What Is Left to Screen On When Every New-Grad Resume Is AI-Built?

What can a new-grad resume still prove?

Almost nothing you could not have assumed. The attributes employers say they look for on a new-grad resume (problem solving, teamwork, written communication) are exactly the claims a model asserts fluently and for free, and nearly 90% of employers report reading for the first of those ¹. With no employment history behind them, those lines carry no dates, no manager and no scope. What survives is anything with an address.

Three things on a student's page still have one:

• An artifact you can open. A repository, a dashboard, a published piece, a campaign that ran.
• An affiliation someone answers for. An internship with a named supervisor, a lab, a team, a competition with a public result.
• A specific accomplishment a person can confirm. The federal accomplishment-record method exists for exactly this, and one variation requires formal verification through references, to discourage applicants from inflating what they submitted ³.

That last one carries a caution the same guidance states plainly: at entry level, give credit for accomplishments gained outside paid employment (school, volunteer work, community service) ³. A rule that counts only paid work screens out the students who could not afford an unpaid summer.

Grades are a separate question with a separate answer; see whether GPA still predicts anything at entry level.

Which artifact should you screen, role by role?

One per role, named in the posting, with an address you can open. Software: a repository with commit history. Analyst: the model file with its formulas live. Marketing: a campaign that ran somewhere public, with a date. Legal or underwriting: a written work product plus the packet it was built from. Four minutes inside one real artifact tells you more than a full page of claims.

Decide what you are reading for before you open the first one, because the artifact was probably built with an assistant too, and that is not the finding. What you can read is whether the choices inside it were made against a constraint: commits spread across weeks with messages that name a decision, rather than one squashed import; a spreadsheet whose formulas are still live rather than values pasted over them; a campaign with a stated audience and a result someone measured.

Then ask one question about a choice you can see: why one function is separate from another, or which number in the model moved the recommendation. That cannot be answered out of a document someone else's model produced. The question is the screen; the artifact is only what makes the question possible. Who actually built the portfolio project is the version of this you will meet most often.

Keep a second route open. A student with no public artifact may have a course deliverable, a lab result or an internship product they can share with permission, and refusing those screens on access rather than on ability.

Should you run a detector on the application?

No. Seven widely used detectors misclassified 61.3% of TOEFL essays by non-native English writers as AI-generated, and at least one flagged 97.8% of them, while essays by US students passed ². Pointed at a campus pipeline carrying international students, that is not a screen. It is an adverse-impact problem you built yourself, and every applicant used AI anyway.

The document is jointly written now, and that changes what it is evidence of rather than making it worthless. It still shows what a candidate chooses to claim, what they believe the role is, and how they organize a page. It was never a writing sample and it is certainly not one today.

A detector output is also a selection procedure in the legal sense. The anti-discrimination laws reach tests and screening tools by what they do rather than by what they are called, and a tool that disproportionately excludes a protected group has to be justified ⁴. A false-positive rate landing almost entirely on non-native speakers is the shape of a problem, not a signal.

Olive does not detect AI-written documents and carries no classifier; the question it answers is how a person works with an assistant when someone can watch, which is a different question from who typed the sentence.

Move the decision to a first-round act

Give the decision to something the candidate does, not to something they submitted. Work samples score high on both content and criterion validity, and applicants tend to perceive them as fair ⁶. The limit matters more for new grads than for anyone else: the same guidance says a work sample fits only where the competency is expected on entry, not where training follows the hire ⁶. So sample the judgment you will not train.

For a graduate that means the reasoning around the work rather than the craft inside it. A first round that holds up runs about 45 minutes on one case, identical for every applicant, with an assistant openly allowed, and asks for four things you can read afterwards:

• What would make an answer to this wrong? Written down before anything is generated.
• Which single claim in the brief did you go and check, and what did checking it change?
• What did you keep for yourself, and why that part?
• What did the assistant produce that you threw out, and on what grounds?

Grade against a key written before the first submission arrives, and grade the four separately, because one summed number hides the candidate who verified nothing and wrote nicely. Keep the round under an hour, because a longer one selects for free time.

If your funnel is heading this way regardless, skipping the resume screen and starting with the work sample is a cleaner design than bolting a fifth stage onto the loop you have.

How do you keep the replacement fair?

A rule that decides who advances is a selection procedure, whatever you call it internally. Anti-discrimination law reaches tests and screening tools by their effect, and a procedure that disproportionately excludes a protected group must be justified ⁴. Write the rule before the first application arrives, apply it the same way to everyone, and keep the artifact requirement satisfiable by more than one route.

Three checks are worth the time they cost:

• Count completions by group, not only scores. A round that only some candidates finish has already made a selection, and it made it before anyone read a submission.
• Give everyone the same clock, and publish it. An open-ended deadline tests free time.
• Accept an alternative artifact by default. Coursework, a lab deliverable, an internship product shared with permission. Requiring a public repository screens on who had a laptop and a spare summer.

There is a wider picture behind the application volume. Employment of workers aged 22 to 25 in AI-exposed occupations now sits about 19% below where it would have been had it tracked their less-exposed peers, and the gap comes from reduced hiring rather than from separations ⁵. Fewer openings mean more applicants each and a thinner basis for telling them apart, which is the pressure that makes a keyword rule tempting.

Non-technical roles need the same treatment with different artifacts; see screening AI judgment where there is no code.

See a sample report >