Minus 19 or Plus 12 Percent? Why AI Studies on Entry-Level Jobs Disagree So Much

What this is about

Within a few weeks, several major analyses have come out on the same question: Is AI costing entry-level workers their jobs? The headlines sound like they're from different worlds. The Stanford Digital Economy Lab reports a 19 percent gap for the US. The IAB, the research institute of Germany's Federal Employment Agency, finds 1.2 percentage points in Germany and gives the "all-clear, with caveats". On November 9, the Reserve Bank of Australia (RBA) sees "tentative signs", and a company analysis by Ramp and Revelio Labs even counts 12 percent more entry-level hires at AI-intensive companies.

Our argument: these numbers contradict each other less than it seems. They come from different data and describe different age groups. Each study also defines "AI exposure" differently, and some calculate relative figures while others use absolute ones. Put the studies side by side, and you can explain much of the spread from −19 to +12 percent.

The studies at a glance

First difference: Who counts as "young"?

Stanford looks at 22- to 25-year-olds, so almost exclusively true career starters. The IAB groups 20- to 29-year-olds together and itself calls this an approximation of the early career phase. That group also includes people who've been working for five or eight years and who, based on the US findings, would barely be affected. A wider age band inevitably dilutes the effect. The RBA, by contrast, uses 15 to 24, which also sweeps in students' part-time jobs.

A comparison within Germany shows how much this matters. The ifo Institute also works with Federal Employment Agency data. But it looks only at 20- to 24-year-olds and only at the most AI-exposed occupations. There, according to ifo, full-time employment fell from around 279,000 to 239,000. By our math, that's 14.3 percent. For 20- to 29-year-olds, the IAB finds −3.8 percent in highly exposed occupations and −2.6 percent in low-exposure ones. So with the same raw data, depending on how you slice it, you get either a 14 percent decline or a gap of just over one percentage point.

Second difference: What does "AI-exposed" mean?

Stanford and the IAB use the same measure from Eloundou and colleagues (2024). It measures exposure, meaning the share of an occupation's tasks that a language model could theoretically complete much faster. It says nothing about whether companies actually use AI. The RBA uses the Australian measure from Jobs and Skills Australia, which is based on an ILO study. That, too, only describes potential. Mike Major stresses that the study measures "only theoretical exposure, not actual use in businesses".

Ramp/Revelio, on the other hand, measure money companies have demonstrably spent on AI tools. That's a completely different question: Stanford asks how occupations that AI could do are faring. Ramp asks how companies that buy AI are faring. On top of that, there's a selection effect: companies that spend heavily on AI tend to be large, tech-savvy, venture-backed and growing anyway. As Forkast reports, Ramp economist Ara Kharazian says so himself:

"Who funded you predicts AI adoption better than what industry you're in."

The authors explicitly write that their work doesn't show AI is creating jobs everywhere. Almost all the gains were in the tech sector. Companies that only had subscriptions and pilot projects didn't grow.

A counterpoint from Harvard is interesting here. Hosseini and Lichtinger also measure AI adoption at the company level, using job ads for "AI integrators". Their basis is the résumés of around 62 million workers at 285,000 companies. They find the opposite of Ramp: starting in the first quarter of 2023, junior employment at adopting firms dropped significantly, while senior employment stayed stable. Reports differ on how big the effect is, with figures ranging from 7.7 to 12 percent. So how you measure an "AI company" can flip the sign.

Third difference: Relative gap, absolute decline or deviation from trend

Stanford's 19 percent isn't a count of people laid off. It's a counterfactual comparison, answering the question: How many fewer young people work in exposed occupations than if they had followed the same path as their peers in low-exposure occupations? According to Stanford, employment in the two most exposed fifths (quintiles) fell by about 11 percent, while in the three least exposed it rose by about 10 percent. Run the numbers, and 0.89 divided by 1.10 gives a gap of 19 percent. Apply the same calculation to the IAB figures (0.962 divided by 0.974), and you get −1.2 percent. So the difference isn't down to the math, but to what actually happened.

The time period matters too. The IAB data ends in late 2025, Stanford's in June 2026. For July 2025, the Stanford revision puts the gap at 15 percent. The first version from August 2025 had said 13 percent; the data has since been revised. So even at roughly the same point in time, the US figure is a good ten times the German one. That's a real difference between the countries, not a measurement artifact. But it's smaller than "19 versus 1.2".

