The usual career advice meets new payroll data
Ask how to stay employable as AI takes on more work and the usual answer is to go deep in one field first and add breadth later. In August 2026, Stanford University's Digital Economy Lab published payroll figures that put pressure on the first half of that advice. They don't do it the way most people would guess.
Polora gave those figures to a group of AI models in a staged experiment. Three models took roles. gpt-5.6-sol played a labor market economist, gemini-3-8-flash a career risk strategist and grok-4.5 an organizational capability architect. A second gpt-5.6-sol seat worked as a researcher, checking claims against sources on the web, and claude-sonnet-5 moderated and gave the closing verdict.
They were asked what current evidence says about going deep versus staying broad. They were told to be explicit about what the numbers cannot show. Then they had to answer for three people : someone choosing a first career, someone ten years into one specialty, and someone whose job is coordinating work across several fields.
Young workers in AI-exposed jobs are 19% behind, which is not the same as 19% fewer
Stanford followed records from the payroll company ADP through June 2026. Among 22- to 25-year-olds in the occupations most exposed to AI, employment was about 19% below where it would have been if it had kept pace with people the same age in less-exposed occupations.
That 19% is the distance between two paths. It is not a drop. In plain counts, from November 2022 to June 2026, employment of young workers fell about 11% in the most exposed and second-most exposed fifths of occupations. In the other three fifths it grew about 10%. The researcher model checked both figures against Stanford's page.
Employers are hiring fewer young people, while experienced workers in the same jobs hold steady
Most of the gap came from weaker hiring. It did not come from more people leaving or losing their jobs. Experienced workers in the same exposed occupations show no comparable gap. The work and the exposure are the same. What differs is age, which roughly tracks whether someone is trying to get in or is already inside.
Stanford also reports that the decline was concentrated where AI is used to do tasks in place of people. Where AI mainly assists people, it found no clear link to weaker employment.
A US Census Bureau working paper from April 2026 reached a similar finding by a different route. It compared groups defined by industry and state rather than occupations, and looked at 22- to 24-year-olds. In the most exposed fifth of those groups, their adjusted employment fell 12% over ten quarters, while it held stable in less-exposed groups, and again hiring was the main channel. The economist model had first cited this paper at about 9%. The researcher model checked the paper and reported that its headline figure is 12%, not 9%.
These numbers cannot tell you whether specialists or generalists are safer
Neither study sorted anyone by how deep or broad their skills are. They grouped workers by age and by how exposed their occupation is to AI, and age only roughly reflects experience. So the data cannot show that specialists are losing ground or that generalists are protected. Every participant in the experiment said so.
The data also do not prove that AI caused the gap. The Stanford authors call their findings descriptive, not causal. The gap narrows once education is taken into account, some of the diverging trends began before generative AI spread, and Stanford's estimates run larger than some national surveys. Their sample of about 25,000 firms does not represent the whole US labor market. The researcher model confirmed that the divergence survives several exclusions and controls. According to the economist and architect models, these include leaving out tech firms and checking for sensitivity to interest rates and to remote work. The Census paper estimates that monetary policy could explain up to a quarter of its employment decline, though not the sharp drop in hiring.
The researcher model summed up what the evidence does support. Since late 2022, young workers in AI-exposed work have seen unusually weak hiring and employment compared with young workers elsewhere, and older workers in the same work have not. AI is a plausible contributor, but how much of the gap it explains has not been isolated.
Depth may still pay, but the first rung is harder to reach
The economist model argued that "go deep first" assumed a first job where beginners did routine work, such as standard code, summaries, translations and basic analysis, and turned it into experience. It pointed out that this kind of work overlaps heavily with what AI can do now.
Put that next to the payroll figures and a narrower reading appears. The older workers in exposed jobs, who are more likely to have built depth, are not the ones falling behind. The ones falling behind are the young, most of whom are trying to start. On that reading, the risk is not that depth stopped paying. It is that the job where depth used to be built is harder to get.
The strategist model put it more bluntly, saying the bottom rungs of the ladder had been sawed off. The researcher model called this a plausible interpretation, not a demonstrated fact. The data show weaker hiring. They do not show why firms hired fewer people, whether senior staff are using AI in place of juniors, or whether the young people affected moved into other occupations.
All three models gave the same answer, and the data did not test it directly
The three role-playing models spoke in turn, and the later two said they were building on what came before. They reached the same answer. The economist and architect models called it a barbell, and the strategist model described it as one hard anchor skill joined to an interface with customers, operations or regulation. One end is enough depth to judge the work and be answerable for it. The other is enough breadth to move as the boundaries of tasks shift. The economist model split depth into three kinds : routine production, judgment built from experience and from answering for decisions, and the skill of framing problems and checking AI output. It argued that depth in routine production is the kind losing value faster. It also warned that knowing a little about five fields is no refuge if you cannot judge quality in any of them.
The moderator did not treat their agreement as evidence. In its verdict it called the convergence the right instinct and the best current inference, but said it is the participants' best reasoning applied to data that does not test it directly. It added that a different reading fits the same numbers just as well : employers are automating routine entry-level tasks, and that says nothing general about breadth. The moderator noted that the conversation never seriously tested that reading.
What the moderator said to the three people in the question
For someone choosing a first career, the moderator said the safest reading of the evidence is not to go broad instead of deep. The point is to avoid depth made only of tasks a model already does. It suggested fields where beginners still get supervised exposure to unclear, high-stakes decisions, paired early with one related skill. It called this a hedge against a plausible but unproven risk, not a guarantee.
For someone ten years into one specialty, it said the data offer no evidence of risk, since the gap is among the young. Its suggestion was to work out what, in your own case, might explain that protection, such as accountability, experience-based judgment or relationships, and to strengthen whichever one really applies, rather than assume seniority alone is enough.
For someone who coordinates work across fields, it said the role looks safer on current evidence, but only because no one has shown coordination itself to be exposed. That is an inference from a gap in the data, not a finding. Earlier, the economist model had argued that coordination is secure only when it is more than forwarding information and writing summaries.
The question the data leave open
One thing to take away fits in a sentence. The new jobs data do not rank specialists against generalists. They show that young people in AI-exposed jobs are being hired less than their peers in other jobs, while older workers in the same jobs are holding steady.
That raises a question the payroll figures cannot answer yet. The architect model warned that firms that permanently cut entry-level hiring without another way to train beginners will later face a shortage of mid-career judgment. That was its own argument, and nothing in the record has tested it. The moderator closed on a wider open question : whether depth or breadth matters at all once you account for which work AI is doing in place of people and which work it is only helping with. Neither the data nor this conversation has answered it.
