Can employers tell if a CV was written by AI?
Not reliably with software. Published tests found AI-text detectors miss edited AI text and flag human writing, especially by non-native English writers. What a recruiter does catch is a claim you cannot explain, so invented details are the real risk.1
Written 13 October 2026. CitedCV. How we write our guides
Before you send an AI-drafted CV
Spend ten minutes on the draft, not on a detector. Whether a tool scores your CV as 20% or 80% AI says little about your chances; whether you can talk about every line says a lot. The audit below is the one check that matters at interview.
- Print the CV or open it next to a blank note.
- For every number, title, tool and certificate, write where it comes from: a payslip, a contract, a report, a certificate, a project you can describe.
- Strike out anything with no source. A model fills gaps with plausible detail; this is where it shows.
- Read each remaining line aloud and ask: could I answer "tell me more about that" for two minutes?
- Replace stock openers such as 'results-driven professional' with one concrete line, or drop them.
- Follow the employer's rule on AI use if the posting states one.
How AI-text detectors make their call
Most detectors score how predictable the wording is. A language model tends to choose likely words, so text with low 'perplexity' (few surprising word choices) looks machine-made to them. Liang and colleagues at Stanford showed the weakness of that idea: careful but plain human writing is also predictable.
What detector makers say about their own tools
OpenAI launched an AI-text classifier on 31 January 2023. On its own English 'challenge set', the page said, it labelled 26% of AI-written text as likely AI-written and wrongly labelled human text as AI-written 9% of the time. It called the classifier very unreliable below 1,000 characters, recommended it only for English, and said very predictable text cannot be reliably identified. On 20 July 2023 OpenAI withdrew it, citing its low rate of accuracy.
Turnitin, whose detector schools use, reported in May and June 2023 a document-level false positive rate under 1% for documents with 20% or more AI writing, and about 4% at sentence level. It raised its minimum from 150 to 300 words, said its goal is long-form writing, and added an asterisk to scores under 20% because false positives are more common there. It says it cannot remove the risk of false positives completely and that highlighted sentences should start a conversation, not settle one.
Vendor figures are the vendors' own measurements on their own data, read on 13 October 2026. They are reported here as statements, not as independent results.
What published tests of detectors found
These are primary studies, read in full on 13 October 2026. None tested CVs; they tested essays, abstracts and short texts, which is the closest evidence there is.
| Study | What was tested | Finding | What it means for a CV |
|---|---|---|---|
| Liang, Yuksekgonul, Mao, Wu and Zou (Stanford), arXiv 2304.02819, 2023 | Seven detectors on 91 TOEFL essays by non-native writers and 88 US 8th-grade essays | Near-perfect on the US essays; 61.22% average false positive rate on the TOEFL essays; 89 of 91 flagged by at least one detector | Plain, careful English from a second-language writer is the profile most often wrongly flagged |
| Same study, second experiment | ChatGPT essays rewritten with a prompt to use literary language | Detection fell from 100% to 13% | One extra rewrite prompt defeats a predictability test |
| Weber-Wulff and seven co-authors, International Journal for Educational Integrity, 2023 | 12 public tools plus Turnitin and PlagiarismCheck on human, machine-translated, AI, AI-edited and AI-paraphrased texts | All below 80% accuracy, only 5 above 70%; human texts judged correctly 96% of the time, about 20% less often after machine translation | A CV written in your language and run through a translator is at higher risk of a false flag |
| Perkins and six co-authors, International Journal of Educational Technology in Higher Education, 2024 | Popular detectors on 805 tests of AI texts with and without evasion techniques, plus 10 human-written controls | 39.5% average accuracy on unaltered AI text; only 67% of tests on human controls were correct; accuracy fell further with simple evasion | Both errors are common: AI text passes and human text is flagged |
| Dugan and seven co-authors (RAID benchmark), ACL 2024 | 12 detectors on over 6 million generations, 11 models, 11 attacks | Detectors claiming 99% accuracy are rarely tested on shared data; most are easily fooled by attacks; few work at a false positive rate under 1% | A vendor's headline accuracy is not evidence about your document |
| Turnitin (vendor research), October 2023 | Up to 2,000 texts per group, English learners and native writers, above and below 300 words | No significant gap at 300+ words; false positives significantly above its 1% target under 300 words | Short text is where errors rise, and a CV summary is short |
Why a CV is an awkward input for a detector
This is reasoning from the sources above, not a measured result. A CV is mostly bullets, dates and names; its prose is a short summary and a few lines per job, often under the 1,000 characters OpenAI called very unreliable and under the 300 words Turnitin requires. CV language is formulaic by design ('Managed', 'Built', 'Reduced'), which makes it predictable to a perplexity test whether a person or a model wrote it.
