Ava Bagherzadeh

Posted on Aug 17

Everyone used ChatGPT on their resume. That is exactly why yours is invisible.

#ai #writing #discuss #career

Everyone used ChatGPT on their resume. That is the actual problem, and it is not the one people think it is.

TL;DR — you are not being caught, you are being averaged. Ask a model to sound professional and it walks to the same centre of gravity every time, so a stack of applications stops having outliers. The fix is specificity, not disguise.

The worry is usually "will they be able to tell?" That is the wrong worry. Whether a recruiter consciously identifies your resume as AI-written barely matters. What matters is that AI writing converges. Ask a model to make a resume sound professional and it walks toward the same centre of gravity every time: the same verbs, the same rhythm, the same shape of achievement bullet. Do that at scale and a stack of applications stops having outliers.

You are not being caught. You are being averaged.

That distinction changes the fix. If the problem were detection, the answer would be disguise: swap words, break up sentences, add a typo. People genuinely do this, and it does nothing, because the flatness was never in the word choice. If the problem is convergence, the answer is specificity, and specificity is something a model cannot invent on your behalf because it does not have your material.

Here is what actually happens, what it looks like on the page, and what to do instead.

Part one: the tells, and why each one is a tell

I want to be precise about the mechanism for each of these, because "recruiters can spot AI" is folklore and the individual patterns are not.

1. Self-description in the third person of a brochure

"Results-driven professional with a passion for excellence." "Dynamic team player."

Nobody has ever said this out loud about themselves. The reason models produce it is that resume corpora are full of it, and it is the safest possible output: unfalsifiable, uncheckable, offensive to no one. That safety is exactly what makes it worthless. A line that could sit on any resume in the stack does no work in a stack.

Instead: delete the summary or make it a claim only you could make. "I have spent four years on billing systems, which means I have been on call for the thing companies least want to break." That is a sentence with a person behind it.

2. Perfect grammar with no rhythm

AI prose is evenly weighted. Every sentence is roughly the same length and carries roughly the same emphasis. Human writing has spikes. Short sentence. Then a longer one that explains why the short one mattered.

This is the subtlest tell and the one people never fix, because nothing is wrong. It reads like a committee approved it, and something a committee approved is something nobody chose.

Instead: read it aloud. Where you would naturally pause, break the sentence. Where you would emphasise, cut words.

3. Verb inflation

"Spearheaded." "Orchestrated." "Leveraged." "Synergised." "Championed."

The tell is not that these words are bad. It is density. When every one of eight bullets opens with a heroic verb, the reader stops reading the verbs, which means every bullet now begins with noise. Worse, inflated verbs invite a question you may not want: if you spearheaded it, who followed?

Instead: use the plainest true verb. "Built." "Owned." "Cut." "Wrote." "Fixed." Plain verbs read as confidence, because only someone comfortable with their work describes it without decoration.

4. Numbers with no denominator

"Improved efficiency by 30%."

Thirty percent of what, measured how, over what period, compared to what baseline? A model produces this shape because the shape is what achievement bullets look like, so it fills the slot with a plausible number. It has no idea whether the number is true.

This is the tell that actually costs you interviews, because a recruiter who half-believes it will ask about it, and you will not have the denominator either.

Instead: either give the full frame or drop the number. "Cut the nightly batch from 6 hours to 40 minutes by replacing the per-row lookup with a single join" needs no percentage. It is more impressive and it is defensible, and defensible is the whole point since you have to survive a conversation about it.

5. The cover letter that is a template with a variable in it

If your cover letter names the company once and never says anything that could only be said about that company, it reads as generated whether or not it was.

Instead: one specific sentence, anywhere in it. Something about the product, a decision they made, a constraint their domain has. It does not need to be flattering. "You are doing document parsing in a regulated industry, which means your hard problem is not accuracy, it is being able to explain a wrong answer" gets read to the end.

6. Identical architecture across a whole stack

Summary, Experience, Skills, Education, same headers, same spacing, same order. Individually invisible. In a pile of fifty, it is the texture of the pile.

Instead: put the thing that makes you plausible for this role in the top third, whatever section it belongs to. If a side project is the strongest evidence, it goes above a job that is less relevant.

7. Skills sections generated from the job description

A model handed a posting will helpfully list every technology in it. Now your resume claims eleven things at equal confidence and a recruiter cannot tell which two you would survive an hour of questions on.

Instead: rank honestly, and split. Things you would defend under pressure, and things you have touched. That second list is not a weakness; it is the sentence "I know the difference", which is rarer than any of the skills.

