Graduates in AI-exposed fields are being hired less. So where did they go?

Three studies about AI and new graduates came out in the past six weeks, and on a quick read they can’t all be true.

• Stanford’s monthly tracker, updated on 21 October, finds that employment of 22- to 25-year-olds in the jobs most exposed to AI is down 12% since ChatGPT arrived, while in the least exposed jobs it’s up 8%.
• The US Census Bureau finds that graduates of the most AI-exposed degrees are 5 points less likely to be in work soon after graduating, and earn 13% less when they are.
• And economists at UCLA find that unemployment among new graduates this summer was 7.3%, lower than in 2024, and no higher in AI-exposed fields than in others once you allow for chance.

So are young graduates losing out to AI or not? Put the three side by side and they tell one story. It just isn’t the one in most of the headlines.

Before you look, have a guess

Stanford’s tracker follows millions of American payslips through ADP, the payroll company. It sorts jobs into five groups by how much of their work AI can do, then counts the young people each group employs, starting from November 2022, the month ChatGPT came out.

Young workers in the jobs AI can do most of

How many 22- to 25-year-olds each group of jobs employs, against November 2022, when ChatGPT came out (that month = 100). From American payslips.

Month	Least exposed jobs	Most exposed jobs
Nov 2022	100	100
Sep 2023	107	100
Sep 2024	108	95
Sep 2025	108	91
Sep 2026	108	88

Employment index (November 2022 = 100) for workers aged 22 to 25 in the least and the most AI-exposed fifth of occupations, and for workers aged 35 to 40: Stanford Digital Economy Lab and ADP Research, Canaries Dashboard, data to September 2026, released 21 October 2026.

Young workers in the most exposed jobs are 12% below where they started. Young workers in the least exposed are 8% above. Look at people in their late thirties in the very same jobs and both lines go up. Whatever this is, it’s happening to the young.

It also didn’t happen all at once. There was no cliff in 2023, the year everyone expected one, just a line drifting a little lower every few months, and still drifting. (The 19% you may have seen in the headlines, and that we’ve quoted before, is these two lines set against each other: 88 is 19% below 108.)

They didn’t stop working. They started somewhere else.

Fewer young people in exposed jobs doesn’t tell you what happened to the ones who’d have been in them. For that you have to follow people, and the Census Bureau did. It took graduates from colleges that award about three in every ten American degrees and linked what they studied to the payroll records of the jobs they took next.

Its most exposed tenth of degrees is mostly computing: computer science, computer engineering, information systems and their neighbours. Graduates of those degrees in 2024 were 5 points less likely than the class of 2022 to be in work in the three months after graduating, set against the least exposed degrees. The ones in work were paid 13% less, which the authors point out is about what graduating into a big recession does to you.

But here’s the part that explains the puzzle. About half of that 13% isn’t lower pay for the same job. It’s a different job.

A hundred graduates, two years apart

Where graduates' first jobs were six months after graduating, for every 100 in work. Switch between the class of 2022 and the class of 2024, and watch who moves.

The most AI-exposed degrees, mostly computing

First job in	Class of 2022	Class of 2024
Cafés, restaurants and hotels	3	5
Shops	6	8
Everywhere else	55	60
Professional and technical firms	26	20
Tech and media companies	10	7

The least AI-exposed degrees

First job in	Class of 2022	Class of 2024
Cafés, restaurants and hotels	7	8
Shops	8	8
Everywhere else	73	74
Professional and technical firms	8	7
Tech and media companies	4	3

First job: the main job held two quarters after graduating, by sector, for the classes of 2022 (July 2021 to June 2022) and 2024 (July 2023 to June 2024). Read back from Figure 5 of Graduating into Disruption, US Census Bureau (CES-WP-26-56, September 2026), the most exposed tenth of degrees against the least exposed six tenths. Rounded to whole graduates.

For every 100 of these graduates in work six months on, about 9 fewer started at professional and technical firms (consultancies, IT services, design and engineering practices) or at tech and media companies. About 4 more started in shops, cafés and restaurants, and the rest spread thinly across everything else. Switch to the least exposed degrees and almost nobody moves: no kind of work gained or lost more than about 1 in 100.

Why unemployment barely moved

That’s how all three studies can be right. Unemployment counts people with no job who are looking for one. A computing graduate behind a café counter isn’t unemployed. Neither is one doing a master’s because the job didn’t come. As far as the unemployment rate is concerned, both are fine.

So UCLA can find that unemployment among new graduates hasn’t jumped, while Stanford sees fewer young people in the exposed jobs and the Census sees them starting lower down and somewhere else. The word for it is underemployment: in work, but not the work you trained for. None of the usual monthly numbers measures it well. UCLA’s wider measure, which also counts graduates who want a job but have stopped looking, reached 10.4% this summer, the highest of the five summers it looked at, though not by enough to rule out chance.

Some did go back to study. Among the most exposed graduates, the share in graduate school rose from 16% to 18%, the Census team found, and a few more went self-employed. In Texas, the Dallas Fed found the same thing: graduates of more exposed degrees were more likely to go on to graduate school, and two in three computer science graduates who went back chose more computer science.

