ANALYSIS: Is AI Closing the First Door to the Job Market?

The door doesn’t slam; it simply stops opening

A figure that doesn’t spell doom

19%.

This number doesn’t mean that artificial intelligence has wiped out one-fifth of all entry-level jobs in the U.S. It tells us something more specific, more cautious, and—precisely for that reason—more troubling: in the payroll data analyzed by Stanford, employment among 22- to 25-year-olds in occupations highly exposed to AI is 19% below the level it would have reached had it kept pace with their peers in less-exposed occupations.

No widespread collapse. No tidal wave of layoffs. No robot snatching an access card from a young worker’s hands.

The silence of job postings that no longer appear

The mechanism at work is more subtle: fewer hires, but no increase in layoffs. The first job doesn’t end. It never begins.

This difference changes everything, because a society can see its overall statistics hold steady while the first step of its social mobility crumbles beneath the feet of those just arriving.

It’s not easy to capture an absence. We can’t interview job openings that were never posted or count the apprenticeships that will never take place. Yet this may be where AI is beginning to displace work: before the paycheck, before the office, even before the rejection.

The first cost of automation could be a start that is withdrawn without any announcement of its cancellation.

What Stanford Actually Measured

Millions of paychecks, not a crystal ball

Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen analyzed an administrative dataset from ADP covering between 3.5 and 5 million employees per month in their balanced panel through June 2026. They compared employment trajectories by age and occupational exposure to generative AI.

Between November 2022 and June 2026, employment among 22- to 25-year-olds in the two most-exposed quintiles declined by about 11%, while it increased by about 10% in the three least-exposed quintiles. Total employment in this age group, however, remained virtually flat.

A Descriptive Indicator, Not a Causal Verdict

The authors say so themselves: their results are early indicators, not a causal estimate. The ADP sample overrepresents large companies, manufacturing, and wholesale trade; it underrepresents, in particular, retail trade, lodging, and food services.

The direction of the signal holds up across several tests, but its magnitude varies across databases, and controlling for education level significantly reduces certain estimates. Herein lies the honest nuance: the canary sings—it does not pass judgment.

Rejecting exaggeration does not mean we must look away. A phenomenon does not need to be perfectly isolated to warrant a response; it needs to be described with the precise accuracy that the data allow.

Scientific doubt does not erase the warning; it gives it shape.

The job market holds steady, while young people slip

The reassuring aggregate

Across the main sample as a whole, employment has grown by about 6% since November 2022. Even in the quintile most exposed to AI, it has increased by about 4%. The study therefore finds no evidence of a widespread shift in employment across the economy.

This finding deserves to be stated plainly. The doomsayers still have too little evidence. AI has not emptied American offices. Experienced workers do not show the same disparity as younger workers.

The Average That Hides an Age Divide

But an average can be as solid as a facade and as hollow as a set. Those aged 22 to 25 account for less than 10% of the sample; their decline can therefore be masked by the positive trend among older cohorts.

The market isn’t collapsing. It’s closing off at the bottom. These are two compatible realities, and confusing them would allow companies to celebrate their productivity while an entire generation searches for a place to learn how to generate it.

The political danger lies here: waiting until the problem becomes visible in the overall unemployment rate. By then, unacquired skills, delayed promotions, and professional networks that were never formed will have already left a lasting scar.

An economy can look good on paper and still fail in its duty to pass on skills.

Codified knowledge becomes the perfect target

What textbooks have made accessible

Entry-level jobs often rely on codified knowledge: rules, procedures, formats, research, summaries, documentation, and repeatable checks. This is precisely what generative models ingest, recombine, and apply with spectacular speed.

The first tasks entrusted to a beginner were rarely prestigious. They served a purpose. They taught the vocabulary of the trade, the discipline of attention to detail, the consequences of a mistake, and how an abstract decision affects a real person.

What Experience Still Preserves

Tacit knowledge comes from elsewhere: practice, mentorship, and repeated exposure to ambiguous situations. It lives in judgment, in the ability to sense that a technically correct result is humanly wrong, and in the memory of exceptions not found in any manual.

AI absorbs what has already been written more easily; it often enhances the value of those who know what is missing from the written word. Thus, the same tool can replace a junior task and amplify the capabilities of a seasoned professional.

This is not a matter of the moral superiority of age. It is an accumulated advantage. Yet if a company stops bringing in novices, it protects its experience today by setting the stage for a shortage tomorrow.

You cannot demand experience after having eliminated the very place where it is forged.

