Is AI taking the first job? Nobody can tell yet

Nobody can yet say whether AI is shrinking young people's way into work, and the data that could settle it don't arrive in time to act on.

• Payroll data: Stanford's figures put employment of 22- to 25-year-olds in highly AI-exposed occupations about 19% below where it would be had it grown as fast as that of peers in less-exposed ones. A year earlier the shortfall was 15%.
• Danish registers: a study rules out average effects of chatbots on earnings and hours larger than 2%.
• Official US surveys: the monthly one reaches only a few dozen young software developers.

Meanwhile a whole cohort is entering the labour market. Policy proposals with automatic triggers are being written around AI's effect on jobs, and no official statistic yet isolates it. That is problem №3 in the registry.

What the data show

Payroll records point to a growing gap. Stanford's Digital Economy Lab works from ADP's payroll data, which cover millions of US workers. Its August 2026 update finds:

• From November 2022 to June 2026, employment of 22- to 25-year-olds fell about 11% in the two most AI-exposed fifths of occupations, and rose about 10% in the three least-exposed fifths.
• The gap comes mainly from fewer hires, not from more people leaving their jobs.
• It is concentrated where AI is used to automate work rather than assist it.
• So far it shows up mainly in employment, not in base pay.

Census data show a lasting drop in hiring. A US Census Bureau working paper uses the Bureau's Quarterly Workforce Indicators, built from employer–employee administrative records. It finds:

• Regression-adjusted employment of 22- to 24-year-olds fell 12% in the most AI-exposed fifth of industry–state cells in the ten quarters after ChatGPT launched. Less-exposed industries held steady.
• In the paper's words, "hiring has nearly recovered to its previous rate, but the number of hires has not". The rate largely recovered because employment had shrunk.
• Up to a quarter of the relative decline may come from monetary policy rather than AI.

Résumé data find a drop where firms adopt AI. A Harvard working paper follows 65 million US workers at more than 280,000 firms through their résumés. At firms that began advertising jobs to put generative AI to work, junior employment was about 9% lower than at comparable firms six quarters later, mostly through slower hiring. Senior employment barely changed. A job ad is a sign that a firm adopted AI, not a measure of how much it uses it.

Other measures find little or nothing:

• Yale's Budget Lab tracker, updated on 14 October with the monthly US household survey through September 2026, finds no clear evidence of AI-related disruption. Churn across occupations, and employment and wages of AI-exposed against comparable unexposed workers, are not consistent with any large-scale disruption so far. Unemployment has tended to grow faster among AI-exposed workers since late 2022, but Yale calls those estimates volatile.
• In Denmark, economists linked surveys of 25,000 workers in eleven highly exposed occupations to administrative records through December 2024. They found precise null effects of chatbots on earnings and hours. The share of early-career workers moved no differently at workplaces that encourage chatbot use. That study measures actual use, not exposure, which makes it unusual.
• Anthropic combined its models' usage with official survey data. It finds no systematic rise in unemployment among highly exposed workers, but suggestive evidence that hiring of young workers into exposed occupations has slowed (Anthropic).

These findings don't simply contradict each other. They measure different things, in different countries, over different periods. That is why they can't simply be added up.

Why the data can't settle it

The official surveys are too small, too slow or too coarse:

• The monthly Current Population Survey reaches between 23 and 48 software developers aged 22 to 25 in a given month (Stanford, online appendix).
• The American Community Survey is annual and arrives late (same paper).
• The Census Bureau's quarterly workforce statistics carry no occupation (Census working paper, §3).
• The October 2025 household survey was never collected, because of the US government shutdown (Yale Budget Lab).

The big datasets disagree on the size. For 2022 to 2024, Stanford compared the gap between young workers in the most- and least-exposed jobs across two sources:

• The gap: 13.2 points in ADP's payroll data, but 2.2 points in the American Community Survey. The survey's figure can't be told apart from zero.
• Where they agree: professional, information and financial services.
• Where they diverge: education, health and public administration.

Estimates of how many firms use AI vary widely. They differ by more than three times:

• The Census Bureau's business survey finds 17% to 20% of US firms used AI in the past two weeks (US Census Bureau).
• The New York Fed finds 61% of service firms in its region (Liberty Street Economics).

The two surveys ask different questions of different firms, and neither links adoption to who firms hire.

