The first rung: what happens when entry-level jobs disappear

Employment of 22-to-25-year-olds in the jobs most exposed to AI is now 19% below trend, according to Stanford researchers who first put the gap at 13% in 2025. For centuries, crafts and professions renewed themselves the same way: someone senior let someone junior do the simple work badly, then less badly, until the junior became senior. That bottom rung is now thinning. In Germany, the Ifo institute counts 14% fewer young full-time workers in AI-heavy occupations. If you remove the first rung, the ladder still stands for a while. Then you notice nobody is climbing it.

Contents · 9 sections

TL;DR #

• Stanford’s payroll-based study now finds employment of 22-to-25-year-olds in the most AI-exposed jobs 19% below trend (data to June 2026), up from the 13% first reported in August 2025. Older workers show no such gap.
• In Germany, Ifo counts 14% fewer full-time workers aged 20 to 24 in AI-heavy occupations. The IAB finds no significant AI effect, and both say causation is not proven.
• The drop comes mainly from fewer hires. That removes the junior work through which people have always learned the tacit skills that are still in demand.
• A firm that hires far fewer juniors today will have a thin mid-career layer in the 2030s, and will pay for it then.

What the numbers say, and how sure they are #

The headline study is “Canaries in the Coal Mine?” by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab, based on ADP payroll data. The figure most people quote, 13%, is from the first version in August 2025. It has been revised since, so it is worth being precise:

• August 2025: a 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations, after controlling for firm-level shocks.
• August 2026 revision, data to June 2026: employment of 22-to-25-year-olds in highly exposed occupations is 19% below where it would be if it had kept pace with similar young workers in less-exposed jobs. In plain levels, it fell about 11% from November 2022 to June 2026, while the least-exposed groups grew about 10%.
• Experienced workers show no comparable gap. The adjustment comes mainly from fewer hires, not more firings.

The authors are careful, and so should I be. They call these “descriptive patterns, not causal estimates of the effect of AI.” They note that the gaps are larger in ADP’s sample than in national surveys, and that some trends started before ChatGPT. So: documented as a pattern, not proven as a cause.

Germany now has its own version. On 9 November, the Ifo institute published a new AI early indicator covering 140 occupations. Full-time employment of 20-to-24-year-olds in the most AI-exposed jobs fell 14%, from about 279,000 to 239,000, since late 2022. LSN News, my news engine, counted 28 articles from 27 German-language outlets on that one story within days. Oliver Falck, who leads Ifo’s innovation centre, put it in one line: “Der Berufseinstieg ist das Frühwarnsystem des Arbeitsmarkts”, career entry is the labour market’s early-warning system.

There is a counterweight. The IAB, the German federal employment research institute, found in October that the AI effect on young job seekers was not statistically significant, and that the business cycle hit young workers more than twice as hard as any possible AI effect. Ifo itself says its indicator does not prove AI caused the drop. My honest reading: the first rung is thinner, AI is one plausible cause among several, and weak demand is clearly another.

The venture firm SignalFire adds a sharper industry signal. In its 2025 talent report, new graduates were just 7% of hires at the largest tech firms, and new-graduate hiring there was more than half below 2019.

How crafts used to make masters #

Why does the bottom rung matter so much? Because that is where expertise was always made.

In my library, Ernst Gombrich’s A Little History of the World describes the medieval guilds in a sentence a child can follow: “Not just anyone could become a master-tailor. First you had to serve your time as an apprentice.” Then came years as a journeyman, and then a masterpiece judged by the guild.

The Cambridge Handbook of Expertise and Expert Performance, edited by K. Anders Ericsson, also in my library, takes the same history seriously. It describes the craft guilds, from around the tenth century, as “a skills-based, performance-assessed, and domain-specific learning community.” The point was not the certificate. The point was thousands of hours of supervised practice on real work, with feedback from people who had done it for decades.

Modern professions kept the shape even when they dropped the name. Junior lawyers do document review. Junior accountants check the ledgers line by line. Junior developers fix small bugs and write tests. Junior editors check references. It is dull, it is cheap to automate, and it is exactly the work through which people learn what “right” looks like.

The Stanford update makes this point in its own language. It separates codified knowledge, the kind written in textbooks and manuals, from tacit, experience-based knowledge. Jobs built on codified knowledge are the ones shrinking for young workers. Jobs built on tacit knowledge are growing for experienced ones. Here is the catch: you get tacit knowledge by doing the codified work for years. Remove that work and you remove the path to the skill that is still in demand.

