Our own head-to-head, with receipts

We tested pasting resumes into a chatbot. It helped two, hurt two, and did something stranger to the fifth.

The honest headline first: on the two weakest resumes in our test, a chatbot’s unguided rewrite raised the score by double digits. That’s the result every competitor page like this one is built to avoid saying. Here’s the rest of what happened, including the parts that argue against us.

Free · No sign-up · No credit card

The actual finding: it’s a coin flip you can’t inspect

“ChatGPT makes your resume worse” is false on our own numbers — it lifted the two weakest resumes we tested. “ChatGPT makes your resume better” isn’t true either — it degraded both resumes that were already strong, by thinning out keyword and skills evidence that was already doing its job, with no rubric telling it that evidence mattered. Add a fabricated number, an invented contact detail, a decade-shifted timeline, and a punctuation habit (below) that every single rewrite carried and no original did, and the pattern is consistent: naive rewriting can genuinely help, and nothing about the process tells you in advance which result you’re going to get. You only find out if something measures the before and the after — which is the actual product here, not a score.

Five resumes, before and after

Each was pasted into a chatbot with the prompt a real person types — “improve my resume,” nothing else — and scored with our own deterministic pipeline both times.

ResumeStarting pointBeforeAfterResult
J. Alex WangWeak — vague summary, few concrete claims6579Chatbot won

Genuinely better. The summary got specific and every real number survived.

Jamar JacksonWeak — similar shape to Wang's6277Chatbot won

Also genuinely better — the second of two clean wins for the chatbot.

Jamal K. WashingtonStrong — metrics and keywords already in place8265Chatbot lost

Worse. Already-strong keyword and skills evidence didn't survive the rewrite intact.

Karla RomeroStrong — a well-built student resume8073Chatbot lost

Worse, for the same reason as Washington's — the rewrite thinned out what was already working.

Mohamed SinghMiddle of the pack7573Roughly a tie

Roughly a wash on the score — but this is the one with the fabricated number and the invented email below.

Five resumes, single-shot, one model standing in for a general chatbot, run in August 2026 — a demo with our own receipts, not a study. See the methodology note in the FAQ below for what that means and doesn’t mean.

What the rewrites actually did

Three excerpts, verbatim, from mock resumes built for this kind of testing — one clean win, and two ways the same run went wrong without anyone asking it to.

The clean win: a vague summary made specific

Before

“Communication expert with six years of experience in non-profit program management, including evaluation, and program design.”

After

“Program evaluation and communications professional with six years of experience in nonprofit program management, qualitative research, and stakeholder engagement.”

Every real number in the original bullets — the 94% response rate, the $120K subaward, the 500 new members — survived untouched. This is the chatbot doing exactly what a resume rewrite should: no invention, just a sharper sentence.

The fabrication: a number nobody wrote

Before

“Communicated with Senate staff and Capitol staff to arrange meetings for incoming international delegates.”

After

“Coordinate weekly meetings between foreign dignitaries and Senate staff, supporting diplomatic engagement for 15+ international delegations.”

“15+” appears nowhere in the source resume. The same output, a few lines down, left an honest placeholder — “managed billing for [X] accounts” — for a number it didn't have. The inconsistency is the real problem: a reader can't tell which figures are real without checking every single one.

The overreach: an email address nobody asked to change

Before

mo_singh_thegoat@aol.com

After

mohamed.singh@gmail.com

The advice underneath this was correct — that address does read unprofessionally on a resume. The execution wasn't: the chatbot didn't flag it and let the person choose a replacement, it invented one and put it straight into the contact header as fact.

The tell that survived, even though the buzzwords didn’t

Across the five originals, em-dashes appeared 0 to 2 times — ordinary punctuation. Every single rewrite came back with 8 to 15 of them. The classic slop vocabulary (“synergy,” “leverage”) barely showed up — a modern model is past that tell — but the punctuation habit stamped every output the same way. That gap is exactly what our own vocabulary-based checks were missing before this test, and it’s why we added a dedicated density check on top of the word list.