Goldman Sachs measures something else again: the deviation from trend of entire industries across all age groups. So the −27 percent at German call centers says nothing directly about entry-level workers. According to City AM, Goldman finds that every additional 10 percent of exposure is associated with 0.1 percentage points less annual headcount growth. For entry-level jobs, it's more than 0.2 percentage points in the US and more than 0.6 in Australia.

The confounders: Interest rates, remote work, the economy

The strongest argument against AI as the main cause is timing. A team led by Morgan Frank (University of Pittsburgh) found in US unemployment insurance data that risk in exposed occupations was already rising in early 2022. That was months before ChatGPT. The Economic Innovation Group points to the interest rate turnaround as an overlooked cause, since tech companies are especially rate-sensitive. Using data from the official Current Population Survey (CPS), the New York Fed estimates that remote work explains 64 percent of the rise in unemployment among young graduates. The authors themselves call this a back-of-the-envelope calculation. Their argument: when people work remotely, it's harder to train newcomers.

Stanford explicitly tested these objections. The gap remains when you exclude tech companies and account for interest rate sensitivity and suitability for remote work. But the authors also concede that some trends began before ChatGPT, and that the gap shrinks when you control for education. In addition, the figures in the ADP sample come out larger than in the official surveys. The Stanford authors partly attribute this to changed weights in the Census Bureau's ACS survey.

The IAB factors the economy in directly, capturing how heavily each occupation was represented in industries and regions with mass layoffs. The result: young workers in weak sectors lost 4.2 percent of their jobs between 2022 and 2025. According to the IAB summary, this effect is more than twice as strong as the AI correlation. AI only becomes significant in manufacturing (around −6 percent in 2025) and in non-academic occupations. That's notable, because the public debate revolves around university graduates. IAB researcher Michele Dinelli therefore draws attention to "occupations with a non-academic profile and manufacturing". The RBA also cautions that youth employment was unusually high in 2022 because of the overheated labor market, so some of the cooling is normal. Youth unemployment rose from about 7 to about 11 percent, but according to the RBA it's still below pre-pandemic levels.

What still fits together

Behind the different numbers, a shared picture emerges. No study finds broad AI-driven job losses, not even Stanford. Where an effect shows up, it works through fewer new hires, not layoffs. That's what Stanford, the Harvard paper and the Dallas Fed show: in the Dallas Fed's data, the missing people are mostly those moving directly into a job from outside the labor market. In March 2026, Anthropic found that young people's chances of landing a job in exposed occupations had dropped by about 14 percent, though this was only barely significant. The RBA likewise sees the sharpest declines in job ads for clerks, programmers and call center workers. In the UK, Indeed job ads for graduates in July 2026 were around 7 percent below the previous year and at their lowest since 2020. Whether AI is to blame is disputed there.

A Danish study by Humlum and Vestergaard covering 25,000 workers found no measurable effects on wages or hours worked. But it looks at people who are already employed, not at entry into the workforce. So it answers a different question.

Our conclusion

The spread from −19 to +12 percent is largely explained by the studies measuring different things. Sometimes it's occupations, sometimes companies. Sometimes theoretical potential, sometimes actual spending. Sometimes 22- to 25-year-olds, sometimes 20- to 29-year-olds. Sometimes relative gaps, sometimes industry trends. Ramp isn't measuring the opposite of Stanford; it's measuring something else.

Our take: part of the difference between the US and Germany is still real. Even at the comparable point in mid-2025, the US gap was 15 percent and the German one around one percentage point. Possible reasons include faster AI adoption in the US, the larger tech sector and Germany's vocational training system. None of this is proven. The ifo figure of −14 percent also shows that Germany is seeing significant shifts in the youngest group and in the most exposed occupations. No study can yet separate how much of that is due to AI and how much to the recession.

The critical point remains: all of these studies are descriptive, and none proves causation. Most measure potential rather than usage, the data lags by months, and the economic situation in Australia (five rate hikes in 2026) and Germany overshadows everything. If you cite a single number as proof, in either direction, you're making it too easy for yourself. Still, it's worth keeping an eye on, as IAB vice director Melanie Arntz advises: beyond 2025, and ideally with data that links actual AI use in companies to employee records.