Everyone has an opinion on which AI writes the best resume. Most of those opinions come from typing one prompt, liking the look of the answer, and moving on.
Looking good is not the test that matters. A resume is a list of claims about you, and every one of them can come up in an interview. So we tested the thing nobody checks: when you hand ChatGPT or Gemini your real history, does the resume that comes back still describe you?
We wrote career notes for ten people across four countries, the way people actually type them into a chat box. Short, a bit messy, some numbers, a lot of things with no numbers at all. We gave each tool the same notes, the same job ad and the same one line request. Then we read all twenty resumes line by line and checked every figure against what the person had said.
The short version: both tools wrote clean, well structured resumes that covered the job ads equally well, and both scored between 85 and 95 on an ATS checker. The difference was honesty. ChatGPT stuck to what it was told. Gemini filled every silence with a confident number, and in four resumes it added qualifications the person never mentioned. Its drafts read stronger, and some of that strength was made up.
What we actually did
The method is simple on purpose, because the simple version is what you do at home.
We wrote notes for ten candidates: an ICU nurse in Leeds, a backend engineer in Austin, a marketing manager in Chicago, a night shift warehouse team leader in Northampton, a staff accountant in Toronto, a primary teacher in Bristol, a customer support rep in Phoenix, an electrician in Brisbane, an operations coordinator in Denver moving into data analysis, and a 20 year old student in Columbus applying for her first retail job. Each set of notes had the facts a real person would give: employers, dates, what they did, and a few numbers they actually knew. Plenty of achievements were described with no number at all, because that is how people talk.
Each candidate got a real looking job ad for a role one step up from where they were. Then both tools got exactly the same message:
Write me a resume for this job. Here is my background: [the notes] Here is the job ad: [the ad]No system prompt, no instructions about honesty, no tricks, default settings. ChatGPT was OpenAI's gpt-5.5 and Gemini was Google's gemini-3.8-flash, both called through their official APIs on 25 September 2026. Each tool wrote each resume once.
Then we measured. Word count, bullet count, em dashes, bracket placeholders, the 14 stock phrases from our 500 resume study, and how many of the job ad's key terms made it in. We ran every resume through the FreeCV ATS checker three times against its job ad. And a script pulled out every number that did not appear in the notes or the ad, which we then reviewed by hand, one line at a time.
Finding one: Gemini invented 13 numbers. ChatGPT invented none.
This is the result that matters most, so here it is in full. Every number Gemini wrote that had no source in the candidate's notes or the job ad, quoted exactly, next to what the person actually said.
| Candidate | What they told it | What Gemini wrote |
|---|---|---|
| Customer support, US | “My CSAT is usually one of the highest on the team” | “consistently achieving one of the highest CSAT ratings on the team (consistently 95%+)” |
| Staff accountant, Canada | “I built an Excel template that made the rent reconciliation way faster” | “reducing monthly tenant reconciliation processing time by 40%” |
| Staff accountant, Canada | “I did bookkeeping and T2 corporate tax returns for small business clients” | “for an assigned portfolio of 25+ private corporate clients” |
| Data analyst, US | “I built a Power BI dashboard for on time delivery” | “performance dashboard tracking 1,000+ monthly shipments” |
| Data analyst, US | “a weekly Excel report with Power Query that used to take me half a day” | “reducing production time from 4+ hours to under 15 minutes weekly” |
| Data analyst, US | “Before that I was a teller at Wells Fargo” | “Balanced daily cash transactions with 100% compliance to strict audit, regulatory, and reconciliation standards” |
| Data analyst, US | “I did a portfolio project analyzing Denver bike share data” | “Cleaned and transformed 100k+ trip records” |
| Backend engineer, US | “it stopped the missed reminders problem” | “eliminating dropped notifications and achieving 99.99% message delivery reliability” |
| Marketing manager, US | “I did the social media, local SEO for 6 clinics, and patient newsletters” | “maintaining open rates above 32%” |
| Marketing manager, US | “local SEO for 6 clinics” | “to increase inbound patient leads by 24% year-over-year” |
| ICU nurse, UK | “staff nurse on a respiratory ward at St James Hospital Leeds” | “Delivered acute medical nursing care to 28+ bed respiratory patients” |
| ICU nurse, UK | Nothing about mandatory training | “NHS Mandatory Training (Adult Safeguarding Level 3, Infection Control, Information Governance) … 100% compliant” |
| Primary teacher, UK | “Year 4 class teacher at Greenway Primary School” | “differentiated to meet the diverse academic, social, and emotional needs of 30 pupils” |
Read down the right hand column and every line sounds like a strong resume. That is the danger. None of these are wild. A 95% CSAT is believable for a top support rep. A 40% time saving is believable for a good spreadsheet. A 28 bed respiratory ward is a normal size. The model is not writing fantasy, it is writing the most typical number for that kind of claim, which is exactly why you might not notice it.
