You posted one job. Two hundred applications arrived. They’re all well-written, all suspiciously well-matched to your wording, and after the fortieth one you can no longer tell who’s real.
Every article you’ll find about this promises to teach you the tells. Em dashes. “Spearheaded.” “Proven track record.” Spot the robot, bin the robot, hire the human.
Please don’t do that. The research on AI-text detection is unusually clear, and it says the tells-based approach will make you reject the wrong people — specifically, immigrants.
There’s a better way to run this, and it’s older than ChatGPT.
The problem is real, and it’s bigger than your inbox
Start with the honest scale of it. Robert Half surveyed more than 2,000 US hiring managers and found 67% saying that reviewing AI-generated applications had slowed their hiring, with one in five reporting delays of over two weeks. 84% of HR teams reported heavier workloads. 65% said AI-polished résumés made a candidate’s actual skills harder to verify.
Two caveats I’d want if I were you. That fieldwork was done in November 2025 and published in March 2026 — it’s been recycled through the trade press ever since, so if you saw it this week, it isn’t new. And Robert Half is a staffing firm; the same release reports that 89% of respondents find staffing firms effective at solving this. Read that as you like.
The underlying pressure is not in dispute, though. Gartner projects that by 2028, one in four candidate profiles worldwide could be fake.
Why “spot the AI resume” is bad advice
Here’s what the evidence actually shows.
Stanford researchers ran seven widely-used AI-text detectors against 91 TOEFL essays written by non-native English speakers, alongside essays by US students. The detectors got the native writers right more than 90% of the time.
They misclassified over half the non-native essays as AI-generated. One detector flagged 97% of them. Nineteen of the essays were unanimously flagged by all seven tools.
The mechanism is unglamorous and important. These tools measure how predictable the word choices are. Someone writing in their second language reaches for simpler, more common vocabulary — which looks statistically identical to a machine. When the researchers had ChatGPT rewrite those same essays with fancier vocabulary, the detectors reclassified them as human.
Read that again. Making the text more AI-touched made the detectors call it more human.
A separate review of 14 detection tools concluded they are “neither accurate nor reliable.” So when a survey reports that eight in ten hiring managers say they can spot an AI résumé, understand what that number is: self-reported confidence, not measured accuracy. Nobody tested them.
And roughly half of hiring managers say they’d automatically dismiss a résumé they suspected was AI-written. Put the two findings together and you get the actual risk: you are not screening out fakers. You are screening out people who learned English as adults.
There’s a fairness problem there, and depending on where you operate, a legal one.
While you were reading prose style, the real fraud walked in
The candidates who are genuinely fake don’t have a prose problem.
KnowBe4 — a security awareness training company, of all places — hired a North Korean operative in 2024. He passed four video interviews, a background check, and identity verification, using a stolen US identity with an AI-doctored stock photo. Within 25 minutes of receiving his work laptop he started loading malware.
That’s not an outlier. The Department of Justice has documented schemes using stolen identities of over 80 Americans to place operatives at more than 100 US firms; a single “laptop farm” in Arizona placed workers at 309 companies, including a Fortune 500 bank and Nike. Palo Alto’s Unit 42 showed that a researcher with no prior experience could build a working real-time deepfake interview identity in about 70 minutes with consumer hardware.
None of these people were caught because their résumé used an em dash. They were caught by behaviour, by verification, and in KnowBe4’s case, by their own security monitoring after the hire.
And your AI screener has a thumb on the scale
If you’ve been thinking “I’ll just have ChatGPT sort the pile” — one more finding worth knowing.
Researchers ran a controlled résumé experiment across 24 occupations, holding quality constant. Large language models systematically preferred résumés written by themselves. The bias against human-written résumés ran between 67% and 82% depending on the model. Candidates whose résumé was written by the same model doing the screening were 23% to 60% more likely to be shortlisted than equally qualified applicants who wrote their own.
So AI screening doesn’t filter out AI writing. It rewards it — and quietly penalises the candidate who sat down and wrote the thing themselves. Which is presumably the person you were trying to find.
What to do instead
Stop assessing the document. Assess the person against the work. This is well-trodden ground in hiring research: what predicts job performance is a work sample and a structured interview, and what barely predicts anything is reference checks and years of experience. (The classic validity numbers from Schmidt & Hunter’s 1998 meta-analysis were revised downward by Sackett and colleagues in 2022 — the rank order held, the magnitudes shrank. Don’t quote the coefficients at people. Do trust the ordering.)
