The first visible AI jobs shock may not look like a layoff announcement. It may look like a vacancy that was never posted.
That is the unsettling possibility raised by a revised Stanford Digital Economy Lab analysis of US payroll records through June 2026. Among 22-to-25-year-olds, employment in occupations highly exposed to generative AI was 19 per cent below where it would have been had it kept pace with less-exposed work.
The gap appears to have opened mainly because employers hired fewer young people into exposed occupations. Experienced workers did not show a comparable shortfall. There was no economy-wide jobs collapse hiding underneath the result.
It is a narrower finding than “AI is taking everyone’s job.” In one respect, though, it may be more consequential. If organisations keep their senior staff but stop replenishing the bottom of the ladder, the immediate numbers can look calm while the supply of tomorrow’s senior analysts, lawyers and coders quietly thins.
The 19 per cent is a relative gap
The headline number needs careful handling. It does not mean that 19 per cent of all young people in AI-exposed jobs lost their employment.
In the researchers’ revised paper, employment among 22-to-25-year-olds in the two most AI-exposed occupational groups fell by about 11 per cent between November 2022 and June 2026. Employment for the same age group in the three less-exposed groups grew by roughly 10 per cent. Comparing those two paths produces a 19 per cent “kept pace” shortfall.
That distinction matters. Calling it a 19 per cent absolute fall would overstate what the data show.
Silicon Canals covered an earlier version of this research, when the best-known estimate was a 13 per cent relative decline after statistical controls and firm-level comparisons. The new 19 per cent figure is not simply that number with another year added. It is a more direct comparison of two diverging employment paths, using data through June 2026.
Different measure, later window, same worrying direction.
The wider labour market still looks ordinary
Nothing in the updated analysis resembles mass technological unemployment. Employment in the study’s balanced sample of ADP client firms rose by about 6 per cent from November 2022 to June 2026. Across all ages, employment in the most AI-exposed group of occupations rose by around 4 per cent.
That is part of what makes the finding easy to miss. Aggregate employment can grow while access to a particular kind of first job deteriorates.
A company does not need to dismiss an experienced analyst for AI to alter its workforce. It can keep the analyst, give her a model that drafts summaries and checks formulas, then decide that the next graduate position is no longer urgent. The team has adjusted. No redundancy has appeared in the statistics.
Repeat that decision across enough firms and the effect accumulates at the entrance, even while the building above it remains occupied.
This looks more like a hiring shock than a firing shock
The Stanford team examined hires and separations separately. The widening gap was driven primarily by reduced hiring of young workers, not a surge in young-worker departures or a purge of experienced staff.
A separate US Census Bureau working paper, published in April 2026, found a closely related pattern in different administrative data. In the industry-state groups most exposed to AI, hiring of 22-to-24-year-olds dropped sharply after ChatGPT’s release. The paper estimated that early-career employment in the most exposed group was 12 per cent lower after ten quarters, with the reduction in hires doing most of the work.
Two studies do not settle the question, and their units and methods are not identical. Their agreement on the mechanism is still notable.
Hiring shocks are socially quieter than layoffs. There is no factory gate, public list or single announcement. A graduate simply applies to a smaller intake. A contract role is not renewed as a permanent position. A team gets approval for one experienced hire instead of two trainees.
The people affected may never know which opportunity disappeared.
What “AI-exposed” actually means
Exposure is not the same thing as replacement. The researchers rank occupations by how readily a language model could perform or accelerate the tasks inside them. A highly exposed job can still contain responsibility, relationships, local knowledge and judgement that a task-based score does not capture.
The revised Stanford work addresses that problem in two ways. It uses both expert ratings of potential language-model exposure and patterns from Anthropic’s Economic Index, which records how Claude is actually used. It also distinguishes activity that looks automating from activity that looks complementary to a worker.
The young-worker declines were concentrated in occupations where the observed use of AI appeared more substitutive. In complementary occupations, employment was flat or rising, especially among experienced workers.
The result also divides along types of knowledge. Young employment weakened most in roles built around codified knowledge: procedures that are formal, documented and relatively easy to check. Employment held up better where work depended on tacit knowledge accumulated through context and experience.
This is not proof that experience itself makes someone safe. It suggests that current models fit more neatly into work where the rules can be written down, which often happens to be the work handed to a beginner.
The first rung was doing two jobs
Entry-level work has always had two functions. It produces something the organisation needs today, and it trains someone the organisation may need later.
A new analyst cleans awkward datasets before learning when a clean-looking number is misleading. A junior lawyer reviews standard documents before recognising the clause that is standard in form but dangerous in context. A programmer fixes bounded bugs before being trusted to make architectural trade-offs.
