Picture, as an illustration, the first week of a first job. The work that used to fill an analyst’s first year — summarizing filings, writing the first draft of a client memo, cleaning the data behind the quarterly deck — now arrives already done by a tool, and the new hire’s job is to check it. The incoming cohort is smaller than the one before. Nobody was let go to make it smaller. The firm simply hired fewer people.

That last detail is the one the data keep finding. Most of what is said about AI and jobs is forecast or anecdote. The one change that has been measured repeatedly, in payroll records and in government surveys, is narrower: fewer young people are being hired into the occupations most exposed to AI. How much of that AI itself has caused is still debated. If you are in your first few years, or about to start, it is worth understanding precisely: it says less about whether you will have a career than about what its first years now need to accomplish.

What has actually been measured

The most detailed evidence comes from Stanford’s Digital Economy Lab, which tracks ADP payroll data covering millions of US workers. In its August 2026 update, employment of 22- to 25-year-olds in AI-exposed occupations stood 19% below where it would have been had it kept pace with their peers in less-exposed occupations. Experienced workers show no comparable gap. In plain levels, employment of 22- to 25-year-olds in the two most exposed fifths of occupations fell about 11% between November 2022 and June 2026, while employment of the same age group in the three least exposed fifths grew about 10%. The gap has widened: it was 15% in data through July 2025 [1].

Two further findings matter as much. The study finds no evidence of widespread, economy-wide job displacement from AI. And the divergence “operates primarily through reduced hiring of young workers rather than increased separations” [1]. Economists at the Dallas Fed, working from a different source, the government’s Current Population Survey, reach the same conclusion about the mechanism: layoffs do not explain the weaker numbers for young workers; fewer of them are entering these jobs in the first place [2].

The surrounding market is consistent with a narrower entrance. The New York Fed puts unemployment for recent college graduates aged 22 to 27 at about 5.6% in the second quarter of 2026, with 42% underemployed, meaning in jobs that do not typically require a degree [3]. Revelio Labs finds US entry-level job postings down about 35% since early 2023, while cautioning that a fall in postings does not by itself establish AI as the cause [4]. Handshake reports postings 2% lower than a year earlier and 12% below pre-pandemic levels, and the share of graduating seniors who are pessimistic about the job market up from 46% to 62% in two years [5].

Where the evidence disagrees

The size of the effect is genuinely contested, and the researchers say so themselves. The Stanford authors report that their estimates attenuate when they control for the education levels of occupations, that some of the divergence began before generative AI, and that the pattern is more pronounced in their payroll sample than in national surveys. They describe their results as descriptive facts, not causal estimates [1]. The Dallas Fed’s survey-based measure is smaller: the share of 20- to 24-year-olds employed in the most AI-exposed occupations slipped from 16.4% in November 2022 to 15.5% in September 2025, and even if all of that decline had turned into unemployment, it would account for about 0.1 percentage point of the national rate [2].

Other measures point the other way. Employers responding to NACE’s spring 2026 survey projected hiring 5.6% more new graduates from the Class of 2026, up from a median of −2.4% in the autumn [6]. In ZipRecruiter’s 2026 graduate survey, the share of recent graduates landing a role within three months of graduating rose from 63% to 77% [7].

These findings can all be true at once. Large employers can plan to hire more graduates while entry-level postings across the market fall; many graduates can find work quickly while fewer find it in the occupations where AI is most capable. The reading the evidence supports is specific: the first rung in AI-exposed fields is narrower, not gone, and the narrowing happens at the point of hire. Nothing in the data says that people who get in are being pushed out. The difficulty is concentrated at the door.

Why the first rung, and not the rest of the ladder

The Stanford study offers a clue to why the change falls on new entrants. Declines are concentrated in occupations where AI is mostly used to substitute for human tasks; where it is mostly used to complement workers, employment is flat or rising, especially for experienced workers. The authors’ suggested mechanism is that AI substitutes for codified knowledge — the formal, documented knowledge at the core of a degree — while complementing tacit knowledge, the kind acquired through practice, mentorship and repeated exposure to real situations [1]. They are careful to call this suggestive, not proven.

Wage data point the same way. A February 2026 Dallas Fed analysis of 205 occupations finds that the premium employers pay for experience is higher in occupations more exposed to AI. Where experience earns very little, AI exposure is associated with weaker wage growth; in the occupations with the highest experience premium, more exposure is associated with faster wage growth. The author’s explanation is that experienced workers hold tacit knowledge that AI cannot replicate [8].

