AI-written CV detectors: false positives and hiring decisions
An AI-writing warning is not evidence of dishonesty, and false alarms can create more review work than useful signal.
· 9 min read

AI-written CV detectors can flag text as likely AI-generated, but that warning does not tell you whether an applicant lied or whether the CV is inaccurate.
Do not reject a CV on an AI-writing warning alone. Check whether AI assistance was prohibited, what the detector percentage actually represents and what evidence would verify the candidate’s experience. A false alarm can create extra review work without improving the hiring decision.
Separate writing help from false claims and rule breaches
Who wrote the words, whether the experience is true and whether assistance was allowed are three different questions. Each question needs its own evidence.
| Application | The relevant hiring question |
|---|---|
| AI edited a truthful account, and editing was allowed. | Does the experience meet the job requirements? The permitted assistance is not a separate fault. |
| A person wrote a false qualification without AI. | Does the appropriate verification establish the qualification? A human-authorship result would not make it true. |
| An answer was generated during a task that clearly prohibited it. | What evidence establishes the rule breach, after considering the instructions and agreed support? |
| The writing process is unknown. | What can be assessed from the work, and is there a specific unresolved concern beyond the writing style? |
Before adding a detector to your workflow, finish this sentence: “A reliable finding of AI assistance would change our decision because…” If there is no job-related or policy reason, skip the authorship investigation and assess the application.
For an events vacancy, ask whether the applicant booked speakers or managed the budget. That distinction can change whom you interview. The writing style cannot establish either responsibility.
What does an “80% AI-written” score mean?
Ask what the percentage counts. “80%” might describe the share of eligible text flagged by a tool. It need not mean an 80% probability that the applicant used AI.
Next, check what qualifies as eligible text. A tool may accept a PDF without being suitable for the bullet points and tables inside it.
Turnitin provides a useful counterexample to the idea that “accepts PDFs” means “works on CVs”. Its AI Writing Report guide requires at least 300 words of long-form prose and says bullet points and tables are not reliably assessed. A bullet-point CV can therefore fall outside the intended writing format.
Its percentage describes qualifying text flagged, not the probability that an applicant used AI. 0% does not verify human authorship; an asterisk replaces results above zero and below 20%; a processing failure provides no authorship finding. Turnitin also warns against using the report alone for an adverse decision.
This is an education product’s documented limit, not proof that every detector fails on every CV. Ask the supplier of the actual tool you are considering for its own tested formats and output definitions.
Ask any provider to show you each state:
What does the percentage measure? What happens to a bullet-point CV, and what will I see when the tool cannot assess it?
Record “could not assess” separately. It means there is no usable authorship result, not that the CV is either suspicious or confirmed human-written.
Was the detector tested on CVs like yours?
Ask for the tested product version, languages, document types and warning threshold. A study of fully generated essays does not establish performance on CV bullets edited for clarity.
Ask how the evaluator established each document’s writing history. Labelling text “human” because another detector said so would make the test circular. Use test writing and edits your team can document; do not ask applicants for private chat histories to compensate for weak evidence.
Check what happened to documents the tool could not assess. An accuracy figure that leaves out short CVs, unsupported languages or failed files tells you nothing about those applications.
What essay-detection studies do and do not show
In a 2023 study by Weixin Liang and colleagues, seven detectors assessed 91 human-written TOEFL essays. About 61% were wrongly flagged on average. After ChatGPT revised the vocabulary, about 12% were flagged. The lower figure describes the AI-assisted versions, not a newly measured false-positive rate on untouched human writing.
By contrast, the abstract of Yang Jiang and colleagues' 2024 GRE-writing study reports near-perfect accuracy and no evidence of disadvantage to non-native English speakers for its purpose-built detectors and test material.
These are different detectors and writing tasks, not two estimates of the same product's accuracy. Neither establishes how a current CV detector will perform in your hiring process.
What if AI edited only part of the CV?
A CV may combine the applicant's own bullet points with an AI-edited summary. Ask how the supplier labels that mixed document. A label covering one edited sentence tells you something different from a label covering an entirely generated application.
Try three versions with the same underlying facts: original writing, a light AI edit and a substantially AI-generated version. Agree how the supplier defines each category before looking at the results. Otherwise, “correctly detected” could mean different things to you and the supplier.
Keep the authorship test separate from the hiring rule. Your policy may permit some or all of that assistance. An out-of-scope or failed assessment belongs in neither the human-written nor the AI-written group.
Use CV formats and languages your team receives. A few examples can reveal mistakes, but will not tell you how often the tool is wrong across all applicants. After the supplier adjusts the tool, try new examples rather than only repeating the ones it has just seen.
Why a 5% false-positive rate can still produce many wrong warnings
A false positive is human-written text wrongly flagged as AI-written. A 5% false-positive rate does not mean only 5% of the warnings your team sees will be wrong. Those percentages count different groups.