Stated reasons can't be audited. AI was the most common reason companies gave for job cuts announced in the US from January to September 2026: about 21% of 573,195 (Challenger, Gray & Christmas). Yet among AI-using service firms in the New York Fed's regional survey, only 4% had laid anyone off because of AI in the past six months; about 15% had hired fewer people because of it, and 13% had hired more.

Other shocks overlap. ChatGPT arrived just after a pandemic that upended hiring patterns, and remote work and education levels are both higher in AI-exposed industries (Census working paper). Stanford's gap persists after controlling for exposure to interest rates and for remote work, and it kept widening after rates peaked. It shrinks once education is taken into account. Even so, controls are not an experiment.

The missing piece is data.

• Stanford's ideal: employment records for the whole workforce, at high frequency, with age and occupation codes.
• What its own data lack: ADP's employers are anonymised, so Stanford's payroll data can't be matched to whether a firm uses AI.
• What we found no one has: a source that links what happens to workers to whether their employer uses AI, nationwide and quickly.

This is not a new gap. In 2018, the Senate Appropriations Committee asked the Bureau of Labor Statistics for a strategy to measure how new technology affects workers (BLS). In March 2026, a bipartisan group of US senators was still asking for AI questions to be added to the official surveys.

Why it matters now

Policy proposals are being built on these numbers:

• OpenAI has proposed safety nets that switch on automatically when public labour data, such as unemployment, cross set thresholds (OpenAI, p. 7).
• Anthropic's policy framework uses the unemployment rate as its main trigger (Anthropic, p. 6).

Those rates are published every month. What no official measure can yet show is how much of them AI causes.

A bad start lasts. Economists have shown that graduating into a weak labour market costs young people earnings for years. In one well-known study of Canadian graduates, a typical recession cut initial earnings by about 9%, and the loss took around ten years to fade (Oreopoulos, von Wachter and Heisz). If AI is closing the entry door, the cost falls on a whole cohort, and it shows only years later.

People expect the worst. 73% of US adults under 30 expect AI to mean fewer jobs in the US over the next 20 years. In 2024 the figure was 61% (Pew Research Center).

The predictions can't be checked yet. In May 2025, Anthropic's chief executive said AI could eliminate half of entry-level white-collar jobs and push unemployment to 10% to 20% within one to five years (Axios). In September 2026, US unemployment stood at 4.2% (BLS). The window he named runs to 2030, so neither side can claim the argument yet. That is the problem.

Why it is №3

We chose the founding problems in October 2026 from 37 candidates. Each was scored on how much depends on it, whether it will still matter in ten years, how many independent experts and institutions name it, whether a third party can check progress, how plainly it can be explained, and how urgent it is.

This one scored highest of the problems about people and society, level with №1 overall:

• It is one of the public's top worries. 71% of US adults expect AI to mean fewer jobs over the next 20 years (Pew Research Center). Across 37 countries, a median of 46% expect fewer jobs and 9% expect more (Pew Research Center).
• It is named widely:
  • the International AI Safety Report 2026;
  • IMF staff;
  • Stanford, Yale and the US Census Bureau's researchers;
  • OpenAI and Anthropic;
  • a bipartisan group of US senators.
• It can be checked. Whether a data series exists, how late it arrives and how precise it is can all be checked objectively. Judging whether a study truly identifies cause takes expert review.

It merges two problems from our research: the blind spot in the data, and the missing first rung of a career.

What would count as progress

• (a) A statistics office publishes, within 90 days of each quarter, hiring of 22- to 25-year-olds by occupation and by employers' reported AI use, for the US and one other large economy.
• (b) Two teams using different official records, with plans published in advance, estimate how much of the change in young people's hiring is due to AI, with overlapping margins of error under 10 points wide.
• (c) It is solved when five years of official records follow people who started work in 2023–25, and a rule set in advance shows whether their hiring and pay gaps at AI-using employers closed or stayed open.

Each test settles a different part of the question:

• (a) gives everyone the same numbers on time;
• (b) answers whether AI is the cause;
• (c) tells a delay from a lasting break.

Take it on

This statement is version 1.0, written by the UNSOLVED editors. If you are a labour economist, work at a statistics office or a payroll provider, or hold administrative data that could settle it:

• take it on as its stater;
• submit a contribution with evidence;
• have your organisation become its patron.

Everything starts on the problem page. To get what changes outside the registry each week, follow it there.

Researched and drafted with AI assistance. Every claim was checked against its source before publication.