Automation is not complementary

When the tool does the work

The Anthropic Economic Index distinguishes between “automating” uses—where the system directly performs a task with little human intervention—and “augmenting” uses—where it collaborates, explains, validates, or refines. Stanford identifies a significant difference between these two worlds.

Job losses among young people are concentrated more in professions where AI use replaces human tasks. Where the tool complements the work, employment is generally stable or on the rise, especially among experienced workers.

When the technology expands the scope of the job

This contrast dispels an intellectual laziness: talking about “AI” as a single phenomenon. A company can use it to eliminate human work or to make that work more challenging. The software doesn’t choose this approach. Management does.

Automating a menial task to free up a novice for a more meaningful task is not the same decision as automating the task and eliminating the novice. The technology is the same; the human intent is opposite.

The debate must therefore move beyond the mystical realm of the model’s capabilities. The real question is organizational: what do we do with the time saved, who benefits from the increased productivity, and what responsibility do we maintain toward the next generation?

The future of work depends not only on what machines can do, but on what employers decide to pass on.

“Seniorization” as an Economy of Patience

Junior positions that already require senior-level experience

PwC has given a name to a trend that many candidates recognize: “seniorization.” Entry-level positions still exist, but they require more autonomy, judgment, and experience. The job title remains at the bottom of the ladder; the requirements, however, have risen.

For the employer, the calculation seems rational. A seasoned professional, equipped with AI, can produce more and supervise fewer people. Coordination costs go down. The immediate profit margin improves.

The Shift of Risk to the Candidate

This model, however, shifts the risk. In the past, companies hired promising talent and funded part of their development. Today, they seek candidates who are already fully trained, then expect the school, internship program, family, or the candidate themselves to have paid for that training.

This is called efficiency when looking at the next quarter. Over the course of a generation, it amounts to the privatization of education: everyone must arrive ready in a world that no longer affords them the right to become ready.

Those with the best support networks will find projects, unpaid internships, mentors, or family capital. The others will encounter a door that requires a key—one that used to be made on the other side.

Seniorization rewards experience while refusing to finance its development.

The degree is losing some of its promise

An education that precisely codifies what AI has mastered

The study shows that controlling for university degrees significantly lowers the estimates. This result may point to an alternative explanation; it may also reveal a channel for the phenomenon. Universities convey a great deal of formalized knowledge, and this formalization is precisely what models can access.

A degree does not become useless. It may simply cease to be sufficient. It proves an ability to learn, but employers now want proof that you have already applied, evaluated, negotiated, and fixed problems.

The Shrinking Social Promise

For decades, a simple pact was sold: study, work, get a job, move up. If that first “job” disappears, student debt remains while the path to advancement fades away.

This isn’t just a curriculum issue. It’s a breach of contract between the institutions that educate, the companies that hire, and the young people who are asked to invest years of their lives without even being guaranteed the minimum space needed to turn their knowledge into sound judgment.

The answer, therefore, cannot be yet another slogan about the skills of the future. It requires pathways where experience is truly gained—with responsibility, a paycheck, and a clear path for advancement.

When school leads to a locked door, it is not the student alone who has lost their way.

Mentorship does not automatically survive productivity

The Invisible Work of Knowledge Transfer

Training someone slows things down at first. You have to review, explain, let them try, correct without humiliating, and start over. In a dashboard obsessed with throughput, this time seems like a waste.

Yet this slowness produced the judgment on which the company subsequently depends. The junior learned the exceptions; the senior learned to articulate their intuition. Both became better because knowledge had to circulate.

The Machine as a Tempting Shortcut

A generative assistant can provide an instant answer, rephrase an email, produce a first draft, or flag an anomaly. It can also free the manager from having to explain why the answer is correct, why the email builds a relationship, or why the anomaly matters.

Productivity without mentorship results in faster outputs and more fragile professionals. It replaces the conversation that used to pass on the craft with an interface that primarily delivers a solution.

The company saves time today and loses its collective memory tomorrow. This cost won’t show up in the June 2026 payroll data. It will become apparent when those who know how to do the job leave and no one has learned to see what they saw.

An organization that never slows down to teach eventually forgets what it knows.

Inequality may worsen before unemployment does

Those who can buy their first opportunity

When official entry points narrow, private avenues gain value: recommendations, volunteer work, personal projects, certifications, willingness to relocate, and periods without income. Each favors those who already have a financial cushion.

The market may then select less on the basis of talent than on the ability to survive long enough without being chosen. AI doesn’t invent this injustice. It can accelerate it by reducing the number of opportunities where an employer is willing to take a chance on imperfect potential.

