How to Write a Resume That Does Not Sound AI-Written

Recruiters can name the tells of an AI-written resume down to the phrase. This finds every one in your document, splits them into the ones you can change and the one you cannot, and hands the rest back as questions.

By Zaira

01The Problem

Some of it is a drafting problem. Most of it is not.

"It reads as AI" is a real observation with a documented cause — and it points at three different faults, only one of which you can edit.

The tells are named, so "it sounds like AI" is not a feeling

Recruiters describe them without hedging: American spellings in an otherwise British document, emphasis doing the work a number should do, "proven track record" and "trusted right hand" and "results-driven", adjectives standing in for evidence, description of responsibility instead of proof of it, and a formatting sameness that suggests one template and one prompt. That list is specific enough to check a document against, which turns a vague worry into a fixable set of lines.

Editing the words does not change who wrote them

Measurable work on language-model text finds writing fingerprints that identify which model produced it, hold across subject areas, and stay consistent within a model family — and they survive being explicitly asked to write in a different style. So part of what you can see on your own document is not a phrase you used badly. It is a signature, and no amount of rephrasing that sentence removes it, only relocates it.

Removing the tells changes which pile you land in

There is a documented tension worth naming rather than resolving. Research on screening models finds they prefer resumes the same model wrote, at rates between 67% and 82%, so a de-AI'd document is not a neutral document — it is one that has moved. Meanwhile human reviewers are explicitly checking for AI phrasing. You are optimising for two readers with opposite preferences, and the honest method tells you which lines serve which one instead of promising a document that satisfies both.

02The How-To

The tells audit, step by step

Copy it into Zaira and paste your draft plus the facts behind it. It lists the tells with the line each one is on, sorts them into fixable and unfixable, and rewrites only the fixable lines.

Audit an AI-assisted draft for machine tells
Automate it with Zaira

Step 0: Set the rules, then take my evidence

Act as my resume editor, not my rewriter. I am giving you a draft that was written with AI help. Do not make it more impressive — I want the machine parts removed and my parts sharpened. - The draft: [paste] - What I actually did, in my own words: [list, plain, no adjectives] - Numbers I can actually stand behind: [list] - The job I am actually going for: [posting text or link]

Step 1: Find the tells before you touch the prose

Go line by line and list every marker of machine-written text: - American spellings where the rest of the document is British or otherwise - bold or emphasis doing the work a number should do - stock phrasing: proven track record, trusted right hand, dynamic, results-driven, leveraged, spearheaded, utilized, passionate - adjectives standing in for a number - description of responsibility instead of evidence of it - a formatting sameness that suggests one template and one prompt Show me the table first: tell | where it is | why a reader notices it. Do not rewrite anything yet.

Step 2: Separate what I can fix from what I cannot

For each tell, tell me which of the three it is: - mine to fix by supplying a fact I already have - mine to fix by cutting the sentence entirely - a fingerprint of the model, which survives any instruction to write differently and which I cannot remove by editing words Be blunt about the third category. Never tell me a rewrite fixes it, and do not pad the first two with lines you are unsure about.

Step 3: Rewrite only the lines that change

For each fix in the first two categories, show the original line and my replacement side by side. Use only facts from Step 0. Where a line needs a number I did not give you, leave the line alone and move it to Step 5. Do not touch any line that is not in the Step 1 table.

Step 4: Read it back

Tell me what job title the rewritten draft is now claiming I am good at, and whether that matches what I put in Step 0. If it does not, say so plainly rather than smoothing it over.

Step 5: What I still have to supply

List every sentence that depends on a fact you do not have, as a question I can answer tonight. One question per line. No sentence gets written for me.

Output rules

Every replacement traces to something in Step 0. Never add a number, a title, a scope, a credential, or a tool I did not give you.
03Why People're Using

Stop polishing a draft that was generic first

Three outcomes, one loop. You stop guessing which sentences are weak, you stop rewriting lines that were never going to change, and you finish the document with five questions rather than five paragraphs.

The tells, with the line each one is on

Step 1 produces a table before any editing happens, so the diagnosis is a thing you read rather than a change you discover afterwards. Telling "it feels AI-written" from "line 14 is an adjective where a number belongs" are different conversations, and only the second one can be finished.

The category you cannot edit

Step 2 names the third category explicitly and refuses to pretend a rewrite fixes it. That is the part of this page that does not exist elsewhere: knowing which lines are structurally not yours to change is what stops you spending three evenings rephrasing a sentence that was always going to read as generated.

Five questions instead of five paragraphs

Step 5 returns every unsourceable sentence as a question you can answer tonight, which is usually faster than the rewrite you were about to do by hand. It is also the only version of this that stays defensible in an interview, because every line on the finished document traces to something you actually said.

04FAQ

Frequently asked questions

The questions people ask when the thing they are worried about is how the document reads rather than what it says.

The documented ones, which is a shorter list than people assume. American spellings in an otherwise British document. Bold or emphasis carrying a claim that should carry a number. Stock phrasing — proven track record, trusted right hand, results-driven, dynamic, leveraged. Adjectives standing where evidence belongs. Description of responsibility rather than proof of it. And a formatting sameness across a document where every section has the same rhythm, which suggests one template and one prompt rather than one person's writing. The last one is the tell that most often gets a document dismissed, because a recruiter scrolling a stack cannot investigate it — they only notice it.

Mostly, and the parts you keep are the parts worth keeping. A stock phrase is doing the work a specific number should do, so replacing it makes the line stronger rather than thinner — that is the common case. Adjectives are the same: "spearheaded a process improvement" carries less than the actual scope of what you changed. The exception is any line whose only content was the fluency, and those are rarer than the panic suggests. The method leaves those alone and reports them as unfixable rather than quietly deleting your experience, because a thin true document beats a full one you cannot defend.

This is genuinely contested and no one should tell you otherwise. Screening models have been measured preferring resumes the same model wrote, at rates between 67% and 82%, which suggests stripping machine phrasing may also remove the signature those models reward. Against that, recruiters explicitly check for AI phrasing and name the tells above. So you are optimising for two readers with opposite preferences. What helps is knowing which lines serve which: a specific number serves both, stock phrasing serves neither, and the residual signature is not something you can edit either way. Decide which reader you are optimising for and stop paying for the other one.

Increasingly, yes — a survey of close to a thousand hiring leaders found 92% of recruitment leaders believe AI-generated resumes are now commonplace, and recruiters describe the tells in the first answer. But commonplace is not disqualifying. The same research found that while two thirds of employers still screen on the resume, only 2% call it their most trusted signal, and 64% have hired someone who misrepresented themselves on one. What a reviewer is now looking for is whether the document reads as true, not whether a tool touched it. That is why this method never invents and returns unsourceable lines as questions — the failure mode being watched for is overclaiming, not assistance.