Those tasks can be repetitive. Some are exactly the kind of codified work a language model handles increasingly well. But repetition is also how people build pattern recognition, encounter exceptions and learn when the written procedure is not enough.
If a firm removes the routine task and then removes the trainee attached to it, the short-term saving can be perfectly rational. The longer-term effect is less tidy. Where does the senior worker with tacit knowledge come from if fewer people are allowed to acquire it?
The Stanford paper does not model that future pipeline. It would be a mistake to present the answer as one of its findings. The data nevertheless expose the question in unusually concrete form.
The novice productivity paradox
The hiring result sits awkwardly beside another body of workplace research. As we have previously examined, several studies have found that generative AI delivers its largest productivity gains to novices and lower-performing workers. A tool that makes expert patterns easier to reach can help a beginner improve quickly.
In principle, that should make junior employees more attractive. A person who once needed a year to reach acceptable output might become useful sooner.
But productivity does not mechanically determine employment. A firm can use the gain to produce more with the same number of people. It can lower prices, shorten turnaround times or improve service. It can also decide that a smaller cohort can handle the previous workload.
This is why the words “augmentation” and “automation” are not properties of a model alone. They also describe a management decision. The same system can be used as a patient tutor that helps a trainee compare alternatives, or as a production shortcut that removes the reason to hire the trainee at all.
Google’s large ATLAS study, which Silicon Canals covered in August, found that most observed non-routine cognitive interactions looked collaborative rather than end-to-end automation. That is encouraging. It does not tell us who gets invited into the collaboration.
Why firms can underinvest in the next generation
There is an old tension inside apprenticeship. The firm bears the cost of training a beginner, but the experienced worker who eventually emerges can leave. That already gives employers an incentive to let someone else pay for the first few years of development.
AI may sharpen the temptation. A senior-heavy team can use software to clear routine work and deliver this quarter’s output without carrying a large training cohort. One company can plausibly make that choice. If many do it together, the market may later discover that it has preserved expertise without reproducing it.
This is an inference from the employment pattern, not something the payroll records directly measure. We do not yet know whether smaller entry cohorts will persist, whether redesigned roles will replace them or whether demand for AI-enabled services will create enough new work to reopen the ladder.
We do know that seniority is not a raw material. Organisations manufacture it slowly, through exposure to real cases, feedback, mistakes and responsibility.
There are serious reasons to be cautious
The researchers are unusually direct about the study’s limits. This is descriptive evidence, not a controlled experiment proving that generative AI caused the entire gap.
The period after November 2022 also included higher interest rates, a post-pandemic correction in technology hiring and changes in remote work. Some of the divergence visible in the data began before generative AI became widely available. When the authors add education controls, parts of the age gap become smaller.
The ADP data are large, covering roughly 3.5 million to five million workers a month in the balanced panel, but they are not a perfect miniature of the US workforce. Manufacturing, wholesale industries and larger firms are overrepresented; retail and hospitality are underrepresented. Around 30 per cent of records lack a usable job title, requiring the researchers to impute occupational exposure.
The estimated gaps are also larger in the ADP sample than in broad national benchmarks. The authors’ robustness checks matter: the pattern survives removing technology firms and computer occupations, and it remains under alternative exposure measures. Those tests make a simple dismissal harder. They do not erase the limits.
Base pay, meanwhile, showed no comparable divergence. That suggests employment quantities moved before salary rates, although the payroll data do not capture every form of compensation.
A warning light rather than a verdict
The 19 per cent gap is not a forecast that every junior knowledge job will vanish. It does not mean young people should abandon software, law, finance or analysis. The Stanford team explicitly declines to say that the trend will continue.
It is better read as a warning light. A healthy-looking labour market can conceal a narrowing path into certain occupations, and a technology that helps current workers can still reduce the number of new workers given a chance to learn.
Companies retain choices. They can redesign junior roles around verification, customer contact, model evaluation and supervised judgement. They can use AI to expose beginners to more examples while keeping a human responsible for feedback. They can measure progression and bench strength, not just output per employee.
None of that requires preserving every old task for nostalgia’s sake. Plenty of entry-level drudgery deserves to disappear. The challenge is to separate work that was merely cheap labour from work that quietly functioned as practice.
The first rung of the career ladder was never glamorous. If firms remove it without building another route upward, though, the damage will not end with the graduate who misses a job in 2026. It will surface years later, when everyone wants to hire an experienced person and too few people were given the chance to become one.