Those Whose Resumes Must Pay Off Immediately

For a young person who has to pay for housing, support their family, or pay off student loans, waiting is not a viable strategy. They take a job that’s less closely aligned with their education. Then time passes, the gap in their resume widens, and entry-level positions demand ever more relevant experience.

The damage isn’t just in terms of salary. It’s the career path that diverges: an initial temporary decision becomes a delay, the delay becomes a signal, and the signal becomes yet another barrier.

That’s why it would be irresponsible to expect a dramatic drop in salaries. Stanford observes an adjustment in employment first, not in base pay. Exclusion precedes price reduction.

The market doesn’t pay certain young people less; it first stops asking them how much they’re worth.

Companies Are Eating Away at Their Own Next Generation

Individual performance masks the collective debt

A team of senior employees equipped with powerful tools can be formidably effective. Each job cut seems justifiable in isolation. But if all companies make the same calculation, none will be willing to pay for the training that all of them will eventually need.

It’s a tragedy of the first step. Everyone cuts back on their junior training programs; a few years later, the entire industry discovers it lacks mid-level professionals capable of taking over.

The talent pipeline isn’t a file you can restore

You can reboot a server. You cannot make up for five years of missing experience in a single quarter. Professional judgment is cumulative, relational, and often context-specific. It cannot be bought at the last minute without outbidding the competition.

Eliminating a junior position because a promoted senior employee is sufficient today is like burning the wood from the roof trusses to heat the office. The gain is real. So is the cost—only it arrives after those who made the decision have left.

Boards of directors should treat training as infrastructure, not as a discretionary expense. Without renewal, current productivity becomes an advance drawn on the future.

A company that no longer hires entry-level employees is quietly announcing that it intends to live off the past achievements of others.

What the Data Doesn’t Reveal

Rates, Remote Work, and the Scars of the Pandemic

The labor market has experienced several shocks at once: rising interest rates, a tech slowdown, remote work, educational transformations, and the post-pandemic return to work. Jobs at risk from AI also vary by education level, salary, and industry.

Stanford is testing several of these factors. The signal persists after certain controls are applied and after excluding IT jobs or tech companies. But education weakens it, and these disparities existed before ChatGPT.

The Discrepancy Between ADP and Public Surveys

The direction of the trend also appears in certain public data, but to a lesser extent. The small cells in the Current Population Survey are highly noisy; the American Community Survey only covers data up to 2024 in the comparisons presented.

Blaming AI for every job lost would be wrong. Claiming that uncertainty exonerates it would be just as lazy. A responsible stance acknowledges both points: the signal is concerning; the exact causality has not been established.

This discipline safeguards the debate. It prevents fearmongers from inflating the numbers and technology salespeople from hiding behind the lack of a perfect experiment.

Caution is not an escape route; it is the way to stay in the room.

The government cannot provide training in place of work

Useful programs, but without real accountability

Public authorities can fund internships, support apprenticeships, modernize community colleges, track hiring by age, and make certain contracts contingent on training programs. They can also highlight companies that are building the next generation of talent.

But no program can fully replace a real assignment, a mistake for which you are held accountable, a client whose needs you listen to, or a colleague who corrects you. Competence is forged through the challenges of real-world work.

Sharing the Cost of Apprenticeships

The best solution might be a partnership: the government reduces part of the initial risk; the employer guarantees a training position, a mentor, clear objectives, and a path for advancement; and the young person contributes their effort without having to pretend they already have five years of experience.

Subsidizing an empty title isn’t enough. Every public dollar should buy measurable skill transfer: mentoring time, tasks that increase in complexity, and a genuine chance of being hired after the apprenticeship.

The challenge isn’t to preserve every old job. Some deserve to disappear. It’s to preserve the pathway through which a person becomes capable of taking on new ones.

We won’t save the first step by painting it; we must rebuild it under the real weight of a job.

AI Should Create Stronger Apprentices

Empower the novice instead of erasing them

There is another approach. Give the beginner the tool, teach them to verify, document, challenge, and explain. Entrust them with complex problems sooner, while keeping a human in charge of quality.

In this model, AI eliminates some of the repetition, but the junior employee remains because they learn to steer the process, make judgments, and take responsibility for the outcome. The productivity gains fund skill development rather than merely funding workforce reductions.

Measure what grows, not just what costs

Companies know how to calculate time saved. They are less adept at measuring the depth of a talent pool, the number of decisions an employee can make without supervision, or the speed at which a trainee becomes a mentor.

If AI only enhances those who were already strong, it concentrates power. If it also enhances those who are just starting out, it can shorten the learning curve without eliminating it. This distinction should become a governance metric.