8. Achievements with no failure anywhere near them

An all-triumph resume is not read as flawless, it is read as unexamined. Interviewers are trying to work out whether you know what went wrong, because that is what predicts the next six months.

Instead: one line, somewhere, that admits a real constraint. "Shipped it three months late because we underestimated the migration" is a sentence that gets asked about, and you will enjoy answering it.

9. Tense and person drift

"Managed a team of five. I am responsible for the roadmap. Responsible for hiring." Three grammatical people in three lines. This happens when you generate sections separately and paste them together, and it is the closest thing to a literal fingerprint on this list.

Instead: read the whole document top to bottom once. This catches it immediately and almost nobody does it.

Part two: the part nobody tells you, which is mechanical rather than stylistic

You can fix all nine of those and still never be read, because before a human forms an impression, a parser has to be able to extract you.

Check that your file contains text. A PDF exported from a design tool can be an image of a resume. It looks perfect. It parses to nothing. Open the file and try to select a sentence with your cursor. If you cannot, no system can read it, and every stylistic improvement above is moot.

I am not speculating. Our own export had exactly this defect, and I found it by counting the text-showing operators in the file rather than trusting that it looked correct on screen. It looked perfect.

Understand what a knockout question is. Most application forms carry a small number of fields that can filter you by rule before a human is involved: degree, work authorisation, years in a named tool. These are worth more of your attention than the entire skills section. Where a form allows "in progress" plus a date rather than a bare no, use it.

Know that one email is one candidate profile per company, not one per job. Apply to six roles at the same employer and someone opens a single record with six applications and whatever you typed the first time. The first application at a place you actually want is the one the other five inherit.

Part three: how to use the model so it helps instead of averaging you

The failure mode is asking for output. The fix is asking for interrogation.

Do not paste the job description and ask for a tailored resume. That is what produces convergence: the model has the posting and nothing of yours, so it writes toward the posting's own vocabulary and invents the rest.

Do paste your own raw material, ugly and unedited, and have the model ask you questions. "Here are messy notes about what I did for four years. Ask me ten questions a skeptical interviewer would ask, and tell me which claims I have not evidenced." You answer in your own words. Now you have specifics, in your voice, and the model never had to invent anything.

The division of labour that works: the model finds gaps, you supply facts, the model tightens phrasing, you approve the voice. It never originates a claim. The moment it originates a claim, you are defending someone else's sentence in an interview.

One more, and it is the highest-leverage thing here: find the requirements that repeat in the posting. If something appears in the intro paragraph and again in the bullets and again in the requirements list, it survived somebody's internal review and it is a real must-have. Something mentioned once near the bottom is a wishlist item. Tailoring to the two or three repeated things beats rewriting everything, and it takes ten minutes.

Part four: the thing that makes all of this feel unfair

You can do every item above and still hear nothing. That is not evidence that you got it wrong.

Most hiring systems only send mail on two events: an interview invitation, or a rejection somebody actually triggered. There is very often no event at all for "we looked and moved on". Roles get filled internally. Requisitions get frozen. So silence carries almost no information, and people rewrite their entire resume off a signal that contains nothing.

If you are going to change something, change it off a signal that exists. The cheapest one: take ten roles you genuinely want, apply normally, and separately get one of them in front of an actual human. If the ten go quiet and the human turns into a conversation, your resume is fine and your channel is the problem. If the human also goes quiet, now look hard at the resume. Almost everyone rewrites first and never learns which of the two was broken.

And there is a failure below that one that nobody checks: whether the application arrived at all. A submit can silently fail, and the confirmation page you saw was rendered by the same software you typed into. It is not the employer saying anything.

Where I am coming from

Full disclosure, I work on AI Applyd. It scores your resume against a specific posting, rewrites it per role from your own material rather than from the posting's vocabulary, and fills and submits on the employer's own hiring system across twelve of them: Greenhouse, Lever, Ashby, Workday, iCIMS, Personio, Teamtailor, SmartRecruiters, Recruitee, Breezy, Workable and Rippling. Then it waits for that employer's system to confirm the application arrived before it counts anything as sent, and shows you landed, still verifying, or did not land.

Free to start, no card, at https://aiapplyd.com

The honest caveat: none of this, mine included, makes a posting want your background. What it can do is make sure the version of you that arrives is specific, defensible, machine-readable, and actually delivered. Four things that are entirely within your control, in a process where almost nothing else is.

DEV Community