Three findings, held together. First, the damage at the entry level is real and it is measured, not vibed: a Stanford Digital Economy Lab study of actual US payroll records — not surveys — finds that workers aged 22 to 25 in the most AI-exposed occupations have seen their employment fall in relative terms by 13 percent in the paper's original version and 16 percent in its revision, with young software developers specifically down nearly 20 percent from their late-2022 peak. Second, the layoff numbers that dominate headlines are far softer: a rising share of announced US job cuts is being blamed on AI, but tracker analysts themselves warn that firms use "AI" as cover for cuts they would have made anyway. Third, the counter-signals are documented and they cut against the doom frame: Australia's largest bank reversed a plan to replace 45 customer-service staff with a voice bot after admitting an "error," and IBM — which paused hiring for AI-replaceable roles in 2023 — announced in February it would triple US entry-level hiring in 2026. The desk's read: believe the payroll data about the bottom rung, discount the press-release layoff counts, and treat the top-line "300 million jobs" figures as exposure estimates, not forecasts of loss.
The number everyone remembers is Goldman Sachs's estimate that generative AI could expose around 300 million full-time jobs worldwide to some degree of automation, and the World Economic Forum's projection that some 92 million roles will be displaced by 2030. Established Those figures are real, and they are also almost useless for understanding what is happening right now, because "exposed to automation" is not "eliminated," and a 2030 projection is not a 2026 measurement. The useful question is narrower: where, if anywhere, can we already see AI moving employment in hard data? The answer turns out to be a specific, findable place — the first rung of the ladder.
1. The one place the data is unambiguous
The strongest evidence does not come from a survey, a CEO's LinkedIn post, or a consultancy forecast. It comes from payroll. In a paper titled Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford's Erik Brynjolfsson, Bharat Chandar and Ruyu Chen used high-frequency administrative records from ADP — the largest payroll processor in the United States, covering millions of workers across more than 730 occupations — to ask a precise question: since ChatGPT's launch in late 2022, what has happened to employment in the jobs most exposed to generative AI, broken down by age? Established
The headline finding, stated in the paper's own terms: early-career workers aged 22 to 25 in the most AI-exposed occupations have experienced a 13 percent relative decline in employment in the original August 2025 version of the study — a figure the authors revised upward to 16 percent in the November 2025 revision as more data arrived. In the same occupations, workers over 25 held roughly steady or grew. In less-exposed occupations, the young were fine. The decline is specific to the intersection of young and exposed. Established
Drill into a single occupation and it sharpens. By the account of one of the study's own authors, employment among 22-to-25-year-old software developers in the ADP data was down nearly 20 percent from its late-2022 peak — around ChatGPT's arrival — to July 2025. Established That is the number worth sitting with. Not because it settles the future — it does not — but because it is the cleanest signal we have that something occupation-specific and age-specific is already underway, and it lands exactly where theory would predict: entry-level coding is the kind of routine, well-specified, heavily-documented work that current models do best.
The paper's other facts matter for reading it correctly. The adjustment is showing up in headcount, not pay — firms are hiring fewer juniors rather than cutting the wages of the ones they keep. And the declines cluster where AI automates a task rather than where it augments a worker, which is the whole ballgame: the same technology that thins the junior ranks in one firm makes a senior worker more productive in the next. Established
2. Why the layoff headlines are the weaker evidence
Set against that payroll signal is a noisier one: the running count of announced job cuts that name AI as the cause. The trajectory looks alarming — a share in the low single digits through 2024 climbing steeply across 2025 and into 2026, with AI cited as the leading stated reason for cuts in several consecutive months of 2026. Assessed
Here the Navigator desk applies its standing scepticism, and it is the spine of this piece: a job cut blamed on AI is a claim made by the company doing the cutting. There is no auditor verifying that the eliminated role was genuinely automated rather than simply deleted in a budget round and then dressed up as forward-looking transformation. Tracker analysts themselves flag this. "AI" is a flattering thing to tell investors — it reframes retrenchment as innovation — and it is a convenient thing to tell the press, because it makes a cut sound strategic rather than defensive. Assessed When a firm cutting staff after a weak quarter cites AI, the honest analyst records two possibilities and does not pretend to know which dominates: real automation, or ordinary cost-cutting in a fashionable costume. Both are almost certainly present in the totals. That is precisely why the payroll study — which measures what actually happened to employment, regardless of what any manager said caused it — is the stronger evidence, and the press-release count the weaker.