Second-language writers and translated CVs
Liang's team then asked ChatGPT to enrich the TOEFL essays' vocabulary; the false positive rate fell from 61.22% to 11.77%. The bias was in the plainness of the wording, not in who held the pen. Turnitin disputes how far this applies to its tool, noting the TOEFL essays were all under 150 words, below its minimum. Both points lead to the same caution: a detector verdict on a short text by a non-native writer is weak evidence.
What a recruiter can actually see
The applicant tracking systems most employers use store, parse and search CVs; their documentation describes no AI-text detection, as the page on ATS-friendly CVs explains. A detector only enters the picture if someone pastes your text into one. What reaches a person is your wording and your claims.
Wording that sounds like everyone else
Prospects, the UK graduate careers service, says recruiters are getting better at spotting AI-written CVs and that you should never pass off a completely AI-generated CV as your own. What gives a draft away is sameness: the stock phrases every model produces, identical structure in every bullet, and polish with no specific detail.
When polished writing stops counting
Galdin and Silbert (arXiv 2511.08785, 2025) studied applications on Freelancer.com. Before large language models, employers paid a premium for applications tailored to the job post; after them, that premium disappeared. Once everyone's text reads well, employers look elsewhere: work history, proof, the interview. The page on AI cover letters covers a related study of a cover letter tool on another platform.
Help with writing is not the problem
Van Inwegen, Munyikwa and Horton (arXiv 2301.08083, 2023) ran a field experiment with nearly half a million jobseekers in an online labour market. Those offered algorithmic writing help on their resumes, mainly spelling and grammar suggestions, were 8% more likely to be hired, with no sign that employers were less satisfied. Clearer writing helped employers judge ability; it did not hide it.
When the screener is itself a model
Xu, Li and Jiang (arXiv 2509.00462, 2025) found large language models used as evaluators preferred resumes they had generated over human-written ones in 67% to 82% of comparisons. In their simulation across 24 occupations, that bias raised shortlisting odds by 23% to 60%. The authors treat it as a flaw to fix and cut it by more than half with simple interventions, so it is no reason to hand your CV to a model.
The real risk: lines you cannot defend
Prospects puts it plainly: the facts on a CV are easy to corroborate, and an invented detail is either obvious or exposed at interview. The page on why AI CV builders hallucinate explains where invented metrics come from. Below is what that looks like in a conversation. The lines are constructed for this page; no real candidate is quoted.
| Line an AI draft produced | The question it invites | A version you can defend (example) |
|---|---|---|
| Spearheaded a cross-functional initiative that lifted team productivity by 35% | "How did you measure 35%?" | Set up a shared ticket board for three support agents; first replies went from two days to one (team report, Q3 2025) |
| Expert in Python, SQL, Tableau and Power BI | "Walk me through a Power BI report you built." | SQL for weekly sales extracts; two Tableau dashboards used by the regional team |
| Led a team of five engineers through a cloud migration | "Who reported to you?" | One of five engineers on the AWS migration; wrote the scripts that moved the billing database |
| Fluent German | "Shall we continue in German?" | German B1 (Goethe-Zertifikat, 2024) |
| Results-driven professional with a proven track record | None; it tells the reader nothing | Delete it, or open with your strongest true line |
Hidden keywords or instructions to an AI screener are a separate risk and a detectable one; the page on hidden text in resumes covers it. Answers to screening questions are checked too, which the page on knock-out questions explains.
Should you say you used AI?
Follow the employer. Prospects notes that some employers say 'no AI' while others encourage ethical use, and lists sensible uses: research, summarising job descriptions, a first draft, proofreading. It says AI should not write the whole CV or invent and embellish details. If the application form asks about AI use, answer it truthfully; if it does not, the CV itself carries no label either way.
Limits of this evidence
- No primary study we found measures how well recruiters or detectors identify AI-written CVs. The evidence above is from essays, abstracts and short texts.
- Detectors change. Every study dates its tests; the OpenAI and Turnitin figures are from 2023.
- This page reports no accuracy figure for any detector sold today, because we found none tested independently on CVs.
- If an employer accuses you of AI use on the strength of a detector score, the studies above are a fair basis for asking how the score was produced.
