Will AI Replace Veterans or Juniors First? The Data Says "Seniority Wins"

When people talk about "which jobs disappear first" as AI spreads, many assume veterans doing routine work are most at risk. What the data of the past few years actually shows is the opposite.

The paper "Canaries in the Coal Mine?" by Brynjolfsson and colleagues at Stanford's Digital Economy Lab uses payroll data covering millions of workers from the payroll firm ADP to show that in AI-exposed occupations, it is employment of 22- to 25-year-olds that has stalled. After the first version in August 2025, the revised version of August 2026 — with data through June 2026 — reports that the gap has widened further. A paper by Harvard researchers analyzing about 65 million résumés (May 2026 version) finds that only junior employment fell at firms that adopted generative AI, and a study by Anthropic (March 2026) that matched Claude usage data with U.S. employment statistics found signs pointing the same way in young people's hiring.

This article lays out, from research and official statistics, how AI's impact on employment differs by experience — junior versus senior workers, and turns it into questions you can use to tell which side your own job is on. Which occupations are most affected is in 15 Jobs Most Affected by AI, the conditions that make work last in Jobs That Survive the AI Era, the gap between white-collar forecasts and actual data in Will AI Eliminate White-Collar Jobs?, and the performance gap among people who keep working in the same workplace in How Does AI Widen the Ability Gap Among Office Workers?

1. The Answer — the Data Says "Juniors First"

No guesswork — here are figures checked against the publishers' original pages.

Two points deserve attention. First, what is falling is hiring, not layoffs. The Stanford paper finds the effect works mainly through reduced hiring of young workers rather than increased separations. Before people in jobs lose them, the entry point narrows. Second, there is no sign of broad unemployment across the economy. The paper's very first finding is "no evidence of widespread, economy-wide job displacement," and total employment in the ADP sample rose about 6% over the period. The effect is concentrated in a narrow slice: young workers in AI-exposed occupations.

The old assumption that "young workers are protected" no longer holds. What is emerging instead is a structure in which what AI automates most easily is the textbook work traditionally handed to newcomers — writing routine code, keeping books, first-line customer inquiries.

2. What Official Statistics Show — U.S. Graduate Unemployment and Japan's Placement Rate

The studies in section 1 rely on company payroll and résumé data. Here we check whether national statistics point the same way. For the U.S., we use The Labor Market for Recent College Graduates, which the Federal Reserve Bank of New York publishes quarterly from the Current Population Survey (CPS); for comparison, we add Japan's graduate employment survey (in Japanese) by the Ministry of Health, Labour and Welfare and the Ministry of Education.

The unemployment rate of recent college graduates rose 1.2 points in four years, from 4.0% to 5.2% — the largest rise of the four groups. For much of the 1990s through the 2010s, recent graduates had a lower unemployment rate than the workforce as a whole; since 2021 it has stayed above it. Among employed recent graduates, 42.5% work in jobs that typically do not require a degree (the New York Fed's "underemployment rate").

These statistics alone do not show that AI is to blame. Rising interest rates and the post-pandemic hiring correction overlap with the same period. That is exactly why the studies in section 1 compare AI-exposed occupations with less-exposed ones — to separate these effects.

The New York Fed also publishes unemployment rates for recent graduates (ages 22–27) by major. Here are selected majors.

Unemployment among new graduates is highest in computing majors, where AI is widely used to write code. Civil engineering, also an engineering field, is low, and nursing — which requires a license and hands-on practice — is lower still. The figures cover a single year, 2024 (American Community Survey, released February 2026), and also reflect factors other than AI, such as the number of graduates in each field and the economy.

Other countries can look different. In Japan, for example, the placement rate for university graduates (bachelor's) who finished in March 2026 was 98.0% as of April 1, unchanged from the previous year (Ministry of Health, Labour and Welfare and Ministry of Education (in Japanese), published May 22, 2026). Graduate hiring as a whole shows no U.S.-style dip there. However, the rate is the share of job-seeking graduates who found jobs and has no breakdown by occupation. Differences in AI use by age group are covered in our article on the ability gap among office workers.

3. Why Senior Workers Hold Up

As a candidate explanation, the Stanford paper offers the hypothesis that AI substitutes more easily for codified knowledge (taught through school, textbooks and written procedures) and complements tacit knowledge (gained through practice and mentorship), and shows the data points the same way: occupations that rely more on codified knowledge have slower entry-level employment growth, while occupations that rely more on tacit knowledge have faster growth for mid-career and senior workers. The authors stress these are descriptive, not causal, results.

A second clue is how AI is used. Using the Anthropic Economic Index (Claude usage data), the same paper reports that in occupations where AI is mostly used to do the work (automation), young workers' employment falls, while in occupations where it mostly helps people (augmentation), employment is flat or rising — especially for experienced workers. Among equally AI-exposed occupations, the way AI is used splits the direction.

Job postings point the same way. In the Harvard study, tasks that AI does well gradually disappeared from junior job postings at firms that adopted generative AI. The decline in junior employment was concentrated in AI-exposed occupations, while senior employment trends barely changed.

