You can't prove a candidate used AI on a two-minute screening quiz, and you don't need to. The practical way to detect AI cheating in an online assessment is to log three behavioural signals, per-question timing, tab switching and pasted text, and treat each as a question for the interview, never a verdict. In Checkr's June 2026 survey, 73 per cent of US business leaders said they'd met AI-generated resumes, cover letters or assessments frequently or occasionally, and 71 per cent would invest heavily in detection tools if proven reliable. None is. Reading the signals fairly is what matters.
- 73 per cent of US business leaders have met AI-generated resumes, cover letters or assessments frequently or occasionally, and 20 per cent name AI-completed skills tests as their biggest fraud worry (Checkr, June 2026).
- Three signals are worth logging on a quiz: per-question timing, tab switching and pasted text. None of them proves anything on its own.
- A flag is a question for the interview, never an automatic rejection. GDPR Article 22 and plain fairness point the same way.
- The EU AI Act's high-risk duties for screening tools were pushed to 02-12-2027, but telling candidates what you capture applies now under GDPR.
How common is cheating on online screening quizzes?
Nobody has a clean number. Treat every figure here as direction, not a rate for your own funnel. The oldest baseline is a May 2024 survey by ResumeTemplates.com, a resume-template vendor, of 1,250 US job seekers: 22 per cent admitted cheating on an online assessment. Of those, 71 per cent googled answers, 37 per cent used ChatGPT when it wasn't allowed and 33 per cent got help from another person. Self-reported and pre-AI wave, so read it as a floor.
The employer side is louder. In Checkr's survey of 1,000 US business leaders (fieldwork April 2026), AI-completed skills tests were the single biggest AI-fraud worry for 20 per cent, second only to AI-generated resumes at 36 per cent. Every application is AI-written now, and screening has to adapt without becoming a witch hunt.
Which behavioural signals suggest outside help on a quiz?
A quiz produces more data than the answers. Three signals are worth logging, plus one pattern to watch across them.
- Per-question timing: a hard scenario answered in four seconds, or an easy one that took ninety, breaks the rhythm a real person shows.
- Tab switching: leaving the quiz page mid-question. In FabricHQ's analysis of 19,368 AI-conducted interviews on its own platform (July 2025 to January 2026), tab switching or a second screen was only 18 per cent of flagged cases; dedicated cheating assistants (45 per cent) and voice mode on chatbots (34 per cent) had taken over. Vendor data from interviews, not quizzes.
- Pasted text: a free-text answer arriving in one paste event. The most concrete of the three, and still not proof: some people draft in a notes app.
- Uniform delays: the same pause after every question regardless of difficulty. FabricHQ calls it a 'lag loop' of a consistent 3-5 second delay; it flagged 38.5 per cent of candidates overall and publishes no false-positive rate. One tell to watch, not a rule.

Which honest candidates trip these signals by accident?
More than you'd think. A candidate on a bus with patchy 4G shows odd timing on every question. Someone who takes a call mid-quiz shows a tab-away. A careful thinker who reads each question twice looks slow on the easy ones. A screen-reader user produces timing and focus patterns nothing like a mouse-and-keyboard user, and in the EU and UK treating those as suspicious invites a disability discrimination claim. Fixing the process for neurodivergent candidates starts with never letting an assistive-tech pattern count against anyone.
No vendor we know of publishes a false-positive rate for these signals. FabricHQ claims detection 'in 85 per cent of cases' and nothing about the honest people it flagged. A flag raises the odds of outside help, nothing more. Two flags on one candidate deserve a closer look. One deserves a note.
What should you do when a signal fires?
Ask a question. That's the only defensible rule. A flag isn't evidence of cheating; it's a prompt to verify at interview. Never let software reject on a flag alone. GDPR Article 22 gives people the right not to be subject to a decision based solely on automated processing, and a quiet auto-reject on a timing anomaly is exactly that.
Qwiza built its integrity signals on this rule: per-question timing, tab switching and pasted text are flagged so the recruiter can take a closer look, the score doesn't move, and nobody is auto-rejected. That means one follow-up per flag. For a timing anomaly, put the same scenario to them live and listen for the reasoning. For a tab-away, re-ask that question in different words. For pasted text, ask them to expand on what they wrote. Honest candidates find all three easy.