That may not be the escape hatch it looks like. New master’s graduates in the exposed fields saw their pay fall behind too.

Find your degree

The drop isn’t spread evenly. It’s concentrated in a handful of degrees right at the top of the exposure scale.

Every degree, one dot each

Across: how exposed to AI the jobs a degree leads to are. Up: how the class of 2024 did against the class of 2022, compared with the average degree. Switch between landing a job and starting pay, or type a degree to find it.

Degree	In work soon after graduating	Starting pay
Computer science	−6 points	−17%
Computer engineering	−7 points	−11%
Computer and information systems	−5 points	−13%
Finance	−2 points	−1%
Economics	0	−5%
Accounting	0	+3%
Marketing	0	−4%
English	0	−3%
Psychology	+1 point	+2%
Nursing	+3 points	+4%

Each degree's change between the classes of 2022 and 2024, compared with the average degree, for the 142 degrees with at least 1,000 graduates in the sample: US Census Bureau, Graduating into Disruption (CES-WP-26-56, September 2026), Appendix Table A3. Exposure is the paper's own: GPT-4 ratings of each task's exposure to AI (Eloundou and others), averaged over the jobs each degree's graduates go into.

Accounting, often the first degree named when people talk about AI and jobs, barely moved: its graduates found work as often as before, and started on slightly more pay than the average degree. Finance slipped a little. The big moves are computer science, computer engineering and information systems, where the first job got harder to land and paid less when it came.

How sure can we be?

None of the three studies claims to have proved that AI did this, and they’re right not to. 2022 was also the year interest rates went up, which hit tech hiring hard, and the end of a hiring spree that had nothing to do with chatbots.

Stanford is the most candid about it, and its numbers are worth a look.

How much of the gap survives a closer look

How much more or less employment grew from November 2022 to June 2026 than in the least exposed fifth of jobs, in points, by age. Each step allows for more of the other things that differ between jobs.

No adjustment

Exposure to AI	22–25	26–30	31–34	35–40	41–49	50+
Second fifth	+4	+4	+3	+5	+4	−3
Middle fifth	−1	+3	+2	+5	+7	+1
Fourth fifth	−15	−6	−5	0	+7	+2
Most exposed fifth	−18	−5	−1	0	+3	−1

Allowing for how many people in each job have a degree

Exposure to AI	22–25	26–30	31–34	35–40	41–49	50+
Second fifth	+7	+6	+5	+6	+3	−4
Middle fifth	+5	+3	+2	+6	+5	0
Fourth fifth	−7	−7	−5	+2	+4	−1
Most exposed fifth	−9	−5	−2	+1	0	−3

Allowing for degrees, interest rates and working from home

Exposure to AI	22–25	26–30	31–34	35–40	41–49	50+
Second fifth	+6	+5	+4	+6	+3	−3
Middle fifth	+6	+3	+2	+7	+5	+1
Fourth fifth	−5	−6	−4	+3	+5	0
Most exposed fifth	−8	−4	−3	+1	+1	−3

Employment-weighted estimates for firms in the sample since 2018, relative to the least exposed fifth of occupations: Stanford Digital Economy Lab, Canaries in the Coal Mine? (August 2026), Table 1, panels A, C and F. A square that is hatched rather than solid is too small, next to its margin of error, to rule out chance.

With no adjustment, employment of young workers in the most exposed fifth of jobs grew 18 points less than in the least exposed fifth. Allow for how many people in each job have a degree, and that roughly halves, to 9 points, and is no longer big enough to rule out chance. Allow for interest rates and working from home as well and it’s 8. Older workers barely register in any version.

That doesn’t make the gap go away. It means part of it belongs to other things that arrived at the same time as ChatGPT. Stanford also notes that the lines had started to drift apart a little before it, and that its payroll sample shows a sharper drop than the national surveys do. The Census team, for its part, points to separate work that tested the interest-rate story and found it doesn’t explain which sectors shrank, and concludes that AI is the most likely reason. Three teams, three different sets of records, all pointing the same way: the evidence is getting firmer, but it’s early.

If you’re choosing a degree now

Three things are worth taking from all this.

It’s the first rung, not the whole ladder. The pay gap for the most exposed graduates narrows from 13% to about 5% two years in, and older workers in the same jobs are doing fine. Getting in is harder. Staying in looks much as it always did.

Students are already adjusting. In Texas, enrolment in the more exposed degrees fell behind the rest between autumn 2024 and autumn 2025. That’s how a market corrects itself: slowly, unevenly, and a few years behind.

Aim for work where AI is the tool, not the worker. Stanford finds the drop is concentrated in jobs where people mostly use AI to do the work for them. Where they mostly use it to help, employment is flat or rising. Courses with real experience built in, like placements, apprenticeships and graduate schemes, matter more when the junior tasks that used to teach people are the ones being handed over. We wrote about that in where apprentices come from now, and about choosing in the first place in choosing a career at eighteen.

The bottom line

AI hasn’t put a generation of graduates out of work. What it seems to have done, so far, is move the first rung for a narrow band of them: from the office to the counter, for a while, on lower pay. That’s a real cost, and it lands on people who did everything they were told. But it’s a different problem from mass unemployment, and it needs different answers. More ways in, not just more warnings.