Three steps, and a solo owner can run all of them.
1. Give a small, real task. Two hours, paid, resembling the actual job. A sample invoice to reconcile. A messy spreadsheet to clean. A customer complaint to answer. It doesn’t matter whether they used AI to help — it matters whether the output is good and whether they can defend it. This is the single highest-signal thing you can do, and it costs you a few hundred dollars per finalist.
2. Run the same interview for everyone, and interrogate specifics. Write your questions before you meet anybody. Then, on every claim in the résumé, go one layer down: walk me through the month you increased efficiency by 40% — what was the number before, who else was involved, what nearly went wrong? Someone who lived it has texture. Someone coached by AI has a summary. Ask “what would you do differently” — invented experience has no regrets.
Also just ask, openly and without judgement: did you use AI on your application, and how? In 2026 the honest answer is usually yes. The answer you’re listening for is whether they used it as a tool or as a ghostwriter, and whether they can tell you the difference.
3. Verify identity before hardware or systems access. Not at the offer stage — before the laptop goes out. Live video, ID matched to the face on the call, address confirmed against payroll. If they refuse a camera, that’s your answer. This is the step that catches actual fraud, and it’s the step almost nobody does.
What this means for you
If you hire two or three people a year with no recruiter: you probably can’t read 200 applications. Don’t try. Screen on the work sample first, résumé second. Invert the funnel.
If you’re a hiring manager inside a bigger company: the leverage is in fixing the job description. A third of HR leaders in the Robert Half survey rewrote theirs to discourage generic AI answers — specific, unusual, hard-to-fake requirements are cheaper than any detection tool.
If you’re tempted to buy an AI-detection tool: don’t. The peer-reviewed evidence says they’re unreliable, and their errors land hardest on non-native speakers. You would be paying for bias.
If you’re a job seeker reading this: use AI, and then rewrite it in your own voice. Not because a detector will catch you — it won’t, reliably — but because the interview will. You’ll be asked to defend every claim on that page.
If you hire remote workers internationally: the identity-verification step isn’t paranoia, it’s the documented attack path. Do it before access, every time, including for the candidate you like.
What this can’t fix
- It can’t give you back the time. A work sample costs you money and attention. It’s cheaper than a bad hire, but it isn’t free.
- It can’t tell you who wrote the résumé. Nothing can, reliably. Stop trying.
- It can’t scale to 200 candidates. You still need a first cut — make it a screening question with a specific, checkable answer, not a vibe check on prose.
- It can’t catch a competent fraud on its own. KnowBe4 ran four interviews and a background check. Layer identity verification on top; don’t rely on interviews alone.
- It won’t stop applicants using AI. That ship sailed. Roughly a third of job seekers already do, and the number only goes one way.
The bottom line
The flood is real. The advice about spotting it is worse than useless — it fails at the thing it promises and succeeds at rejecting immigrants.
So change what you’re grading. The résumé was never evidence; it was always just a claim. Ask for two hours of the actual work, ask everyone the same questions, and check that the person on the video call is the person you’re about to pay.
That’s it. It worked before AI, and it’s the only thing that still works after.
Want to understand what candidates are actually doing on the other side? Our resume writing course shows the tools they’re using, which is the fastest way to calibrate what you’re reading. If you’re building an AI-assisted hiring process, the recruiter’s AI audit course covers where automated screening goes wrong.
Related reading: AI Hiring Bias 101 and Behavioral Interview Questions, Redesigned for AI.
Sources
- Robert Half survey: 67% of HR leaders report AI-generated applications are slowing hiring (n=2,000+ US hiring managers, fielded Nov 2025)
- GPT detectors are biased against non-native English writers — Liang et al., Patterns (Cell Press), 2023
- AI-Detectors Biased Against Non-Native English Writers — Stanford HAI
- Testing of detection tools for AI-generated text — International Journal for Educational Integrity
- AI Self-preferencing in Algorithmic Hiring — arXiv:2509.00462
- How a North Korean operative got hired at KnowBe4 — Axios
- Deepfake candidates in the hiring pipeline — Palo Alto Networks Unit 42
