What a Reviewer Can Actually Infer From a Resume's Prose

If you review applications, the honest answer is that you cannot establish from a document who or what typed it. The signals people circulate — clean grammar, uniform formatting, corporate phrasing, no typos — are signals of polish, and polish has several other sources. What prose can tell you is whether the page is specific, whether the claims are ownable, and whether the writer understood the role. Those are the things worth reading for, and they are worth reading for regardless of how the document was produced.

This matters because the authorship question is a dead end that feels like progress. You look at a fluent resume, feel a suspicion, and spend attention on a question you cannot answer instead of on the question you are being paid to answer.

Why the standard tells do not work

Take the four indicators that appear in almost every guide on this subject, and ask what else produces them.

Consistent, clean formatting. Produced by every resume template ever shipped, by every builder tool, and by anyone who has been told once that a resume should look tidy. A well-formatted resume is evidence that the applicant cared, or that they used a template. It is not evidence about authorship.

Flawless grammar and no typos. Produced by proofreading, by a friend reading it over, by a spell-checker, and — importantly — by careful applicants writing in a second language, who often over-correct precisely because they are aware of the risk. Treating fluency as suspicious inverts the ordinary meaning of care.

Generic, buzzword-heavy phrasing. Produced abundantly by humans. Resume convention has rewarded that register for decades; it is what people imitate when they are unsure. A page full of “leveraged cross-functional synergies” tells you the writer was copying a genre, not which tool they used.

A uniform voice with no personality. Produced by professional resume writers, whose entire service is smoothing an idiosyncratic history into a standard register, and by anyone who has read a “how to write a resume” article and followed it.

So each tell has at least two innocent explanations, and one of them — non-native fluency, template use, or having paid for help — correlates with groups you have no business penalising. A screening habit that quietly disadvantages careful writers and second-language speakers is a worse problem than the one it was meant to solve.

The one thing prose does reliably indicate

There is a real signal in the text, and it is not stylistic. It is whether any sentence on the page could only have been written by someone who did the work.

A generated draft, or a heavily templated one, tends to describe a category of work rather than an instance of it. “Managed stakeholder relationships to drive project delivery” describes thousands of jobs. “Ran the weekly release meeting for a team of six and owned the go/no-go call” describes one. The difference is not fluency; both sentences are fluent. The difference is whether the writer had a specific thing in mind.

That signal is useful because it is the same thing you actually want to know. A page of category descriptions is uninformative whether a person or a tool wrote it, and a page of specific ones is informative either way. You are no longer guessing at authorship — you are assessing whether the document tells you anything.

What to read instead, in order

Read for one concrete instance per role. Not per bullet — one is enough. Something with a scope, a decision, or an artefact attached. If a whole role passes without one, that is a finding you can act on.

Read the claims as claims. Which lines assert sole responsibility for a shared outcome? Which numbers arrive without any indication of where they came from? Overstatement is the failure mode worth catching, and it is catchable, because a claim with no evidence behind it behaves differently under questioning than one with evidence.

Read the fixed fields against each other. Dates, titles, and qualification status either cohere or they do not. A summary claiming a level of seniority that the timeline below it does not support is a substantive inconsistency, and generated summaries produce that pattern often — but so does ordinary self-flattery, so treat it as a question rather than a verdict.

Read for whether they answered the posting. Not keyword overlap — whether the document engages with what the role actually requires. This is the cheapest discriminator you have and it is the one most often skipped.

Detectors do not close the gap

The obvious move is to run the text through a tool that scores it for machine authorship. This does not help you, for a reason that has nothing to do with how good the tools are: a score is not a finding. You cannot show the applicant the score and ask them to account for it, you cannot record it as a reason, and you have no way to check whether it is right in this instance. What you have afterwards is the same suspicion with a number attached to it, which is worse than the suspicion alone because it feels like evidence.

Resume text also fights the premise. It is short, it is fragmented, it is written in a register that is deliberately impersonal, and much of it is proper nouns and dates. That is the least distinctive prose a person produces.

Resolve it later, where it resolves itself

The authorship question answers itself at the next stage, without anyone raising it. Someone who wrote their own page can walk through any line on it from memory: what the situation was, what they chose, what happened. Someone reciting an approved draft cannot, and the failure is unmistakable — not because they hesitate, but because the detail is not there to produce.

So the efficient sequence is: do not adjudicate authorship on paper, advance or decline on substance, and let the conversation test what the page claims. The questions a resume invites are largely predictable from the page itself, which means you can write two of them down while reading and ask them later. That is a better use of the same suspicion.

If you are on the other side of this

The applicant version of this question is narrower and has a different answer: you cannot know what any given employer looks at, so the only thing worth doing is making the document specific and defensible. And if your concern is what happens once someone does suspect it, that is a separate question about what a suspicion can and cannot lead to.

For reviewers, one last practical note: whatever you conclude, keep it out of the written record unless it is a substantive finding. “Seems AI-written” is an impression about authorship you cannot support. “No concrete example of ownership in three of four roles” is a reason. Only one of those survives being read back to you.