This is the realm of tacit knowledge — knowledge you can't put in a manual and can only internalize on the job. AI can reproduce textbook knowledge instantly, but tacit knowledge is still largely carried by people. The data's "codified versus tacit" gradient points the same way as this intuition.

4. Three Questions to Assess Your Own Job

The research so far shows that jobs whose entry point narrows tend to share some traits. The three questions below are this article's framework for applying those traits to your own job — not a conclusion of the research itself.

Here is what juniors and seniors can do, depending on how many questions you answered "yes".

The Stanford paper's case examples include software development and customer service, among others. Ten-year projections by occupation (U.S. Bureau of Labor Statistics) are in 15 Jobs Most Affected by AI.

5. The Vanishing Training Ground — When Juniors Can't Develop

The serious risk here is that companies may be setting a time bomb: stop hiring juniors → juniors don't develop → a shortage of seniors a few years later.

Traditionally, new engineers and accountants read their seniors' code, worked through routine tasks, and internalized tacit knowledge on the job. When AI takes over those entry-level tasks, the place where juniors learn disappears with them.

Put differently, a company that decides "AI lowers costs, so we'll stop hiring graduates" is also deciding not to hire its own future seniors. The Stanford paper, too, flags the slowdown in entry-level hiring as an early sign that generative AI may be affecting the career ladder. Note that "a shortage of mid-career staff in a few years" is a concern based on current data, not a measured fact.

6. The Counterargument — "AI Isn't Necessarily the Cause"

There are strong counterarguments. An analysis by The Budget Lab at Yale (October 2025) concluded that in the 33 months after ChatGPT's release, the broader labor market had not experienced a discernible disruption. The occupational mix was changing a little faster than in past technology shifts, but not markedly so — and the change predated widespread AI use. Anthropic's own study (March 2026) also found no systematic increase in unemployment for highly exposed workers.

The Stanford paper addresses its own weaknesses head-on:

• The gap shrinks when controlling for education: it cannot tell whether education is an alternative explanation or the very channel through which AI works
• Some divergence predates ChatGPT: more- and less-exposed occupations moved differently around the COVID-19 pandemic
• The gap is larger in ADP data than in national surveys: national statistics point the same way, but less sharply

At the same time, the paper notes that the gap persists when controlling for interest-rate exposure, when excluding tech firms and computer occupations, and when controlling for remote work; that it kept widening after interest rates peaked; and that it concentrates in occupations where AI is used for automation, with a clear age gradient — patterns that interest rates or remote work alone struggle to explain.

So AI cannot be called the sole cause of stalled youth employment. But the relationship — the more an occupation is exposed to AI, and the more AI is used there to automate, the more young hiring falls — keeps showing up in several independent datasets. This article's position: AI is not the only cause, but it is one of the pressures acting on the entry point.

7. Strategies for Early-Career Workers

"I get the data — so what should I do?" Here's a practical list.

1. Choose workplaces where AI assists people

Within the same occupation, some workplaces hand whole tasks to AI and others use AI to assist people. In the Stanford paper, young workers' employment was flat or rising in occupations where AI mainly augments work. In interviews, ask "What do newcomers do with AI, and who checks it?" — the answer shows which way a workplace leans. Being able to use AI is a given; what sets people apart is the judgment of which parts of the output to doubt and which to fix.

2. Stake out ground where AI is weak — early

• Physical presence: on-site work, healthcare and nursing, real-world communication
• Accountable judgment: compliance, governance, ethics
• Designing creative questions: new businesses, UX design, brand
• Moving people: sales, coaching, leadership

Growing occupations and the education or training needed to enter them are tabulated in Jobs That Survive the AI Era.

3. Build combined skills

Being able to "write code" or "do accounting" on its own is no longer enough. Create scarcity through combinations like "domain × AI" or "design × data" — for example, "clinical experience + prompt engineering" or "legal practice + AI output verification."

4. Create your own place to gain tacit knowledge

If your company won't give you a chance to read seniors' code, get it yourself through open-source contributions, side projects, communities and mentors. As the data shows, occupations that rely on tacit knowledge see faster growth for experienced workers. In the AI era, early-career workers can no longer rely only on being trained by their employer.

5. Move up a level — toward management, design and business judgment

Pure implementation seats are the easiest to cut. But if you step early into roles where you make design decisions, understand the business and can move people, you ride on the side where AI amplifies you. In the Stanford paper, too, employment was flat or rising in occupations where AI is used to help people.

8. Where Senior Workers Can't Relax

"Seniority wins" does not mean every senior is safe. These groups are arguably more at risk than juniors:

In short, it isn't "senior = safe" but "a senior who uses AI well and exercises tacit judgment = safe." What is growing in the data is not "experienced people in general" but experienced workers in occupations where AI is used to help people. The dividing line is attitude, not ability.

9. What Companies Should Do Now

Beyond individuals, companies also need to think about their long-term talent structure.

1. Treat junior hiring as an investment in future seniors, not a cost

AI lowers short-term costs. But closing the entry point risks a shortage of mid-career and senior staff a few years out. Cutting graduate hiring can mean cutting your future competitiveness.

