The most important AI labor story may not be a wave of dramatic layoffs. It may be the job that never gets posted.

An updated Stanford study using payroll records from millions of U.S. workers finds a widening employment gap for workers ages 22 to 25 in occupations highly exposed to generative AI. Their employment now stands about 19 percent below where it would be if it had kept pace with similarly aged workers in less exposed occupations.

That number needs to be read carefully. It does not mean AI has erased 19 percent of all entry-level jobs, and the researchers explicitly say their findings are descriptive rather than proof that AI caused the entire gap. What it does show is that the divergence first spotted in 2025 has continued to grow through mid-2026.

The gap has widened, not disappeared

The research comes from Stanford Digital Economy Lab economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen. Their revised working paper uses anonymized payroll data from ADP through June 2026.

Last year's version found a significant relative decline for young workers in AI-exposed fields. The updated data show the simpler employment gap widening from about 15 percent in the July 2025 data vintage to 19 percent by June 2026.

In raw levels, employment for 22 to 25-year-olds in the two most AI-exposed groups fell about 11 percent between November 2022 and June 2026. Employment for workers of the same age in the three least-exposed groups rose about 10 percent over the same period.

Older workers do not show a comparable pattern. That is one reason the result is getting attention. A broad economic slowdown should usually leave fingerprints across more age groups. Instead, the weakness is especially visible among people trying to get their first few years of professional experience.

It looks more like reduced hiring than mass firing

The Stanford team says the adjustment is happening primarily through lower hiring of young workers, not a surge in separations. In other words, companies may be leaving some junior seats empty rather than replacing entire departments with a chatbot and sending out a celebratory press release.

That fits how businesses often adopt new technology. A company does not necessarily fire three junior analysts because an AI tool can handle part of their workload. It can simply hire one analyst next year instead of three.

The distinction matters because reduced hiring is harder to see in conventional layoff statistics. A worker who loses a job appears in employment data immediately. A graduate who never gets the opening that would have existed three years ago is much harder to count.

Automation and augmentation are producing different outcomes

The researchers also separated occupations where AI use is more likely to automate tasks from those where it tends to complement human work. Employment declines for young workers were concentrated in the automation-heavy group. In jobs where AI is used more as a tool alongside workers, employment was flat or rising, particularly for experienced employees.

The study also finds a split between codified and tacit knowledge. Codified knowledge is the kind that can be written down, standardized, and taught through documents or formal training. Tacit knowledge is the judgment people accumulate through practice, context, and repeated exposure to real situations.

Young workers in jobs built heavily around codified knowledge have seen weaker employment. Experienced workers in occupations that depend more on tacit knowledge have generally held up better. That is uncomfortable news for anyone whose career ladder traditionally began with routine research, drafting, analysis, or support work, because those tasks are exactly where generative AI tends to be most capable.

Fresh ADP data show the trend is still alive

ADP Research's July update adds another month to the picture. Employment in occupations with high AI exposure rose just 0.1 percent from a year earlier, compared with 1.1 percent growth in the least-exposed occupations.

Among workers ages 22 to 25, employment in highly exposed jobs was down 3 percent year over year. ADP says that marked 34 consecutive months of contraction for that group. The dashboard update also showed employment for young workers overall down 1.3 percent from a year earlier.

At the same time, neither Stanford nor ADP finds evidence of economy-wide AI job destruction. That is an important counterweight to the more apocalyptic version of this story. The effect so far looks uneven, age-sensitive, and concentrated in particular types of work.

The researchers are careful about causation

The paper tests several alternative explanations and finds that the broad pattern survives when technology companies and computer occupations are excluded, and when the analysis accounts for remote work and exposure to interest-rate changes.

There are still limitations. Some differences between highly exposed and less exposed occupations began before the generative AI boom. The results become smaller when education is controlled for, and the employment gap is more pronounced in the ADP analysis sample than in some national survey benchmarks.

That is why the authors call the results early indicators rather than a causal estimate of exactly how many jobs AI has removed.

Even with those caveats, the direction is becoming harder to dismiss. The entry-level labor market is weakening most where AI can do a meaningful share of codified junior work, while experienced workers remain much more insulated. If that persists, the long-term problem will not just be fewer junior jobs. Companies may eventually discover that senior employees do not spontaneously materialize at age 35 with ten years of experience. The career ladder still needs a first rung.