And every one of these would come up. “How did you measure the 40%?” is a completely normal interview question. So is “tell me about the 99.99% delivery reliability” to a backend engineer, where the interviewer is almost certainly going to ask how it was monitored. The candidate would be defending a number they never claimed.
ChatGPT, given the same notes, did not do this once. When an achievement had no figure, it wrote the achievement without one. The only numbers it produced that were not copied directly were fair workings, like five plus years of team leading for someone who had been a team leader since 2019.
Finding two: Gemini also handed out qualifications
Numbers were not the only thing it filled in. In 4 of its 10 resumes, Gemini added credentials or details the candidate never mentioned.
- Degree classes. The nurse and the teacher both received a 2:1. Neither said what grade they got.
- School qualifications. The warehouse team leader was given “GCSEs including English & Mathematics”. His notes never mentioned school.
- An accrediting body. His forklift licences became “RTITB Licences”. He said he had counterbalance and reach truck licences, not who issued them.
- A licence class. The electrician's EWP licence became “Over 11m / Under 11m”, a specific class he never claimed.
- Referees. The nurse's references line read “Current Band 7 Sister and Critical Care Matron”. She never named a referee.
A made up degree class is worse than a made up percentage, because it can be checked in one phone call to a university. It is also the kind of detail you skim past when proofreading, because it looks like formatting rather than a claim. We searched the ChatGPT resumes for the same kinds of additions and found none.
Finding three: Gemini still writes in stock phrases
Back in June we counted the words ChatGPT could not stop writing on resumes: proficient, proven track record, results driven, leverage, spearheaded and the rest. We ran the same 14 word list over this batch.
The newer ChatGPT has clearly been cleaned up. Across ten resumes it used the list three times in total. Gemini used it in all ten, most often with “proven track record”, “utilize”, “spearheaded” and “dynamic”. If you use Gemini, run the buzzword swap list over the result before you send it.
This is not a like for like comparison with our June study, which used a different prompt and gave the model no real notes. But the direction is plain. The phrases recruiters roll their eyes at have mostly left ChatGPT's writing and are alive and well in Gemini's.
Finding four: ChatGPT's problem is length and dashes
ChatGPT did not win everything. Its resumes had their own tells, and they are the ones that make a recruiter think a machine wrote it.
Both tools wrote about 500 words per resume, but ChatGPT chopped them into far more bullets, many of them short restatements of the same duty. Nearly 29 bullets is a lot of scrolling for a recruiter who spends seconds on a first read. And the em dash, the long dash hardly anyone types by hand, was still in 70% of ChatGPT's resumes. In June it was in 92%, so it has improved, but it is still the fastest giveaway on the page.
The fix for both is quick. Cut each job to your three to five strongest bullets, and replace every em dash with a full stop or a comma.
Finding five: an ATS checker cannot tell a real number from a fake one
Both tools covered the job ads almost equally. ChatGPT got 89% of each ad's key terms into the resume on average and Gemini got 88%. Both are good, and both are well above what most people manage by hand.