What downsizing did to knowledge last time #

We have seen a version of this before, from the other end of the career.

In my library, Bryan Bergeron’s Essentials of Knowledge Management tells how Northrop Grumman started its knowledge-management programme in the 1990s defence downturn. After massive layoffs among the people who designed and built the B-2 bomber, the firm sent knowledge engineers to capture what it could of “the irreplaceable knowledge that was walking out of the door”, from employees “who had already been given pink slips.”

Think about that scene: interviewing people on their way out, trying to write down what took them twenty years to learn. Bergeron is blunt about timing as well. He notes that many knowledge-management initiatives fail because they run in parallel with re-engineering, and that they work best “at a time of corporate stability.”

Wayne Cascio, in his 2009 guide for the SHRM Foundation, explains why the damage is bigger than the headcount suggests. In knowledge-based firms, relationships and networks hold the institutional memory, so “indiscriminate downsizing has the potential to inflict considerable damage on learning and memory capacity.” He lists software, biotech and pharma among the industries at greatest risk.

The 1990s lost knowledge at the top of the ladder. The 2020s risk losing it at the bottom, which is slower to notice and slower to fix.

What a firm loses in ten years #

Here is a simple piece of arithmetic. It is an illustration, not a forecast.

Suppose a firm hires 30% fewer juniors each year from 2023 to 2027. Nothing breaks in 2027. The seniors are still there, and AI tools handle much of the junior work. But:

• By 2030–2032, the firm has a thin layer of staff with three to five years’ experience. That is where team leads and reviewers usually come from.
• By 2033–2037, the group with eight to twelve years’ experience, the people who become principal engineers, senior editors and partners, is roughly 30% smaller than it would have been, unless the firm buys them from competitors.
• At the same time, today’s seniors are retiring. The people who could have trained the next generation leave without anyone to hand over to.

Every firm that did the same thing will be bidding for the same small cohort. Mid-career salaries rise. The firm that saved on graduates in 2025 pays it back, with interest, in 2035. Some firms will notice sooner: the SHL survey that found 19% of US workers had been replaced by AI and then asked back also found that young workers were the most affected.

There is a quieter loss too. Managers who train people need narrower teams. As firms flatten, with fewer managers and more direct reports each, the time for teaching shrinks just as there are fewer people to teach.

Checklist #

If I ran a team, I would keep hiring juniors, fewer perhaps, but deliberately, and change what the first rung looks like:

• Make the junior job a review job. Let the AI draft; let the junior check it against the source, with a senior checking the junior. That is still supervised practice on real work.
• Protect teaching time. Count mentoring as output, not overhead.
• Write things down while people are still here, not after the pink slips, as Northrop Grumman had to.
• If you are 23, look for employers who still run a real first rung, and treat it as worth more than the starting salary.

What I don’t know #

The causal link between AI and junior hiring is not established. Both the Stanford and Ifo teams say so, and the IAB finds no significant effect in Germany. The economic cycle, interest rates and demographics all play a part. My ten-year arithmetic is an illustration, not a model. And my own news data starts on 7 November, so it shows what is being reported this week, not the trend.

Frequently asked questions #

Is AI really taking entry-level jobs? #

The pattern is documented; the cause is not. Stanford’s data show young workers in AI-exposed jobs falling behind, mostly through fewer hires. The authors call their results descriptive, not causal, and Germany’s IAB finds no significant AI effect. Weak demand is clearly part of the story.

Where does the 13% figure come from? #

From the first version of the Stanford “Canaries in the Coal Mine?” paper, in August 2025. Later versions use more data and different specifications. The August 2026 update reports a 19% shortfall against the trend of similar young workers in less-exposed jobs.

Why does it matter if firms hire fewer juniors for a few years? #

Because juniors become the next mid-career staff. Tacit skills are learned by doing junior work under supervision for years. A gap in hiring now becomes a gap in team leads and senior specialists a decade later.

What can a firm do instead? #

Keep hiring juniors, perhaps fewer, and redesign the first job around reviewing and checking AI output under a senior’s eye. Count mentoring as real work, and capture knowledge while people are still there.

The lesson #

The first rung looks like the cheapest part of a profession to cut, and that is exactly why cutting it is dangerous. It is where codified work turns into tacit skill. A traditional English apprenticeship ran for seven years, and today’s path from graduate to trusted senior is not much shorter. A gap in hiring takes at least that long to fill, and the firms that save on juniors now will pay for it in the 2030s.