3. The counter-signals nobody frames into the story
The doom narrative survives partly by ignoring the reversals, which are documented and specific.
In August 2025 the Commonwealth Bank of Australia, the country's largest lender, told its union it would cut 45 customer-service roles that a new AI voice bot would replace, claiming the bot had reduced call volumes by some 2,000 a week. Within weeks it reversed the decision and admitted an "error": call volumes had in fact been rising, the bank was paying overtime and asking team leaders to cover phones, and the roles were not redundant after all. The affected workers were offered their jobs back. Established The episode is not proof that AI cannot do the job. It is proof that the confidence with which firms announce it can is frequently premature — and that the "AI replaced them" headline sometimes precedes the quieter "actually, it didn't."
The larger tell is IBM. In 2023 its chief executive said the company would pause hiring for roles it believed AI could take, projecting thousands of positions gone. In February 2026 IBM's chief human-resources officer announced the company would triple its US entry-level hiring in 2026 — in her framing, for exactly the roles "we're being told AI can do." Established The stated logic is the second-order effect this desk exists to surface: if you stop hiring juniors because AI can do the junior work, you eventually have no seniors, because seniors are just juniors who were given years to grow. A firm that automates its bottom rung out of existence is eating its own leadership pipeline. IBM did the arithmetic and hired more at the bottom, not fewer.
4. What the signal actually says
Put the three pieces together and a coherent picture emerges — not the one either camp is selling.
The entry level is genuinely being squeezed, and the young are bearing it. The payroll data is too clean to wave away: in the exposed occupations, the door that used to admit twenty-two-year-olds is narrower than it was in 2022, and for young developers it is markedly narrower. Established This is the real damage, and it is concentrated on the people least able to absorb it — those without the experience that AI currently complements rather than replaces.
But "narrower door" is not "closing factory." The same studies that find the entry-level squeeze find experienced workers in those very occupations holding steady or growing, wages not collapsing, and the effect confined to where AI automates rather than augments. Assessed That is the signature of a labour market reorganising — shifting the value of a career from doing the routine task to supervising the machine that now does it — not one in freefall. The 300-million and 92-million figures describe exposure and a 2030 horizon; they are not a tally of jobs already lost, and treating them as such is the central error of the doom frame.
And the loudest evidence is the least reliable. The layoff counts that generate the scariest headlines are self-reported, unaudited, and structurally biased toward crediting AI. The reversals — CBA, IBM — show how often the confident replacement story does not survive contact with a rising call queue or a hollowed-out pipeline. The correct posture is to weight the payroll microdata heavily and the press releases lightly, which is the reverse of how the story is usually told. Assessed
Bottom line: Believe the payroll data, not the press releases. AI is measurably thinning the entry-level rung — 13 to 16 percent for exposed young workers, nearly 20 for young developers — while a large share of "AI layoffs" is ordinary cost-cutting in costume, and flagship replacements keep getting reversed. The ladder is losing its bottom step, not falling over. Watch who gets hired at 22, not how many jobs a consultancy says are "exposed."
5. Prediction — logged in the Ledger
The desk predicts that when the Stanford Digital Economy Lab next updates its "Canaries" data — via the paper's revisions or its public Canaries Dashboard — the relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations will still be present and no smaller than the 13 percent of the original August 2025 paper (it already stands at 16 percent in the November 2025 revision) at the next reading on or before 30 April 2027. This resolves correct if the exposed-young cohort's relative employment decline is at or beyond 13 percent in the lab's next published figure; incorrect if it has recovered above the −13 percent line (i.e. the gap has closed to shallower than 13 percent) or the age-and-exposure effect has disappeared.
Confidence: Assessed (Medium-High). The basis: the effect has strengthened, not faded, across the paper's own revisions, and the underlying capability driving it is not being withdrawn. The main ways this resolves wrong are a macro rebound in tech hiring that lifts all young cohorts, or a methodological revision that reclassifies exposure. Resolution: the next Stanford Digital Economy Lab "Canaries" figure published on or before 30 April 2027.