A New-Grad AI Requirement Selects for Access, Not Skill

What does an AI line on a new-grad req select for?

The line selects for access, and after that for very little. Nearly every graduating student now uses AI, so the requirement removes almost nobody from the pile. The differences that survive sit upstream of the person: which university bought tools, which subject they studied, how advantaged their background is.

None of those describe how a candidate works.

In December 2025, Savanta polled 1,054 full-time UK undergraduates for the Higher Education Policy Institute. Some 95% reported using AI in at least one way, and 94% said they used it to help with assessed work ¹.

A line that 95% of undergraduates clear is not a bar.

That poll is British. The posting and labour-market evidence in the rest of this piece is American, and nothing here counts a US graduating year the same way.

So the 95% is read across a border, and the nearest US check is the resume.

Handshake compared the resumes graduating seniors uploaded to its own US platform. On that platform, Class of 2026 resumes name AI skills more than nine times as often as Class of 2022 resumes, with two-thirds of those mentions now coming from students outside computer science ².

Handshake is counting labels, not use or authorship, so the document cannot break the tie either. What is left to screen on if every new-grad resume is AI-built is the question the posting hands you, unanswered.

Where the HEPI survey does find gaps, they are not gaps in the person. The authors report persistent divides by subject, by socio-economic group in favour of more advantaged students, and by gender, with men still the more frequent users ¹.

Institutional supply is a fourth difference, and it sits wholly outside the student. Asked whether their institution currently provides generative AI tools, 38% of students said yes in 2026, against 23% in 2025 and 9% in 2024 ¹.

The graduate whose university provided nothing did not decline the tools. Nobody offered them.

Why does a graduate's AI experience predict so little?

Because experience measured at the moment of hire barely predicts job performance, across the studies pooled to test it. Chad Van Iddekinge and three co-authors, writing in Personnel Psychology in 2019, pooled 81 independent samples on prehire work experience. The corrected correlation with later job performance was .06, and with turnover it was .00 ³.

Experience with the relevant tasks did no better for job performance, at about .07 ³.

Van Iddekinge's meta-analysis covered work experience in general rather than experience with AI tools. It describes experience as a resume reports it, not experience built up in the current job, which is a separate and more favourable literature.

Entry-level postings have been asking for more anyway. PwC's 2026 Barometer looked at skills that appeared in US entry-level postings in 2025 and had barely appeared in 2019. In the most AI-exposed quarter of occupations, 52% of those new skills were ones previously asked of senior staff, against 7% in the least exposed quarter ⁴.

PwC is reading advert wording, not job content. An employer writing stakeholder management into a graduate posting has stated a demand the role may not contain, and an AI line on the same list reads the same way.

For a graduating student the line has even less to read. There may be no prior job to report, so the answer comes from coursework, a club, an internship or a side project.

The resume records the label rather than the act.

GPA can have the same shape, which is why whether GPA still predicts anything for entry-level roles now depends on the role. Where it has that shape, the proxy is cheap to read and thin on signal.

Ask what a model got confidently wrong, and how they caught it

One question does the work the requirement cannot. Ask the candidate for a specific occasion when a model handed them something confident and wrong, and how they found out. A graduate who has been through it answers with the claim, the check and what the check cost in time.

A graduate who has not moves to a principle about verifying outputs and stays there.

The answer needs no career behind it. A course citation that resolved to a real paper saying something else. A summer job where a generated formula quietly changed a total. A club budget that stopped adding up. All of them contain the same act.

Grade the catch, not the story.

How hard was the thing to see? A fabricated citation is caught by clicking. A plausible wrong number inside a correct-looking spreadsheet is caught only by somebody who went back to the source and re-derived it.

A story like this is easy to have ready, which is why how hard the catch was separates candidates better than whether a story exists at all.

Then ask the follow-up that is harder to rehearse: what would have happened if nobody caught it. A candidate who traced the consequence has thought about the work. One who says it would have been embarrassing has thought about the interview.

The question tests verification, and hiring someone who can tell when AI is wrong is a narrower brief than the posting admits.

Ask every applicant the same question, write the answer down in their own words, and score the catch on one scale everybody on the panel can see.

Where does the AI line belong in the posting?

In the description of the work, and out of the knockout questions. Written as a hard requirement, the line can put off applicants who read requirement lists as eligibility rules, and it does so before a name ever reaches you. Written as a description of what the job involves, it tells a graduate what to prepare.

Tara Mohr surveyed more than a thousand professionals, mostly American, about a job whose qualifications they did not fully meet and had decided not to apply for. The top reason, twice as common as any other, was that they did not think they would be hired without the qualifications and did not want to spend the effort. It was named first by 41% of women and 46% of men ⁵.

The least common reason of all was doubting they could do the job ⁵.

Mohr's survey is a blog-published self-report with no control condition, and it measures what people say about a decision already made. Read it for the shape of the problem, not for a coefficient.

So write the line as work rather than as a credential. Swap tool experience for the act you want evidence of: finding a claim a model got wrong and correcting it before it went out. Then say what counts as evidence, because a graduate's version of that came from a course, a shift or a side project.

The or-equivalent line is where that belongs, and on a new-grad req it carries more weight than the requirement above it.

Why the rewritten AI line can return the same shortlist

Rewriting a requirement moves the posting, and on its own a posting moves hiring only a little. The share of hires without a bachelor's degree rose about 3.5 percentage points inside 11,332 large-firm roles that had dropped the degree requirement ⁶. The Burning Glass Institute and Harvard Business School reached that figure by matching those roles against the career histories of roughly 65 million US workers.

Across the market it came to almost nothing. Only 3.6% of roles dropped the degree requirement at all, so the net effect was 0.14 percentage points, or roughly 97,000 workers a year out of 77 million hires ⁶.

Requirement text is not harmless. The wording can change who applies, while the rest of the loop decides who is hired, so an edited posting behind an unchanged interview returns much the same shortlist.

In a small number of entry-level roles the requirement earns its place. For an evaluations engineer, or a research assistant on a model team, the AI line names something specific, scarce and checkable. Near-universal AI use says nothing about whether a graduate has ever built an evaluation set.

Those postings should keep the requirement and describe it exactly.

Open the requisition and read the line again. A line asking for experience with AI tools describes the year group about to apply; it does not narrow it. A line asking for one wrong claim from a model, found and fixed before it shipped, does not narrow it either.

But it tells whoever runs the screen what to ask about. The screen is where the hire is made.

See a sample report >