What ChatGPT Should and Should Not Write on Your Resume

Where to Draw the Line

Draw it at who supplies the facts. Write the raw material yourself first: the projects, the numbers, the decisions, even in messy bullet-point fragments that don't read well yet. Then hand that material to a model and ask it to tighten the phrasing, fix inconsistent tenses, order the bullets by relevance, and cut anything repetitive. That half of the work is compression and structure, and a model is genuinely useful at it.

The line sits exactly where the model would have to invent something to keep going. If you ask it to "make this bullet stronger" and it adds a percentage you never gave it, that is not editing your material; that is generating new material and presenting it as yours. The test is simple: did you supply every fact in the final sentence, or did the tool fill in the gap on its own?

That gap-filling is worth naming plainly, because it rarely announces itself. A model asked for a stronger sentence about a project will often reach for a plausible number, a typical outcome, a phrase that sounds like the kind of thing someone in that role would have achieved. None of that is a lie in the ordinary sense. It is closer to autocomplete operating past the edge of what you actually told it, and the fix is the same either way: check every number and every claim against what you know to be true before it leaves your document.

What a Model Can and Cannot Know About You

A model knows exactly what you paste into it and nothing more. The strongest evidence on writing assistance and hiring comes from a field experiment on an online freelance labor market, with nearly half a million jobseekers: those randomly given an automated writing-assistance tool on their profile text were hired 8% more often, with no evidence employers were less satisfied afterward ¹. That is causal, from one platform's setting, and a real reason to get help with the writing itself.

Read what was actually tested before generalizing it. The tool was explicitly not generative: the paper describes a grammar and style checker that suggests fixes to text a person already wrote, and states that unlike a chatbot, it cannot be prompted and cannot generate new content on its own ². The causal evidence supports polishing your own material. It says nothing about a model producing the material for you.

Honesty is not the only reason to write in your own voice. In a June 2026 ZipRecruiter survey of a panel of over 1,000 hiring decision-makers, 24% said they can almost always tell when a candidate used AI in an application, and a further 60% said they can sometimes tell ³. That measures belief rather than accuracy: it says nothing about whether that guess is ever correct, only that a large share of readers act as though it is. Belief alone is enough to change how a generic-sounding application lands, independent of whether a tool actually touched it.

The practical consequence has nothing to do with hiding the tool. What matters is making sure the writing carries your specifics rather than the smoothed, generic phrasing a model defaults to when it has little to work with. "Led cross-functional initiatives to drive impact" could describe almost anyone and almost nothing. "Rebuilt the intake form that cut average response time from four days to one" could only be written by someone who actually did that. The second sentence reads as human because it is specific, not because you avoided a particular tool.

Run the Two-Minute Test Before You Send It

Before you submit anything, pick three lines at random and talk through each one out loud for two minutes: what you actually did, what the number means, how you'd defend it if someone pushed back. If you stall on a line, that line is not ready to send, regardless of how well it reads on the page.

This test matters more than any question about whether using AI is allowed at all, because authorship is not the thing that gets an application rejected; a specific claim that fails a check is. And if an interviewer or a screener wants to verify something you wrote about your AI use specifically, the honest version of that conversation takes about ten minutes on real evidence, which is exactly why the evidence has to be real before the conversation happens.

Getting there starts with what you paste in. Start from your own rough bullets, not a blank prompt asking the model to write your resume for a given job title. Paste in the messy version: what you did, roughly, in your own words, with numbers if you have them and a plain "I don't have a number for this one" where you don't. That distinction matters, because a model asked to fill a gap will usually fill it with something plausible rather than admit the gap.

Ask for narrow, checkable edits rather than broad ones. "Tighten this to one line and use an active verb" keeps the model inside your facts. "Make this sound more impressive" invites it to add weight that was not there, which is exactly the request that produces a claim you did not make and now have to defend. Run the same bullets through a second pass asking specifically "does every sentence here contain only things I actually told you," which catches additions the first pass introduced and takes under a minute.

The practical habit that follows: keep the raw, ugly first draft you wrote yourself somewhere, even after the model has smoothed the language. It is your proof, to yourself as much as anyone, of where the facts came from. The same rule extends past the resume itself: a message to a recruiter, a cover letter paragraph, a short follow-up note. Your facts first, its phrasing second, and never the reverse order.

See a sample report >