You Suspect a Resume Was AI-Written. Now What?

Guides on spotting AI-written resumes stop at the list of signs. That leaves out the part that decides anything: what you do next. You now hold a suspicion you cannot establish, about a question the employer probably has no policy on, concerning a candidate who may well have done nothing wrong. Acting on it directly is the mistake. The useful move is to leave the suspicion where it is, decide the application on its substance, and change what you ask for at the next stage so the question stops mattering.

This article assumes you have already read the page and something felt off. It is about the twenty minutes after that.

First, name what you suspect

“AI-written” covers several situations with very different weights, and collapsing them is why reviewers end up stuck.

The prose was drafted or polished by a tool, and the content is true. This is by far the most common case, it is now ordinary practice, and it is roughly equivalent to having a friend tighten your sentences. There is nothing here to act on.

The document is fluent but says nothing. Category descriptions, no instances, no scope. This is a real, actionable deficiency — and note that it is actionable as vagueness, which you can write down, not as authorship, which you cannot.

The document asserts things the candidate cannot support. Unsourced figures, seniority the timeline contradicts, ownership of shared outcomes. Also real, also actionable on its own terms, and testable later.

The application is not from a real, interested candidate at all — bulk-submitted, unrelated to the posting, or inconsistent in ways that suggest it was assembled without anyone reading the advert. This is a volume problem, not a writing-style problem, and its tells are mismatch and indifference rather than fluency.

Only the last three are grounds for anything, and none of them requires you to determine authorship. That is the useful discovery: every version of the suspicion that matters can be restated as something you can actually observe.

What not to do

Do not run it through a detector and treat the output as a result. Setting aside how well such tools work, the deeper problem is procedural: you cannot show a candidate a score and ask them to answer it, you cannot record it as a reason for a decision, and you have no way to check it in this instance. It converts an honest uncertainty into false confidence. The reasons prose analysis fails on resume text specifically make this worse, not better — resumes are short, fragmented, and written in a deliberately impersonal register.

Do not accuse. An accusation you cannot support costs you a candidate, invites a fair complaint, and — if you are wrong — has fallen on someone whose writing is simply careful. It is also unpleasant to be on the receiving end of, and word travels.

Do not write it in the notes. “Feels AI-generated” is an impression about authorship you cannot defend if the file is ever reviewed. “No concrete example of individual ownership in three of four roles” describes the same reading and survives being quoted back to you.

Do not decline on style alone. If the only thing wrong is that the prose is smooth, you are filtering for people who write awkwardly, which includes filtering against second-language speakers who over-correct and candidates who paid someone to smooth their history. That is a worse screen than the one you were trying to run.

What to do instead: change the next question

The productive response is not an investigation, it is an adjustment. Whatever you suspect, the fix is the same — make the next stage require something a document cannot supply.

Ask for one instance, out loud. Pick the strongest bullet on the page and ask what specifically happened: what the situation was, what they decided, what changed. Someone describing their own work produces detail without effort — names, constraints, the thing that went wrong. Someone reciting an approved draft produces the same sentence again in different words. You do not have to interpret hesitation; you are looking for whether the detail exists.

Ask about the parts a draft cannot know. Why they left, what they would do differently, who disagreed with them, what the hard part was. Generated content has no access to any of that.

Ask a scoping question with a number in it. How many people, over what period, what your share of it was. Vague claims collapse here quickly and honest ones get more precise.

Use a short work sample where the role justifies one. Not a large unpaid task — a bounded exercise close to the actual work. This is the only genuinely decisive tool in the list, because it measures the thing you care about instead of proxying it.

All of these are reasonable questions to ask any candidate, which is the point: you have converted a suspicion you could not act on into a stage design you would defend to anyone.

Write the questions down while you read

The practical version of all of this is that suspicion is a good prompt and a bad conclusion. When a line reads as unsupported, do not decide anything about it — write the question it raises in the margin and carry it to the interview. Most of the questions a resume invites are predictable from the page, so this costs nothing while reading and makes the conversation sharper.

By the end of a stack you will have, per advancing candidate, two or three specific things to test. That is a better artefact than a private list of who you thought used a chatbot.

The standard that makes the question moot

Consider what you would want to be true of a resume regardless of how it was produced: every claim on it traces to something real, the scope is stated honestly, the numbers have a source, and the candidate can walk through any line from memory. A document meeting that standard is fine if a tool wrote every sentence of it. A document failing it is a problem even if it was typed by hand.

So the screen to build is the one that tests that standard — which is a screen for whether a claim can be backed up and for whether individual contribution is distinguished from the team’s result. Authorship drops out of it entirely, and you stop spending review time on the one question your process cannot answer.

If you also handle cover letters, the equivalent question there has a different shape and a slightly better answer, because a letter makes claims about the reader rather than the writer.