03 · Labour market

AI-exposed jobs and the young

Since ChatGPT, hiring of young workers into the most AI-exposed jobs has dropped sharply, and their employment there is down 8.6%. But the whole effect sits in remote-capable roles, so treat it as a warning sign rather than proof of AI. This track follows Stanford's ‘canaries’ study (Brynjolfsson and co-authors, 2025), rebuilt on UK data.

Exposure

Q5 is the desk and clerical work AI can do (data entry, payroll, insurance). Q1 is manual work.

Two views to compare

One chart counts everyone in these jobs; another counts only the past year’s hires. Hiring falling while the total holds means the change is at the hiring gate, not layoffs.

The big caveat

These jobs are also the most home-workable, so an AI effect and a remote-work effect can’t be fully told apart.

-8.6%
Young in most-exposed jobs
age 22-25 · employment vs 2022 Q4
-28%
Hiring into exposed jobs
age 22-30 recent hires vs 2022 Q4 · the mechanism
+5.0%
On-site young, WFH-adjusted
decline vanishes among on-site workers
+1.6%
Older workers (control)
age 35-49, most-exposed · ~flat
What’s happening: employment by AI-exposure (2022 Q4 = 100)
Counts everyone in the job, so the total moves slowly. Use the dropdown to change the age band: the gap sits in the youngest, and closes for older workers.
Is it AI? Young employment where AI does the task vs assists
Where AI does the task, employment fell; where it only assists, it grew.
Why: the hiring door. Recent hires by exposure (toggle the age band)
Counts only the past year’s hires, so it reacts fast: down 28% into the most-exposed jobs vs 15% the least. Slide the age band up and the gap fades by the 40s.
The catch: among on-site young workers, the decline disappears
Restrict to on-site young workers and the most-exposed line recovers. The fall lives in home-workable roles.

Data & methodology

This track follows an early warning sign from Brynjolfsson and co-authors (2025): if AI is changing hiring, it should show up first among young workers in the jobs AI can do most of.

The data

Labour Force Survey microdata from the UK Data Service: 19 quarterly files with single-year age, four-digit occupation, and a person weight that grosses the sample up to the population. Employment is that weighted headcount, indexed to the quarter before ChatGPT (late 2022) and smoothed over four quarters to remove seasonality and survey noise.

Measuring AI exposure

Each occupation gets an exposure score from the ILO's 2025 generative-AI index, mapped onto UK occupations and sorted into five equal-employment groups, from least exposed (Q1) to most exposed (Q5). For exposed jobs we also split automation, where AI does the task, from augmentation, where it assists, using the Anthropic Economic Index.

The hiring measure

Recent hires are people with under a year at their employer. Counting only them strips out long-tenured staff and shows the hiring decision directly. The hiring chart widens the young band to ages 22 to 30 so the recent-hire sample is large enough to split five ways; the stock charts keep the study's 22 to 25 band, where the entry-level signal is sharpest.

What it can and cannot say

This is descriptive. It shows patterns and their timing, not cause. The most-exposed jobs are also the most home-workable, and that overlap is not a coincidence: AI has landed hardest on knowledge and desk work, which is the work that can be done remotely. So an AI effect cannot be cleanly separated from a remote-work one. That is a reason for caution, not a reason to dismiss the signal. A firm-level causal test would need the ONS Secure Lab, and the 2023 to 2024 survey samples are noisy, so read the trajectory rather than single points.