Every deployment should answer a simple question: How many tasks are eliminated, how many new responsibilities emerge, and who will actually have the opportunity to learn them?

A good tool doesn’t make the next generation redundant; it gives them something worthwhile to take on sooner.

Leaders must make their human choices public

Take AI Out of the Vokabulary of Magic

Saying that “AI is transforming jobs” often hides a chain of human decisions: budget cuts, unfilled positions, merged teams, raised experience requirements, and eliminated training. Technology becomes a grammatical subject that allows all the real subjects to disappear.

We need to flip the sentence around. Who chose to automate? Who reaped the benefits? Who safeguarded mentoring? Who assessed the impact on hiring people under 26?

Making the First Step Visible

Companies could publish the percentage of entry-level positions, their conversion rates, the time dedicated to mentoring, and trends in experience requirements. Not to create a brochure-worthy virtue, but to make a responsibility that is currently invisible comparable.

A management team that announces billions in productivity gains should be able to say how many new professionals it will train with that power. Otherwise, innovation resembles extraction: taking accumulated knowledge without paying for its renewal.

Investors themselves have an interest in asking this question. A margin gained through underinvestment in people isn’t necessarily a sign of strong performance; it may be a debt in disguise.

What isn’t measured often ends up being sacrificed, especially when those who stand to lose it aren’t even in the room yet.

The first step determines the rest

Don’t confuse a warning with inevitability

19% is neither a prophecy nor a reason to panic. It is a descriptive deviation in a specific sample, following a turbulent period, with acknowledged limitations. It may evolve, dissipate, or take on a different form.

But it already reveals an organizational truth: the codified tasks of newcomers are easy targets; experienced workers are better at recognizing complementarity; and adjustments are reflected in hiring before they are seen in layoffs.

Choosing Who Will Be Allowed to Become Experienced

We can let the market optimize every position until careers no longer have a beginning. We can also decide that productivity must buy learning time, that experience is not an individual miracle, and that the first job constitutes a social infrastructure.

The question, then, is not whether AI can do a beginner’s work. It already can for certain tasks. The question is whether we want an economy capable of accomplishing a great deal, but incapable of producing those who will accomplish tomorrow’s work.

The Stanford canary isn’t saying the mine is collapsing. It’s saying the air is changing—first at the level where those who have just entered are breathing.

19%.

How much longer will we call this a gain in efficiency before admitting that we may be erasing the very beginning?

A door closed without a sound remains a closed door.

Signed, Maxime Marquette, columnist

Columnist’s Transparency Box

Editorial Stance

I am not a journalist, but a columnist and analyst. My expertise lies in observing and analyzing the geopolitical, economic, and strategic dynamics that shape our world. My work consists of dissecting political strategies, understanding global economic trends, contextualizing the decisions of international actors, and offering analytical perspectives on the transformations that are redefining our societies.

I do not claim to possess the cold objectivity of traditional journalism, which is limited to factual reporting. I strive for analytical clarity, rigorous interpretation, and a deep understanding of the complex issues that affect us all. My role is to make sense of the facts, situate them within their historical and strategic context, and offer a critical analysis of events.

Methodology and Sources

This text respects the fundamental distinction between verified facts and interpretive analyses. The methodological rule is consistent: factual information is published only if it is supported by a verifiable source, and the sources actually used in this article are listed under “Sources,” never here.

Categories of primary sources used by the publication, when applicable: official press releases from governments and international institutions, public statements by political leaders, reports from intergovernmental organizations, and dispatches from recognized international news agencies.

Types of secondary sources: specialized publications, internationally recognized news media, analyses from established research institutions, and reports from sector-specific organizations.

When an article cites statistical, economic, or geopolitical data, it comes from data-producing institutions (intergovernmental organizations, central banks, national statistical institutes), and the specific institution is listed under “Sources.”

Nature of the Analysis

The analyses, interpretations, and perspectives presented in the analytical sections of this article constitute a critical and contextual synthesis based on available information, observed trends, and expert commentary cited in the sources consulted.

My role is to interpret these facts, contextualize them within the framework of contemporary geopolitical and economic dynamics, and give them coherent meaning within the broader narrative of the transformations shaping our era. These analyses reflect expertise developed through continuous observation of international affairs and an understanding of the strategic mechanisms that drive global actors.

This article describes a situation documented as of its publication date, not a prediction: subsequent developments may alter these perspectives. No updates are promised in advance; when an article is corrected or supplemented, the change is dated within the text.

ANALYSIS: Is AI Closing the First Door to the Job Market?