The first generation of GPT-written resumes is now well-known to recruiters. They can spot one fast, and the spotting isn't because the resume is bad — it's because the resume has a specific signature. Same rhythm, same buzzwords, same em-dash patterns, sometimes numbers that don't cohere across bullets.

Using an LLM on your resume is fine. Letting it write the final version isn't. This post is about the line between the two and how to use the tool without producing the recognizable artifact.

Where GPT helps and where it hurts

The useful frame is that LLMs are good at editing and poor at inventing. Specifically: given a rough bullet you've written, GPT can usually tighten it, suggest a stronger verb, or compress two lines into one. That's a sharpening function and the model is genuinely good at it.

What it's bad at is generating content from facts it doesn't have. If you give it your job title and ask for bullets, it will produce plausible-sounding bullets — but the numbers will be invented, the projects fictional, the outcomes pulled from training data about similar roles. Some of these will look right at first; almost none will survive a behavioral interview where a hiring manager asks for specifics.

The line to hold: GPT can rewrite anything you've already written. It should not write anything from scratch except draft scaffolds you intend to fill in yourself.

A specific antipattern: pasting a job description and asking GPT to generate a resume tailored to it. The output will be keyword-rich, plausible, and largely fabricated. The recruiter notices, the interviewer notices, and the hiring manager notices when they ask you to walk through the project you supposedly led.

For the parallel question on cover letters, see ai-cover-letter-detection-truth. For how AI screeners on the other side process resumes, see ai-resume-screeners-how-to-think.

What recruiters notice

Recruiters who screen hundreds of resumes a week have developed an internal classifier for AI-written content. The tells are specific and increasingly consistent.

The most common is identical sentence rhythm across every bullet — verb, action, quantified outcome, verb, action, quantified outcome. Real resumes have variation: some bullets are descriptive, some are quantified, some are short, some are long. GPT defaults to a single template and applies it uniformly. The result is mechanical and the recruiter feels it before they articulate why.

The second is default buzzword choice. GPT reaches for "leveraged," "spearheaded," "orchestrated," "drove," "streamlined" unless explicitly told not to. These aren't wrong words, but their density in a generated resume is unnatural. A real candidate uses one or two of them; a generated one uses six.

The third is numbers that don't cohere across bullets. The model doesn't track the implicit story your career is telling. So bullet one says "supported 100M users," bullet two on the previous job says "scaled to 2M MAUs," and a senior recruiter reads it twice and notices the implied user-count regression makes no sense. Manual review catches this; GPT doesn't.

The fourth is summary sections that read like LinkedIn Abouts. "Passionate about delivering scalable solutions in cross-functional environments." It's not wrong; it's just indistinguishable from every other generated summary.

The fifth is small but telling: identical em-dash usage in every bullet, often in places where a comma would be more natural. GPT's punctuation patterns are remarkably consistent across outputs.

The workflow that produces real-sounding output

The workflow that works is human-first, GPT-second.

1. Draft each bullet yourself first. Ugly is fine. "Built data pipeline thing for sales team, it processed about a billion rows a day eventually." Get the facts down, your voice, your phrasing.
2. Ask GPT to tighten one bullet at a time. "Rewrite this in 22 words or less, keeping the specific numbers and removing 'leveraged' and 'spearheaded.'" Specific constraints produce useful output. Vague prompts ("make it sound better") produce the recognizable artifact.
3. Reject anything it added that you didn't supply. If GPT adds a number, a stakeholder, or an outcome you didn't tell it about, take it out. Especially numbers.
4. Vary the structure manually. After the rewrite pass, go through and break up the sentence rhythm. Different bullet lengths. Different starting verbs. Don't let the file read like a single voice.
5. Read the final version out loud. This is the single most effective de-AI-ification step. If a sentence sounds like a website wrote it, rewrite it.

A specific use case that works well

GPT is unusually good at one task: given a tired bullet, generate five different action-verb alternatives so you can pick. "I want to replace 'managed' in this bullet — give me five verbs that are more specific and don't include 'leveraged' or 'spearheaded.'" That kind of constrained prompt produces useful options and isn't fabricating anything.

For the underlying verb work, see resume-action-verbs-that-arent-cliche.

On AI-detection tools for resumes

There's a small industry of "AI detector" tools that claim to flag GPT-written resumes. Most of them are unreliable — false-positive rates on human-written content are high enough that no serious recruiter uses them as a filter.

What recruiters do use is the experienced eyeball pattern described above. That's the actual signal, and that's what you're trying to avoid producing.

If you're worried about a specific resume, the test is simpler than running it through a detector: read it out loud. If three bullets in a row sound mechanically identical, you have an AI artifact regardless of what tools say.

The honest case for using GPT anyway

The above isn't an argument against using LLMs on your resume. Used well, they save real time on rewrites, verb selection, summary drafting, and grammar cleanup. The argument is against the specific failure mode — letting the tool write what you should be writing yourself.

The right mental model: GPT is an editor you have on standby, not a ghostwriter. Drafts are yours; tightening is shared.

What this isn't

A few clarifications:

• It's not a recommendation to avoid AI entirely. Used as an editor, it's a real help.
• It's not a guarantee of "human-written" status. The signature can be removed with effort; the goal isn't to pass an AI detector, it's to write a resume that represents your actual career.
• It's not the same problem on every resume. Some industries (academia, legal, medicine) are less sensitive to AI signals because the resume formats are more standardized to begin with.

The short version: edit with GPT, don't generate with it. Draft your own bullets, ask it to tighten, reject what it adds. Vary the sentence rhythm manually. Read it out loud before sending. The line between "used a tool" and "let a tool write it" is the line between a recruiter reading on and a recruiter pattern-matching the artifact.