2. Redesign early-career training for the AI era

If "read seniors' code" and "learn by doing routine tasks" no longer work, you need a new curriculum: critique AI-generated solutions, debate with AI, experience AI's limits.

3. Use seniors as "AI amplifiers"

Senior tacit knowledge × AI scalability = a big productivity gain. Redesign teams on the premise that seniors use AI to increase their output and hand their way of judging on to juniors.

4. Formalize "a human signs last" governance

Always put a human checkpoint before AI output goes to production. It preserves both work and learning opportunities for juniors and quality assurance at the same time.

Summary

• The data says juniors are being affected by AI first. In Stanford's revised paper (August 2026), employment of 22- to 25-year-olds in AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed occupations. Ages 35–49 in the same occupations grew about 10%.
• Harvard's analysis of résumé data likewise finds junior employment about 9% lower six quarters after firms adopted generative AI, with seniors barely changed. In both, the decline comes from less hiring, not more layoffs.
• In official statistics, the U.S. unemployment rate for recent college graduates rose from 4.0% (Sept. 2022) to 5.2% (Sept. 2026), the largest rise of the four groups. In Japan, by contrast, the placement rate for March 2026 university graduates was 98.0%, with no comparable dip.
• A likely reason: AI replaces codified knowledge and complements tacit knowledge. Juniors decline in occupations where AI automates work; experienced workers grow where AI augments people.
• Assess your own job with three questions: can a newcomer's work be written into a manual, is AI used to do the work instead of people, and are graduate hiring or unemployment getting worse?
• Counterarguments: Yale's Budget Lab finds no discernible disruption in the overall labor market. Stanford itself acknowledges that the gap narrows when adjusted for education and that some differences predate AI.
• Longer term, the worry is a vanishing training pipeline: organizations that close the entry point may lack mid-level and senior staff in a few years.
• Seniors are at risk too if they don't use AI, have shallow hands-on skills or can't pass on tacit knowledge. Companies should rethink junior hiring as an investment in the future.

FAQ

Q1. Is "juniors first" only a U.S. phenomenon?

The research cited here (Stanford, Harvard and Anthropic) and the New York Fed statistics are all U.S. data, and analyses on this scale are still scarce for other countries. In Japan, for example, the placement rate for university graduates who finished in March 2026 was 98.0% (as of April 1, unchanged from the previous year), with no U.S.-style dip in graduate hiring overall. In countries with strong traditions of mass graduate hiring or long-term employment, hiring adjusts differently, so even if the same shift appears, its form and timing may differ. Because such rates rarely break down by occupation, the reliable way is to check how graduate hiring in your own industry — and the work described in junior job postings — is actually changing, using national statistics or industry data.

Q2. So should young people stop becoming engineers?

The opposite. Young people who can use AI well right now are scarce. What is shrinking is hiring for entry-level work AI can easily replace; the value of juniors who can use AI and take on judgment is rising. In the U.S. BLS projections, software developers grow +10.2% through 2035, while computer programmers — whose work centers on writing code — shrink −7.3%. Rather than giving up on computer science, change what you learn and how you approach it.

Q3. How can people in their 30s and 40s become "AI-savvy seniors"?

Just three things:
(1) Spend at least 30 minutes a day with Claude Code, Cursor or Codex
(2) Learn firsthand where AI's capabilities end in your own specialty
(3) Strengthen domain knowledge to the point where you can critique and fix AI output
Over the long run, putting your judgment criteria into words matters more than tool skills.

Q4. Can the training-pipeline collapse be prevented?

At the company level, yes: (1) keep hiring graduates, (2) build training programs that work with AI, and (3) make "time to teach" part of seniors' jobs. What happens society-wide is still open; the Stanford authors are at the stage of publishing indicators so others can keep tracking it.

Q5. Are hands-on jobs (construction, nursing, delivery) really safe?

Relatively safe for now. In Anthropic's usage data, about 30% of workers were in occupations with essentially no observed AI use (cooks, mechanics, lifeguards and similar). Jobs with a strong element of a person on site making the call are far from being replaced. Over the long term, autonomous robots and self-driving vehicles may spread, so a 20-year view is a different story.

Q6. Who is stronger — a senior who uses AI or a junior who doesn't?

In most workplaces, probably the senior who uses AI. When senior tacit knowledge is combined with AI's throughput, the productivity gap becomes large. Conversely, a junior who doesn't use AI is likely to lose ground as entry-level work moves to AI.

Q7. I feel like "half my job disappeared" because of AI. What should I do?

Move in three stages. (1) Short term: focus on the remaining half to raise quality, and use AI to double productivity on the rest. (2) Medium term: deliberately raise the share of work AI is weak at (judgment, people, physical). (3) Long term: read how your industry is changing and redefine what your job looks like five years after AI is widespread. The mindset of "re-staking your place within the industry" is key.

Related: 15 Jobs Most Affected by Generative AI (projections by occupation) / Will AI Eliminate White-Collar Jobs? (forecasts vs. actual data) / How Does AI Widen the Ability Gap Among Office Workers? (the gap among people who keep working) / What AI Did to IT Freelancers — this article is about employment, but for freelancers, experience works the other way around