Illustrative only: the numbers come from Qwiza demo campaigns and pilot targets, and the names are invented. Marko runs a 40-seat cafe in Zagreb and is hiring two baristas for the autumn. His quiz ad brings in 84 completed quizzes in a week. Six show a signal: three answered the milk-steaming scenario in under five seconds, two left the page mid-quiz, one pasted a 90-word answer to the rush-hour question.
He shortlists on score, as normal, and four of the six flagged candidates make the top 20 anyway. At interview he puts the steaming scenario to them verbally. Three explain it fluently; one can't. The paster turns out to draft everything in her notes app and expands on her answer without pause. The two tab-aways both mention a call from a current employer. He hires one of the flagged four. Nobody was rejected on a flag; one was rejected because the live answer fell apart.
How do you write quiz questions that are hard to outsource?
Detection is the second line. The first is asking things a chatbot can't answer well in the time allowed. Generic knowledge questions ('What is FIFO?') are googleable in five seconds. Questions rooted in your own operation aren't: 'A regular's card is declined during the 8am rush and the queue is out the door. What do you do first?' There's nothing to look up, only judgement, and two minutes leaves little room to run it through a phone.
Three habits help. Ask for a first step, not a full plan. Ask about the candidate's own experience with a specific hook ('the busiest shift you've worked, and what went wrong'). Keep the quiz short enough that copying takes longer than answering. Quiz-based screening filters for the interview; it doesn't replace it.
What do the EU AI Act and GDPR require from you?
Less than the headlines suggest, for now. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27-07-2026) pushed the AI Act's high-risk obligations for recruitment tools that filter applications and evaluate candidates from 02-08-2026 to 02-12-2027. Article 50 transparency duties did apply from 02-08-2026, but they cover telling people they're interacting with an AI system and marking synthetic content. A quiz that logs timing isn't a chatbot.
What applies today is GDPR. Tell candidates, before they start, what the quiz records (answers, time per question, page focus, paste events) and why. Keep the decision human, per Article 22. Gartner predicts one in four candidate profiles will be fake by 2028 (as reported by HR Dive, August 2025); the answer is verification you can explain, not surveillance you can't. Our EU AI Act guide has the timeline.
Signals a human can act on
Qwiza flags per-question timing, tab switching and pasted text so your recruiter can take a closer look. Scores stay explained, nobody is auto-rejected, first campaign free.
Try the live demoFrequently asked questions
Can a screening quiz prove a candidate used ChatGPT?
No, and any tool that claims it can should be asked for its false-positive rate. What a quiz can record is behaviour: how long each question took, whether the page lost focus, whether text was pasted. Each pattern is consistent with outside help and also with a slow connection, an interruption or assistive technology. Treat a flag as a reason to verify at interview by asking the same question live. That's cheaper than a detection tool, fairer to the candidate and, under GDPR Article 22, the only defensible way to let a signal influence a hiring decision.
Should I tell candidates the quiz records timing and tab switching?
Yes, before they start, in plain words. GDPR requires you to explain what personal data you collect and why, and time per question, page focus and paste events are personal data once tied to a name. A one-line notice ('this quiz records your answers, time per question and whether you leave the page, to help us prepare interview questions') does the job. It also changes behaviour: a candidate who knows the page is watched is less likely to open a second tab, which cuts your flags before you read one. Hiding it buys you nothing and costs trust.
Is a consistent 3-5 second delay after every question proof of AI help?
No. The figure comes from FabricHQ, a vendor analysing 19,368 AI-conducted interviews on its own platform between July 2025 and January 2026, where 38.5 per cent of candidates were flagged. The vendor describes the uniform delay as one tell it watches for, not a proven rule, and it publishes no false-positive rate. On a written quiz the equivalent is per-question timing that doesn't vary with difficulty. Note it, compare it with the other signals, and put the hardest question to the candidate live. If the reasoning holds, the delay meant nothing.