The ATS scores were close as well. Every resume landed between 85 and 95.
| Candidate | ChatGPT score | Gemini score |
|---|---|---|
| ICU nurse, UK | 90 | 93.3 |
| Backend engineer, US | 85 | 85 |
| Marketing manager, US | 85 | 95 |
| Warehouse team leader, UK | 85 | 85 |
| Staff accountant, Canada | 90 | 86.7 |
| Primary teacher, UK | 85 | 93 |
| Customer support, US | 85 | 85 |
| Electrician, Australia | 93.3 | 95 |
| Data analyst, US | 85 | 90 |
| Retail associate, US | 85 | 85 |
Gemini averaged 89.3 against ChatGPT's 86.8. That looks like a win until you know two things. First, in our separate test of the checker itself, the same unchanged CV scored 85 half the time and 90 the other half, so a gap of two or three points means nothing. Second, look at where Gemini scored higher. It beat ChatGPT on five candidates, and four of those five resumes contained invented numbers.
That is not a coincidence. ATS checkers, ours included, reward bullets with numbers in them, because a quantified bullet is usually a better bullet. The checker has no way of knowing the number was made up. So the tool that invents more figures can score a little higher, and the score quietly rewards the one thing that will hurt you in the interview.
Finding six: both left gaps for you to fill
Bracket placeholders were rare, and both tools had them. ChatGPT left two in one resume: “[Add LinkedIn URL]” and “[Add portfolio URL]”. Gemini left seven across six resumes, including “[Insert NMC PIN]” on the nurse and invented LinkedIn addresses in brackets like “[linkedin.com/in/wei-zhang-placeholder]”. None of them were the old “increased sales by [X]%” kind. Search your draft for a square bracket before you send it, every time.
The full scoreboard
| Measure, 10 resumes each | ChatGPT | Gemini |
|---|---|---|
| Invented numbers | 0 | 13, in 7 resumes |
| Invented credentials | None found | In 4 resumes |
| Stock buzzwords | 3, in 2 resumes | 31, in all 10 |
| Em dashes | 28, in 7 resumes | 6, in 3 resumes |
| Bullets per resume | 28.8 | 19.1 |
| Words per resume | 501 | 512 |
| Bracket placeholders | 2, in 1 resume | 7, in 6 resumes |
| Job ad keyword coverage | 89% | 88% |
| ATS score, 3 runs each | 86.8 | 89.3 |
The raw numbers for every resume are in our CSV file, free to reuse with a link back.
So which one should you use?
If you want the draft you have to fix least, use ChatGPT. It stayed inside the facts it was given, it barely touched the stock phrases, and its problems are cosmetic: too many bullets and too many dashes, both fixable in five minutes.
If you prefer how Gemini writes, and plenty of people will, because its bullets are punchier and more confident, use it with your eyes open. Treat every number in its draft as a question rather than a fact. If you cannot say where the figure came from, delete it or replace it with your real one.
And whichever you pick, the single best thing you can do is give it more real numbers to start with. Both tools do best when the notes are rich. The candidates in our test who gave figures got those exact figures back, correctly, from both tools. The invention happened in the gaps.
3 prompts that keep either tool honest
These work in ChatGPT and in Gemini. Paste your notes and the job ad where shown.
1. The honest first draft
The last two rules do the heavy lifting. Without them, Gemini in particular will fill every quiet line with a number.
2. The source check
Run this on any draft, from any tool, including one you wrote yourself.
3. The final cleanup
This handles the cosmetic tells from both tools: the bullet sprawl, the dashes and the stock phrases.
What this test does not tell you
Ten candidates is enough to see a pattern and not enough to measure an exact rate. Gemini inventing numbers in 7 of 10 resumes is a strong signal, but on another day, with other notes, it could be five or nine.
We used the APIs, not the chat apps. The apps can add their own instructions, memory and settings, so your results in the app may differ. We tested one model from each company, OpenAI's gpt-5.5 and Google's gemini-3.8-flash, and other models from the same companies may behave differently.
Each tool wrote each resume once. We did not ask for revisions, and we did not tell either tool to be careful with facts. That was deliberate, because most people do not. With prompt 1 above, both tools behave better.
Claude is not part of this round. We ran what we could run properly on the day, with identical settings and a full record of every output.
Everything above, in seven lines.
- 1Same notes, same job ads, same prompt. ChatGPT invented 0 numbers. Gemini invented 13, in 7 of 10 resumes.
- 2Gemini also added degree grades, school results, a licence class and referees nobody mentioned, in 4 resumes.
- 3Gemini used stock phrases 31 times across all 10. ChatGPT used them 3 times.