Software shows the pattern in job postings. Indeed Hiring Lab finds US software development postings up almost 15% since February 2025 while overall postings fell 7%, but 71% of the increase between May 2025 and May 2026 came from senior roles, and software postings remain 27.5% below their pre-pandemic level [9].

Put together, these findings describe a market that pays for what is learned on the job. A graduate arrives with codified knowledge. Junior roles used to turn that knowledge into judgment through years of routine work, reviewed by someone more experienced. The routine work is what AI now does well; the judgment is still what employers pay for.

That raises an obvious difficulty. If routine work is no longer how judgment gets built, it has to be built some other way: through regular review of your work, decisions given to you early, and time with experienced colleagues. Not every employer has worked out how to provide those. The first years are now less about doing the work nobody else wants and more about deliberately acquiring what the tools do not have, which is why the choice of first role matters as much as it does.

The first job carries more weight than it seems

The first role matters for longer than most people expect, and that was true before generative AI. In a study by the Strada Institute for the Future of Work and the Burning Glass Institute, 52% of bachelor’s graduates were underemployed a year after graduating and 45% were still underemployed ten years on. Graduates who started out underemployed were 3.5 times as likely to be underemployed a decade later as those whose first job required a degree, while 79% of those who started in a college-level job were still in one five years later [10].

The same evidence points to work experience before the first job as a lever. Completing an internship was associated with a 48.5% lower chance of becoming underemployed [10]. In ZipRecruiter’s survey, graduates who worked during school landed jobs at twice the rate of those who did not: 82% against 41% [7]. Both are associations, not controlled experiments, and students who work or take internships differ in other ways. But the direction is consistent, and it fits the mechanism above: evidence of real work has long counted with employers, and it is likely to count for more when there are fewer entry-level roles to go round.

If you are still studying, paid work and internships in or near your intended field are worth more than another line of coursework. And if your first job does not use your training, treat it as a detour with a planned exit, not a holding pattern. The data suggest the longer it lasts, the harder it is to leave.

Choose the first role for what it teaches

When entry is harder, the temptation is to accept the first offer on title or salary. The better test is what the role will teach you, and how quickly. The scarce input is feedback from experienced people, and it has a cost. In a study of software engineers at a large company, engineers who sat near their teammates received 18.3% more feedback on their code and wrote better code, with the gains concentrated among younger and less-tenured engineers. The experienced engineers giving that feedback wrote less code themselves [11]. Mentoring is real work, and organizations ration it.

That makes it something to ask about before you accept, not something to hope for after. Questions that tend to reveal the answer:

  • Who reviews my work, and how often? A named person and a regular rhythm is a good sign. “The team is very supportive” is not an answer.
  • What decisions will I make in my first year? Roles that include judgment calls, client or customer contact, or ownership of a small outcome build tacit knowledge. Roles that consist entirely of producing inputs for someone else build less of it.
  • How do junior people here use AI? The useful answer is one where the tool takes the routine work and the junior person takes on review, client contact or analysis earlier than before. The worrying answer is one where the routine work has gone and nothing has replaced it.
  • Where did the last two people in this role go next? A role whose holders move up or across within the organization is a role that builds something the organization values.

Use the tools, and learn where they fail

Fluency with AI is an asset early-career workers can build quickly, and the evidence suggests the tools help less experienced workers most. In a study of customer-support agents, access to an AI assistant raised productivity by 14% on average and by 34% for novice and lower-skilled workers, with minimal effect on the most experienced; the authors found suggestive evidence that the tool helped newer workers move faster down the experience curve [12]. In an experiment with 758 consultants at Boston Consulting Group, below-average performers improved by 43% with AI on tasks within its capabilities, against 17% for above-average performers. But on a task deliberately chosen to fall outside those capabilities, consultants using AI were 19 percentage points less likely to reach the correct answer than those working without it [13].

Employers say they want the skill. In Microsoft and LinkedIn’s 2024 Work Trend Index, 71% of leaders said they would rather hire a less experienced candidate with AI skills than a more experienced candidate without them, and 77% said that, with AI, early-in-career talent would be given greater responsibilities [14]. Those are 2024 intentions, not hiring outcomes. The fluency itself is no longer rare: Handshake finds that 85% of the Class of 2026 use AI and more than a third use it daily, while 58% say they will need stronger AI skills to succeed and only 28% say their program has meaningfully integrated it [5].

