The case that AI could destabilise societies through work runs in three steps: displacement of work at scale, then mass hardship, then social and political breakdown. It is the slow path to the kind of failure that nuclear war or a pandemic reaches quickly. The question this year is whether the first step has started. The answer depends heavily on where you look.
The reassuring numbers
The aggregate evidence is consistent. Goldman Sachs estimates that a 10% rise in an occupation's AI exposure is associated with only about a 0.1 percentage point drag on annual headcount growth.1 Anthropic's own analysis of US survey data finds no effect on unemployment in the most exposed occupations; the company has an interest in the answer.2 The Stanford Digital Economy Lab, whose work is the most cited on the other side, also says it sees no widespread economy-wide displacement.3
Layoff announcements tell a similar story once scaled. Employers have cited AI for 120,136 announced US job cuts so far in 2026, about 21% of the total and the most common stated reason for much of the year, according to Challenger, Gray & Christmas. That is still under 1% of the layoffs the Bureau of Labor Statistics counts over the same months, and Challenger itself notes that "AI" often means a budget being redirected, not a job being done by a machine.4
The entry level
The articles saying AI is not affecting jobs are right about the whole economy. They miss where the pressure is landing: on people trying to get their first job.
- US graduates. Unemployment among recent college graduates aged 22–27 was 5.6% in the second quarter of 2026, against 3.6% in 2019, and underemployment edged up to 42%, according to the New York Fed. Recent computer engineering and computer science graduates had some of the highest unemployment of any major, at 7.8% and 7.0%.5
- UK graduates. Adzuna's count of advertised graduate vacancies fell to 8,383 in July 2026, a record low for the series, down 45.6% in a year and from more than 55,000 in 2017; 2.14 jobseekers were chasing each one. Non-graduate entry-level vacancies fell 8%. Indeed's narrower measure fell 7%, and the Institute of Student Employers, which surveys large employers, recorded an 8% fall in 2024/25 with a further 7% projected, and a record 140 applications per vacancy.6
- Young workers in exposed jobs. Stanford's payroll-based research finds employment of 22- to 25-year-olds in the most AI-exposed occupations now 19% below that of less-exposed peers, up from 13% a year earlier, with the adjustment coming through fewer hires rather than layoffs.3
Software: the sharpest case
Developers' accounts of how hard it has become to find work are borne out by the data, with one twist. Indeed's Hiring Lab reports US software-development postings up nearly 15% since early 2025 even as overall postings fell, but still about 27.5% below their pre-pandemic level. Of the increase between May 2025 and May 2026, 71% was in senior roles and 37% was in postings with AI in the job title.7 Stanford's payroll data puts employment of programmers aged 22 to 25 down nearly 20%.3
That is the pattern one would expect if AI coding tools let experienced developers do more: demand holds up for people who can direct and check the work, and falls for the people who used to learn by doing it.
| Measure | Latest | Direction |
|---|---|---|
| US recent-graduate unemployment | 5.6% (3.6% in 2019) | Up |
| US recent-graduate underemployment | 42% | Up |
| UK advertised graduate vacancies (Adzuna) | 8,383 in July, −45% in a year | Sharply down |
| Employment of US programmers aged 22–25 | about −20% | Down |
| US employers' graduate hiring plans (NACE) | +5.6% for the class of 2026 | Up |
| AI-cited US job cuts, 2026 to date | 120,136; under 1% of all layoffs | Falling monthly |
Is it AI?
Not only, and possibly not mainly. The strongest counter-evidence comes from the New York Fed, whose June 2026 study attributes nearly two-thirds of the rise in young graduates' unemployment to remote work: employers are reluctant to train new graduates in jobs done from home. The authors note that the rise began before generative AI spread and persists when AI exposure is controlled for.8 In Britain, Adzuna itself points to higher employer National Insurance contributions and minimum-wage rises, which make junior hires more expensive.6 And not every survey agrees on direction: the National Association of Colleges and Employers' spring update has US employers planning 5.6% more graduate hires this year, led by the largest firms.9
Remote work and AI are not independent explanations. The jobs that can be done remotely and the jobs AI can do overlap heavily, which is one reason the two are hard to separate. The defensible summary is that graduate hiring has weakened sharply in some markets and occupations, that AI is one cause among several, and that software is where the AI signal is clearest.
Does it pass the mechanism test?
The first step is visible, at the entry level. It is not yet mass displacement, and nothing in the data shows the second or third steps. But the entry level is where careers, skills and expectations are formed, and a generation that cannot get a foothold is a political fact long before it is an economic statistic. The indicators to watch are graduate unemployment and underemployment, the entry-level share of postings in exposed fields, and whether the squeeze spreads from new entrants to mid-career workers. If it does, this stops being a projection.