Ways to use AI that leave you in control
| Approach | What it gives you | Watch out for |
|---|---|---|
| Write it yourself, use AI to proofread | Your own wording, fewer errors | Accepting rewrites that add detail you did not write |
| AI first draft, then the line audit above | Speed, with every claim checked | Skipping the audit when the draft reads well |
| Tailor by hand to each posting | The posting's true terms in visible text | Time per application; see the page on tailoring |
| A builder that writes only from your facts | Lines tied to what you supplied | It cannot add experience you do not have |
CitedCV is one example of the last row: it writes a CV only from the facts you give it, each line carries the ids of the facts behind it, and a deterministic verifier drops lines whose numbers or names it cannot match. It does not detect or disguise AI text. The free ATS CV checker flags stock phrases such as 'results-driven' because readers skip them, not to predict a detector score.
Questions people ask
Can employers tell if your CV is AI generated?
Not reliably by software: published tests found detectors miss edited AI text and wrongly flag human writing. People notice generic wording and claims you cannot explain, which is where AI drafts usually fail.
Do recruiters run CVs through AI detectors?
Some may paste text into one, but the applicant tracking systems most employers use document no AI-text detection. A detector verdict on a short CV is weak evidence, as the studies above show.
How accurate are AI detectors?
Less accurate than their marketing suggests. Weber-Wulff and colleagues found all 14 tools below 80% accuracy, and the RAID benchmark found most detectors easily fooled; OpenAI withdrew its own classifier for low accuracy.
Why was my own writing flagged as AI?
Probably because it is plain and predictable. Liang and colleagues found detectors flagged 61% of essays by non-native English writers on average; machine translation also raised false flags in another study.
Is it wrong to use ChatGPT to write a CV?
No, as long as every line is true and yours to explain. Prospects lists research, drafting and proofreading as sensible uses and warns against letting AI invent or embellish details.
Will rewording an AI CV make it undetectable?
Detectors are easy to fool, which is why they are weak evidence, but that misses the point. A reworded invented claim is still invented; the question that exposes it comes from a person, not a tool.
Should I disclose that I used AI on my CV?
Yes, if the employer asks. Some employers ban AI and others welcome ethical use, so follow the rule on the application; where none is stated, a truthful CV needs no label.
Do AI screening tools prefer AI-written resumes?
One 2025 study found language models preferred resumes they generated in 67% to 82% of comparisons. It was a simulation, the authors call it a bias to fix, and simple changes halved it.
Sources
- Liang, Yuksekgonul, Mao, Wu and Zou, GPT detectors are biased against non-native English writers (arXiv 2304.02819, 2023)arxiv.org/abs/2304.02819
- Weber-Wulff et al., Testing of detection tools for AI-generated text (arXiv 2306.15666; Int J Educ Integr 19, 26, 2023)arxiv.org/abs/2306.15666
- Perkins et al., GenAI detection tools, adversarial techniques and implications for inclusivity in higher education (arXiv 2403.19148, 2024)arxiv.org/abs/2403.19148
- Dugan et al., RAID: a shared benchmark for robust evaluation of machine-generated text detectors (arXiv 2405.07940, ACL 2024)arxiv.org/abs/2405.07940
- OpenAI, New AI classifier for indicating AI-written text (31 January 2023, withdrawn 20 July 2023)openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
- Turnitin, Understanding the false positive rate for sentences of our AI writing detection capability (14 June 2023)www.turnitin.co.uk/blog/understanding-the-false-positive-rate-for-sentences-of-our-ai-writing-detection-capability
- Turnitin, AI writing detection update from Turnitin's Chief Product Officer (23 May 2023)www.turnitin.co.uk/blog/ai-writing-detection-update-from-turnitins-chief-product-officer
- Turnitin, New research: Turnitin's AI detector shows no statistically significant bias against English language learners (26 October 2023)www.turnitin.ca/blog/new-research-turnitin-s-ai-detector-shows-no-statistically-significant-bias-against-english-language-learners
- Galdin and Silbert, Making talk cheap: generative AI and labor market signaling (arXiv 2511.08785, 2025)arxiv.org/abs/2511.08785
- van Inwegen, Munyikwa and Horton, Algorithmic writing assistance on jobseekers' resumes increases hires (arXiv 2301.08083, 2023)arxiv.org/abs/2301.08083
- Xu, Li and Jiang, AI self-preferencing in algorithmic hiring (arXiv 2509.00462, 2025)arxiv.org/abs/2509.00462
- Prospects, How to write a CVwww.prospects.ac.uk/careers-advice/cvs-and-cover-letters/how-to-write-a-cv