The Gist: There is now credible evidence that AI has slowed hiring for young workers in highly exposed jobs. It is not yet a crisis in the unemployment data. But the pattern is real enough to warrant serious attention from educators, policymakers, and anyone who cares about how the next generation enters the workforce.
If AI were reshaping the labour market dramatically, where would we expect to see it first?
Not necessarily in aggregate unemployment figures. Those are blunt instruments. They blend industries, age groups, occupations, and business cycles into a single number. A structural shift concentrated in a particular corner of the workforce could be invisible in the headline statistics for years.
The Anthropic Report Massenkoff and McCrory (2026) focuses on precisely the question of where to look and what counts as signal. Its answer to the first question is: among younger workers in highly exposed occupations.
An Early Warning
The paper’s most arresting finding concerns job entry rates for 22-to-25-year-olds. Using the panel dimension of the Current Population Survey, the authors track the monthly rate at which young workers begin new jobs in high-exposure versus low-exposure occupations. Before 2024, the series move together. After 2024, they diverge. Entry into high-exposure jobs decreases by roughly half a percentage point per month. Entry into low-exposure jobs holds steady at around 2% per month.
Across the post-ChatGPT period as a whole, the paper estimates a 14% drop in the job-finding rate for young workers entering highly exposed occupations. This result is just barely statistically significant, and the authors are appropriately cautious about alternative interpretations. Some young people may be delaying entry into the labour market. Some may be redirecting to less exposed fields. Some may be returning to education.
But the pattern is consistent with a separate and more statistically robust finding from Brynjolfsson, Chandar and Chen (2025) at Stanford’s Digital Economy Lab. Using payroll data from ADP covering millions of workers, they document that employment of 22-to-25-year-old software developers fell by nearly 20% between late 2022 and mid-2025. Customer service representatives in the same age group show a similar pattern. Workers aged 35 and over in the same occupations show no such decline.
The canary is not dead. But it looks uncomfortable.
What the Unemployment Data Say (and Do Not Say)
The Massenkoff and McCrory paper’s main finding on unemployment is a non-result, and it is important to read that carefully. They find no statistically significant increase in unemployment for workers in the top quartile of observed exposure since the release of ChatGPT in late 2022. The gap between exposed and unexposed workers has remained flat.
This does not mean nothing is happening. It means the kind of stark, broad-based unemployment shock we would expect from a major economic crisis has not materialised. The authors estimate that a scenario equivalent to the Great Recession doubling unemployment in the most exposed group would be visible in their data. That has not happened.
What may be happening instead is something slower and harder to see: not mass layoffs of existing workers, but a gradual reduction in the intake of new ones. Separations are not rising sharply. Hirings are softening at the entry level. This is how a technological substitution can take hold quietly, without triggering the sort of spike that makes it onto the front page.
Why Entry-Level Matters for Education
From an educational perspective, this is where the Anthropic findings have the most immediate relevance. Computing graduates and students currently in training are not competing for mid-career positions. They are competing for the entry-level roles that AI appears to be displacing first.
A sustained slowdown in entry-level hiring for software developers and customer service roles does not represent some abstract risk in the future. It is already shaping the graduate employment landscape. The Financial Times reported in 2025 on growing concerns about graduate unemployment in AI-exposed fields, particularly in software and data roles.
This raises questions that education systems are not yet set up to answer. If AI automates the tasks that junior workers learn on the job, how does the next generation acquire the experience and contextual judgement that more senior roles require? The argument that exposed workers will move up the value chain assumes there is a pathway from entry-level to senior work. If AI forecloses the entry point, that pathway may not exist.
Noy and Zhang (2023), in their experimental work on AI and productivity in Science, showed that AI tools raised the output quality of lower-skilled workers more than that of higher-skilled workers. This is consistent with the substitution hypothesis: AI is most valuable where it can replace the work that is currently done least well, which is often the work done by novices.
The Augmentation Question
Not all exposure is equal. One of the more hopeful threads in the Brynjolfsson, et al. (2025) research is the distinction between automative and augmentative AI use. In occupations where AI primarily augments human work rather than substituting for it, early-career employment is more stable. This distinction is drawn from the same Anthropic Economic Index data that Massenkoff and McCrory use.
The implication is that how AI is deployed matters as much as whether it is deployed. Design choices made by employers, software developers, and platform providers will shape whether AI becomes a tool that empowers entry-level workers or one that replaces them before they have had a chance to develop.
These are not purely market decisions. They are choices that education and training policy can help to shape, through the skills that graduates bring to the workplace and through the expectations we build about what responsible AI deployment looks like.
We should be watching the data closely.
We should be asking harder questions now,
before the pattern is no longer ambiguous.
Thanks for reading Data in Motion.
Subscribe for free to receive new posts and support my work.
Watching the data is necessary. But data only measures what the current system values. If entry-level coding jobs decline by 14%, that shows up as a problem — because we measure employment. What doesn't show up: the 19-year-old in Manila who could never afford a CS degree but can now build production software with AI tools in a weekend. The retired teacher with a novel she never had time to write, producing it in a month. The kid with cerebral palsy whose mind contains symphonies he could never play — until now.
Every previous technological revolution liberated a physical capacity. The printing press freed knowledge from monasteries. The engine freed movement from muscle. This one is different. It liberates cognition from circumstance. For the first time in history, the gap between what you can imagine and what you can produce is collapsing.
The entry-level funnel is narrowing for tasks that were never the point. They were the hazing ritual we confused for training. The real question isn't how do we preserve the old pathway in. It's what happens when the pathway is no longer necessary — when a mind that couldn't get through the door can now build the building.
We've never had data for how many Hawkings were trapped in bodies, circumstances, or zip codes that never let them reach a keyboard. AI doesn't just displace workers. It liberates creators who were never counted in the workforce to begin with.