.

The uncomfortable part for experienced workers

The paper’s cleanest finding is that experienced workers show no comparable gap, and the natural reaction from anyone over thirty in an exposed occupation is relief. I think the relief is partly earned and partly misread, and the misreading is where the risk sits.

It is earned because the data really does say what it says. Employment for experienced workers in exposed jobs has held, and pay has held, and the mechanism the authors identify is a hiring one. Nobody is coming for the senior engineer’s job this quarter on the strength of these numbers.

It is misread because a hiring freeze at the bottom is not a stable state. The firms that stopped hiring juniors did so because the junior tasks could be automated, and the boundary of what counts as a junior task moves every year. The senior who is safe today is safe because the tasks they do still require context a model does not have. That is a description of a moving line, and the experienced worker is on the wrong side of it eventually unless their job keeps acquiring the kind of context the line cannot cross. The 19% gap is what it looks like when the line moved across the junior tier. There is no reason in the paper to believe it stops there, and the authors themselves are careful to say only that they do not see economy-wide displacement yet.

The second thing experienced workers should stop assuming is that the missing juniors are someone else’s problem. Every senior in an exposed occupation was once a junior who was given a first job by a firm that did not strictly need to give it. The pipeline that produced the current senior tier has been switched off, and the senior tier is now a stock that is not being replenished. In five years, the firms that skipped six graduates to buy two seniors later will discover that the seniors they wanted to buy were never made. That is good news for the seniors who exist, in the narrow sense that scarcity raises prices. It is bad news for anyone who runs a team and will need to hire into it, and it is a problem the senior tier can see coming and is best placed to argue against inside their own firms. I am not optimistic that many will, because the incentive to argue for a hiring line that costs money and produces nothing for three years is weak in any quarter and weaker in this one.

Reading the next update

The Stanford team measured to June 2026 and published in August, so the next update will likely land in early 2027 with data through the end of this year. I want to say what I would look for in it, because the temptation will be to read the single headline number again.

If the 22-to-25 gap widens again and the experienced gap stays at zero, the hiring-freeze reading holds and the advice above stands. If the experienced gap starts to open, the line has moved and the story changes from a first-job problem to a displacement problem, which needs a different response and a different essay. And if the 22-to-25 gap stops widening, which is possible if the firms that froze hiring in 2024 and 2025 discover the bench problem sooner than I expect, that would be the first evidence that the freeze was a one-time reset rather than a permanent removal of the bottom rung.

Whichever of those we get, the current data is clear enough to act on. The job that is missing is the first one. The context it provided is still needed and still learnable. The doors that lead to it have moved, and the graduate who keeps knocking on the one that is shut is spending the most valuable year of their career on a queue that is not moving. I would rather see them walk around the building.

The Canaries gap hit 19 percent, and it only widened for people who never got the first job was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.