- 4ChatGPT's tells are length and dashes: almost 29 bullets per resume and em dashes in 7 of 10.
- 5Both covered about 88% of each job ad's key terms and both scored 85 to 95 on an ATS checker.
- 6A checker rewards numbers and cannot tell a real one from a fake one. Check figures before scores.
- 7Give the tool your real numbers up front and tell it not to add any. The inventing happens in the gaps.
Frequently asked questions
Is ChatGPT or Gemini better for writing a resume?
In our test ChatGPT was the safer writer and Gemini was the more confident one. We gave both the same ten sets of career notes and job ads. ChatGPT invented no numbers at all. Gemini invented 13 numbers across 7 of its 10 resumes, and added credentials the candidates never mentioned in 4 of them. Gemini also used far more stock buzzwords, 31 against ChatGPT's 3. ChatGPT's weaknesses were length and punctuation: nearly 29 bullets per resume and 28 em dashes across 7 of its 10. If you only want one tool, ChatGPT needs less fixing. Whichever you use, check every number.
Does Gemini make up numbers on resumes?
It did in our test. When a candidate described something without a figure, Gemini often supplied one. A candidate who said her CSAT was usually one of the highest on the team got "consistently 95%+". An accountant who said his template made reconciliation way faster got "reducing processing time by 40%". A data analyst who built an on time delivery dashboard got "tracking 1,000+ monthly shipments". In total 13 numbers in 7 of 10 Gemini resumes had no source in what the person told it.
Does ChatGPT make up achievements on a resume?
Not in this round. We checked every number in all ten ChatGPT resumes against the candidate notes. The only numbers not copied directly were fair calculations, like "5+ years" worked out from the dates given. That is a change from what most people remember about older ChatGPT versions, which were known for inventing metrics and leaving [X]% placeholders. The model we tested was gpt-5.5.
Which AI writes resumes that pass ATS better?
Neither had a meaningful edge. Both covered about 88 to 89 percent of the key terms in each job ad, and both scored between 85 and 95 on the FreeCV ATS checker. Gemini averaged 89.3 and ChatGPT 86.8, but our separate test found the checker itself moves by 5 points on an unchanged CV, so that gap is inside the noise. It is also worth knowing that an ATS score cannot tell a real number from an invented one, and Gemini's higher scores came mostly on resumes that contained invented figures.
Why does AI add numbers I never gave it?
Resume advice everywhere says to quantify your achievements, and the models have read all of it. When a bullet has no figure, a model trained to write a strong resume will reach for one, because a bullet with a number looks like a better bullet. It has no way of knowing your real figure, so it writes a plausible one. The fix is to tell it plainly not to add any number you did not give it, and then to check anyway.
Is it OK to use an AI written resume?
Yes, as a first draft. Both tools produced clean, well organised resumes that covered the job ad well. The risk is sending the draft without reading it. An invented number or a degree class you never earned is not a style problem. It is a claim you will be asked about in an interview or a background check, and you will not be able to back it up.
Did Gemini invent qualifications too?
In 4 of 10 resumes, yes. It gave the nurse and the teacher a 2:1 degree class that neither mentioned, gave the warehouse team leader GCSEs in English and Maths and an RTITB accreditation, gave the electrician a specific EWP licence class, and described the nurse's referees as a Band 7 Sister and a Critical Care Matron. Every one of those is plausible, and that is exactly the problem.
How do I stop ChatGPT or Gemini from inventing things on my resume?
Give it every real number you have up front, and add one line to your prompt: do not add any number, percentage, credential or detail that is not in my notes. Then ask it to list every figure it used and where it came from. The three copy and paste prompts in this article do exactly that. Finally, read the result once, slowly, asking of every figure: could I prove this in an interview.
Why is Claude not in this test?
We ran the tools we could run properly on the test day, with identical settings and a clean record of every output. Claude was not part of this round. Two tools tested carefully, with the raw numbers published, is more useful to you than three tools tested loosely.
Can I see the raw data?
Yes. Every resume in the test has a row in our CSV: tool, model, word count, bullets, em dashes, placeholders, buzzwords, keyword coverage, invented numbers, invented credentials and all three ATS scores. We did not publish the resume texts themselves because they contain made up contact details that could collide with real people.