The consulting experiment shows where the advantage actually lies. Using the tool is common. Knowing when its answer is wrong is not, and that judgment is the tacit knowledge the labor-market data say employers pay for. The early-career worker who is worth hiring is the one who uses AI to do the routine work faster and then spends the time saved on checking, questioning and understanding the output. When you describe your work, describe it that way: not “used AI to draft reports” but what you caught, what you changed, and what the result was.

Search wider than the job title

The decline is concentrated in particular occupations, and occupations where AI complements workers have held up [1]. That argues for a wider search. Many graduates target only the job title their degree points to, which in an exposed field may be exactly where entry-level hiring has narrowed.

There is experimental evidence that a wider search pays. In a field experiment with job seekers in Edinburgh, showing half of them relevant alternative occupations and the jobs available in them significantly broadened the set of jobs they considered and significantly increased their job interviews. The effect was strongest for people who had otherwise searched narrowly [15]. The participants were unemployed job seekers in general, not new graduates, so the result is a guide rather than a guarantee. But the principle carries: list the occupations where the same knowledge is applied with more judgment, more contact with clients or colleagues, or more responsibility for an outcome, and search those too. How you look matters as much as where; the route into an organization often counts for more than the application.

Start the record in year one

The weakness every early-career candidate shares is the absence of a track record. Experienced professionals can point to outcomes; a graduate usually points to a degree. The way to close that gap is to start recording outcomes from the first month: what you were asked to do, what you decided, what changed as a result, and who saw it. A year of that turns into the evidence for a first promotion case, a strong self-review or an outside move, and it is far easier to keep as you go than to reconstruct later, as the article on documenting impact sets out.

Persona Map is a career intelligence platform built to hold that record: a private workspace where outcomes, feedback and the relationships behind them are kept as they happen, with an AI advisor that works from your own record. The data describe a narrower door, not a closed building. Getting through it takes evidence of real work; staying and moving up takes the judgment that comes only from doing that work in front of people who can teach you. The first years are for acquiring both, on purpose.

Sources

  1. Erik Brynjolfsson, Bharat Chandar & Ruyu Chen. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Stanford Digital Economy Lab, August 2026. Link
  2. Tyler Atkinson & Shane Yamco. “Young Workers’ Employment Drops in Occupations with High AI Exposure.” Federal Reserve Bank of Dallas, January 6, 2026. Link
  3. Federal Reserve Bank of New York. “The Labor Market for Recent College Graduates.” Data through the second quarter of 2026. Link
  4. Revelio Labs. “Entry-Level Job Postings Are Down 35%.” August 2026. Link
  5. Handshake. “AI and the Workforce Ahead: What the Class of 2026 Tells Us About the Future of the Labor Market.” April 2026. Link
  6. National Association of Colleges and Employers. “Employers Expect to Hire 5.6% More New College Graduates This Year.” Job Outlook 2026 Spring Update, April 15, 2026. Link
  7. ZipRecruiter. “ZipRecruiter Finds Grad Job Market Improving Despite Fears That AI Is Eliminating Jobs.” April 16, 2026. Link
  8. Scott Davis. “AI Is Simultaneously Aiding and Replacing Workers, Wage Data Suggest.” Federal Reserve Bank of Dallas, February 24, 2026. Link
  9. Guillermo Gallacher. “AI and Job Postings: From Destruction to Creation?” Indeed Hiring Lab, July 8, 2026. Link
  10. Andrew Hanson, Carlo Salerno, Matt Sigelman, Mels de Zeeuw & Stephen Moret. “Underemployment and Its Long-Term Effects for College Graduates.” Higher Education Today (American Council on Education), April 15, 2024, summarizing Strada Institute for the Future of Work and Burning Glass Institute, “Talent Disrupted.” Link
  11. Natalia Emanuel, Emma Harrington & Amanda Pallais. “The Power of Proximity to Coworkers.” NBER Working Paper 31880; Quarterly Journal of Economics 141(3), 2026. Link
  12. Erik Brynjolfsson, Danielle Li & Lindsey R. Raymond. “Generative AI at Work.” NBER Working Paper 31161; Quarterly Journal of Economics 140(2), 2025. Link
  13. Fabrizio Dell’Acqua et al. “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Harvard Business School Working Paper 24-013, 2023. Link
  14. Microsoft and LinkedIn. “AI at Work Is Here. Now Comes the Hard Part.” 2024 Work Trend Index, May 8, 2024. Link
  15. Michele Belot, Philipp Kircher & Paul Muller. “Providing Advice to Job Seekers at Low Cost: An Experimental Study on Online Advice.” IZA Discussion Paper 10068, 2016; Review of Economic Studies, 2019. Link