For most people, AI won't arrive as a question about machine minds. It will arrive as a job application that never gets a reply.

Whether that is already happening depends on which study you read. One, from Stanford, says young workers in AI-exposed jobs are falling well behind. Another, from UCLA, says recent graduates are doing about as well as in previous years. Both come from serious researchers using serious data. Read closely, they disagree much less than the headlines suggest, and the narrow ground where they overlap is the most useful thing either one says.

The public has already decided

In a Pew survey of 3,488 US adults taken in June, 71% said AI will lead to fewer jobs over the next two decades, up from 64% in 2024. Only 5% expect more jobs. Among adults under 30, the share expecting fewer jobs rose from 61% to 73% in two years. Gallup found 79% expect AI to reduce US jobs over the next ten years, up from 73% in 2025. The two polls ask about different time frames, so their numbers can't be compared directly, but they point the same way.

The research is much less settled than the public.

Stanford: a widening gap for young workers

The most cited evidence is Stanford's "Canaries in the Coal Mine" research by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, built on payroll records from ADP, one of the largest payroll processors in the US. Their August 2026 update finds employment for workers aged 22 to 25 in highly AI-exposed occupations about 19% below where it would be had it kept pace with similar young workers in less-exposed jobs. At the July 2025 data vintage, the same measure was 15%.

That 19% is routinely misread as "19% of young workers lost their jobs." It is the distance between two paths. From November 2022 to June 2026, employment of 22 to 25 year olds in the two most exposed groups of occupations fell about 11%, while in the three least exposed groups it grew about 10%.

The details are what make the finding persuasive. The change comes mainly from reduced hiring, with little sign of young workers being let go. Experienced workers in the same occupations show no comparable gap. Declines cluster where AI is used to automate work, while employment holds steady or rises where AI complements workers. And the losses fall on young workers in codified-knowledge jobs, the kind that can be learned from books, while experienced workers with tacit knowledge, the know-how picked up by doing, gained.

The authors are careful about what this shows. They call the patterns descriptive, not causal. The gaps shrink when education is accounted for, and they are larger in ADP's sample than in national survey benchmarks.

A different method points the same way. Goldman Sachs economists estimated in April that AI was erasing roughly 16,000 net US jobs a month over the previous year, about 25,000 lost to substitution and 9,000 added through augmentation. That is an inference from statistical models, not a count of layoffs, and the economists say the true effect is likely smaller.

UCLA: no unemployment spike for graduates

In September, UCLA researchers Robert Fairlie and Jane Wu released a working paper based on the Current Population Survey, the large government household survey behind the official unemployment rate. Summer 2026 unemployment for bachelor's degree holders aged 22 to 25 averaged 7.3%, inside the 6.3% to 7.8% range of the previous four summers.

They found no statistically significant rise compared with earlier summers, older graduates or young people without degrees, and no significant link between a field's AI exposure and 2026 graduates' unemployment. The link they did find was with remote work: occupations where remote work is more available showed higher 2026 unemployment, which hints that something other than AI is moving that market. They also warn against reading their paper as a forecast, since classes graduating from 2027 on could face a different market.

Where the two studies actually meet

The disagreement shrinks once you line up what each study counts:

  • Different people. Stanford follows all 22 to 25 year olds in particular occupations. UCLA follows college graduates across all fields.
  • Employment is not the opposite of unemployment. A graduate who wanted a job in software or customer support and took an unrelated one counts as employed. They can show up in Stanford's shortfall without ever raising UCLA's unemployment rate.
  • A hiring freeze produces no layoffs. If a company quietly stops hiring juniors, nobody is fired. The effect is a door closing, and it takes time to reach official figures.
  • Different data. ADP's clients are not a mirror of the whole economy, and household surveys carry their own noise.

Put together, the defensible summary is narrow. There is early, real but unproven evidence that AI is narrowing the way into some white-collar professions, and there is no evidence yet of a broad unemployment crisis among young graduates. Both studies support the first half of that sentence, and neither contradicts the second.

The first rung

The finding with the longest reach is the split between young and experienced workers. Junior roles are how people become senior. They are where tacit knowledge, the thing this research suggests protects workers, gets built.

A company that keeps its experienced staff and stops training new ones can look efficient for years. The cost arrives later, as a shortage of people who know how to do the work and no pipeline to produce more. Slow failures like that are hard to see in monthly data and harder to respond to politically.

The data also lags. The models behind these studies have been replaced many times since November 2022, and METR's latest assessment suggests AI progress at one leading lab is now running at nearly twice its earlier pace. A 21 year old choosing a major today will enter a job market shaped by systems that don't exist yet. Past waves of technology created jobs nobody predicted, and that may happen again, but only 5% of Americans expect it.

Reasonable bets for someone starting out

No one can honestly promise a safe career path, but the evidence points somewhere. Work where AI complements people is where young workers' employment is holding up, so it is worth asking how a company uses AI before taking the job. Internships, real projects and time alongside experienced colleagues build the tacit knowledge that the Stanford data suggests protects people. And the ability to tell when a model's output is wrong is only possible if you understand the work yourself, which makes it valuable and hard to automate.

The single most practical step may be the least dramatic one. Stanford's Canaries Dashboard is designed to update monthly, which makes it a better guide than any one headline, and a better reason than any single study to stay in a field or leave it.