Take a worked set of 10,000 CVs: 1,000 AI-written and 9,000 human-written. Assume the tool can assess all of them, correctly flags 80% of the AI-written group and wrongly flags 5% of the human-written group. These are calculation inputs, not a named detector's results:
| Group | CVs checked | Warnings |
|---|---|---|
| AI-written | 1,000 | 800 correctly flagged |
| Human-written | 9,000 | 450 wrongly flagged |
| Total | 10,000 | 1,250 flags |
Of the 1,250 warnings, 450 are wrong: 36%.
5% counts errors among human-written CVs; 36% counts errors among warnings. For a recruiter opening the warning queue, more than one in three flags would be wrong under these assumptions. Even the correct flags still need to be checked against what assistance was allowed.
The detector also misses 200 AI-written CVs. An unflagged CV is therefore not proof of human authorship either.
These calculations assume the writing history is known. In a live queue, a denial, a withdrawal or failure to reply does not establish whether the detector was right. Keep unresolved cases separate from confirmed detector errors and confirmed use of assistance.
How much review time would those warnings add?
At an assumed three minutes per warning, checking all 1,250 would take 62 hours 30 minutes. That includes 22 hours 30 minutes on the 450 false alarms. Use your actual review time when estimating the workload; an investigation may take longer.
What changes when fewer CVs are AI-written?
Keep the same detection rates, but change the mix: 100 of the 10,000 CVs are AI-written and 9,900 are human-written.
The tool correctly flags 80 AI-written CVs. It also wrongly flags 495 human-written CVs. That gives you 575 warnings, of which 495 are wrong: about 86%.
The detector's success and error rates have not changed. AI-written CVs are simply rarer in this group, so false alarms make up a larger share of the warnings.
Recalculate the warning queue with different assumptions
Multiply the assumed number of AI-written CVs by the detection rate for correct warnings. Multiply the assumed number of human-written CVs by the false-positive rate for incorrect warnings. Add the two counts, then divide incorrect warnings by all warnings to find their share. If there are no warnings, that share is not applicable.
This calculation estimates warnings, not dishonest applicants, and assumes every CV can be assessed. Record mixed-authorship material, unreadable files and failed checks separately rather than forcing them into either group.
The workload is worthwhile only if investigating authorship helps a decision you need to make. Permitted editing remains permitted even when detected correctly. A qualification check or one question about project responsibility may resolve the real concern more directly.
What should your application policy say about AI help?
Your policy should distinguish clearer writing from invented experience. Editing “I arranged the speakers” is different from changing it to “I controlled the event budget”. The concern in the second sentence is the added responsibility, not its polish.
Emma Wiles, Zanele Munyikwa and John Horton studied nongenerative writing assistance in an online labour market. Their 2025 Management Science paper reports an 8% relative increase in hiring among jobseekers given that assistance. The study concerns that writing aid and labour market. It does not show that fully generated CVs have the same effect or verify any applicant's experience.
State the permitted help beside each task. Where your application policy allows editing, a plain version could read:
You may use writing tools to improve the wording of your CV. Keep the experience and qualifications accurate, and distinguish your own contribution from work colleagues did. We will explain any different rules before an assessment or interview.
Approve that wording against your actual process before using it. Application editing, interview preparation and live answer generation need separate rules; “be authentic” leaves candidates and reviewers guessing.
In an AI-assisted exercise, assess what the applicant checks before accepting the output. To assess unaided work, state that condition beforehand and choose a task you can observe. A “no AI” instruction on a take-home task does not prove how it was completed.
Apply the rule candidates received, not a stricter one invented after an answer sounds unexpectedly polished.
What should you check instead of relying on the detector?
Return to the events applicant. They explain that they booked speakers, while their manager controlled the budget. For an application where editing help was allowed, the note could say:
Booked speakers and arranged their attendance. Their manager controlled the budget. Ask about a last-minute change and how they handled it. We have not yet assessed budget-management experience.
The manager can now explore a specific responsibility. Any remaining concern about a false claim or prohibited assistance should be recorded and checked separately. “Possibly AI-written” does not explain either issue.
In Screened, Resume → Quality Signals → Authenticity can identify concerns such as instructions in a CV aimed at influencing the assessment. Our authenticity guide explains how to inspect the warning beside the original CV. This is not an AI-authorship percentage or independent verification of identity and qualifications. Contact support when the explanation doesn't match the document.
Inspect the text behind that warning. It can establish that an instruction is present in the document, not who inserted it or whether the applicant's experience is false.
Write what you learned in the hiring notes: what the applicant did, which qualification is still being checked or why the permitted writing help was not a breach. Correct any brief or decision that relied on a mistaken warning.
The useful outcome is an answer to the hiring question. “The detector says 80%” is a description of the tool's output, not that answer.