What a chat window structurally can’t do

None of this is about one chatbot being worse than another. A chat interface is missing five things by design, regardless of which model or how good the prompt is:

It sees the file you're actually sending

A chat window only ever sees text you paste in. Paste-and-reformat destroys the layout, and nothing in a chatbot can tell you whether your PDF is text a parser can read or a picture of text it can't.

see what a parser reads from your file →

The score doesn't move unless the resume does

Ask a chatbot to score your resume twice and you can get two different numbers — large language models are measurably unstable graders of their own writing. The same file run through a fixed rubric scores the same every time, which is the only kind of number worth trusting.

how the deterministic score is computed →

The honesty rules are enforced, not requested

“Don't invent anything” in a prompt is a request a model can quietly ignore — which is exactly what put a fake number and a fake email on Mohamed Singh's resume above. A rule enforced in code either lets a rewrite through or it doesn't; it can't be talked out of it.

the honesty contract, and what enforces it →

The critique doesn't flatter you

Ask a chatbot to review your resume and the well-documented pull is toward agreement — it's the same tendency that made a major model roll back an update for being too flattering. A fixed rubric has no incentive to make you feel good about a document that isn't working yet.

the rubric, published in full →

It knows what a resume-shaped chatbot habit looks like

The tells above — heavy em-dash use, every bullet in a role landing on the same shape — are exactly what a general chatbot's default output carries, because it's never had to grade against them. Ours is built to catch its own category's fingerprints.

the guardrail prompts, if you'd rather DIY it →

The pitch isn’t “AI resumes fail” — it’s “guarded beats unguarded”

A field experiment covering roughly ~480,000 job seekers (van Inwegen, Munyikwa & Horton, Management Science, 2025) found that algorithmic writing assistance raised hire rates by about 8% and wages by about 8.4%, with no drop in how satisfied employers were afterward. The detail that matters for this page: the assistance in that study was structured and constrained — fix the writing, never invent the content — which is a much closer relative of a rubric- enforced checker than of an open chat window with no rules attached. The evidence doesn’t say AI hurts your resume. It says the AI that helps is the kind with guardrails, and a chat window is the one place those guardrails are optional.

Related reading

Want the guardrails without a checker?

Six copy-paste prompts that put the honesty rules directly in the chat window — the same rules our own paid rewrite enforces in code.

The free prompt library →

Worried a recruiter can tell you used AI?

They can’t detect the tool — no major applicant tracking system runs AI detection. What gets noticed is generic writing, with or without a chatbot involved.

The detection myth, with sources →

Already used a chatbot on yours? Find out which flip you got.

Paste in whatever you ended up with — chatbot-drafted, hand-written, or somewhere in between. SkimProof scores it the same way either way: same rubric, same checks, no credit toward or against you for how it was written.

Drop your resume here (PDF, DOCX, TXT, or RTF — max 5 MB)

or click to browse

Your resume is private. We parse it, score it, and discard the file — delete everything with one click.

ChatGPT resume rewrites — common questions

Can I just use ChatGPT to fix my resume?

Yes, and it can genuinely help — in our own test it raised two weak resumes' scores by double digits. Watch for three things it did without asking: inventing a metric that wasn't in your original text, quietly dropping a keyword or skill your resume already had, and reformatting dates or contact details based on its own assumption. None of those are hypothetical; all three happened in our test run.

Does ChatGPT make resumes worse?

Not as a rule — our own numbers say the opposite for weak resumes. What it did to the two already-strong resumes in our test was trim keyword and skills evidence that was already working, which lowered the score on paper even though every sentence in the rewrite was true. The risk isn't that AI writes badly; it's that nothing tells you which direction you got.

Is this based on a real study?

It's our own demo, not a study, and we'd rather say that plainly than borrow the credibility of research we didn't do. Five resumes, one prompt, one run, one model standing in for a general chatbot, done in August 2026. Run it again with a different resume or a different day and the specific numbers will move — that variance is the actual finding, not a caveat on it.

Which chatbot did you test?

We ran a frontier language model with the same naive prompt a person actually types into a chat window — “improve my resume,” no job description, no metrics supplied, single try — as a stand-in for a leading AI chatbot in its default configuration, rather than a literal session inside one specific chat product. We're deliberately not naming a single company's tool as the one true test subject, because the finding isn't about which product you use — it's about what happens when nothing checks the output, and that risk is the same wherever you paste your resume.

So should I never use a chatbot on my resume?

That's not what the evidence says, either. A large field experiment covering roughly ~480,000 job seekers found structured, guarded writing assistance raised hire rates — the treatment there looked more like an editor with rules than an open-ended “make this better.” Guarded beats unguarded; unguarded doesn't reliably beat doing nothing.

Menu