Interview cheating: what catches each method

In August 2026, when we switched on AI-assist signals for our voice interviews, we also ran them backwards over every interview we already had. 2,507 past calls were stamped with a band. 97.4 percent came back low, 45 elevated and 2 high. I opened the two high ones first, expecting a clear case of interview cheating. One of them was a test call made by a member of our own team, who had been genuinely juggling another window while answering.

The signal was right about the behaviour and completely wrong about the intent. That is the most useful thing I can tell you about detecting cheating in interviews, and it shapes everything below. This is the second post in a series on AI proctoring for interviews. Here I map the six ways candidates actually cheat in a remote voice screen to the signal that catches each, including the two that nothing in a browser can see.

What counts as interview cheating?

Interview cheating is any hidden help that makes a candidate’s answers misrepresent what they know or can do. In remote interviews it usually means looking answers up during the call, reading a chatbot’s output, reading a prepared script, being fed answers by someone else, or having someone else take the interview entirely. Preparing in advance is not cheating.

That last sentence matters more than it looks. A candidate who researched your company, practised common questions and wrote notes the night before is doing exactly what every career guide tells them to do. The line is crossed when help arrives during the interview, in a form the interviewer cannot see, and changes the answer. Your invitation should say where that line sits, because honest candidates genuinely do not know whether a notepad is allowed.

How common is interview cheating?

Nobody can measure it precisely, because successful cheating is invisible by design. The public signals all point the same way, though. In March 2025 CNBC reported on a founder who opened a coding challenge to 700 applicants and judged that more than half of them cheated. The same report said Amazon now asks candidates to acknowledge that they will not use unauthorised tools, and that Google’s CEO had suggested at a February town hall that some interviews move back in person.

Our own numbers are calmer, and they come from a different kind of interview. Across 1,447 completed voice screens we measured in August 2026, 3.4 percent showed any tab switch at all. In the 2,507-call backfill, fewer than 2 percent landed in the elevated or high band. A 10-20 minute conversation about the candidate’s own experience is a harder thing to cheat than a coding puzzle with a known answer, and that difference is the strongest lever you have, which I come back to below.

Six cheating methods and the signal that catches each

The pattern in that table is the whole argument of this series. The events a browser can observe directly, a hidden tab and a switched-off camera, are the ones we enforce. Everything else is a flag or a gap.

1. Searching in another tab

The oldest move and the easiest to catch. When the interview page stops being visible, the browser fires a standard visibility event and HireQwik shows a warning that counts 1 of 3, 2 of 3, 3 of 3. The fourth time, the call ends and the record says why. Because every candidate sees the rule in the lobby before the first question, most people who trip it once do not trip it again. What Tab Switch Detected means covers the screens and the record.

2. A chatbot in a side window or on a second screen

Smarter candidates keep the interview tab visible and put ChatGPT next to it, or on a second monitor. The tab never hides, so the tab rule never fires. Two things still show. First, since 15 October 2026 we record when the interview window loses focus to another application for at least two seconds while the tab stays visible, and on browsers that expose it, whether a second screen is attached. Second, the timing gives it away: typing a question, waiting for the answer and reading it back produces a long silence followed by an unusually fluent reply. We call that “slow but fluent”, and the ChatGPT tell explains how its thresholds were set. Both are flags, never automatic decisions.

3. Reading a prepared script

Some candidates write model answers and read them out. The transcript of a read answer looks excellent. The audio does not: real speech has small hesitations, restarts and repeated words, and read speech is unnaturally clean and quick. The possibly-scripted flag fires only when hesitation, pace and self-repetition all point the same way over at least 150 words, so short answers never trigger it. Candidates reading answers covers the three signals behind it and the cases it misses.

4. Hiding from the camera

Switching the camera off to look something up, or to let someone else sit in, is caught mechanically. The microphone mutes and a blocking screen stays up until the candidate turns the camera back on, so the interview cannot continue in the dark. Three warnings, and the fourth shutoff ends the call. The camera rule counts the control, not the picture, so a hand over the lens does not trigger it. A reviewer watching the recording sees that, which is one reason the recording matters.

5. Invisible overlay apps

This is the method that worries engineering hiring managers most. Tools in this category sit in a floating layer on the candidate’s screen, listen to the question, and show an answer that does not appear in a screen share and does not take keyboard focus from the browser. No tab hides and no window blurs, so neither the tab rule nor the app-switch flag fires. What remains is the candidate’s own behaviour: answers that arrive at the same speed whether the question is easy or hard, and answers that collapse when you ask for one more level of detail. The third post in this series is entirely about Cluely and other invisible helpers, including what we can and cannot see.

6. A helper in the room, or a stand-in

Someone off-camera whispering answers, an earpiece, or a different person taking the entire call. No browser event fires for any of these. A camera that is on shows a face, but it does not prove whose. We do not verify identity, and I would rather say that plainly than imply otherwise. The defence here is process across rounds, covered in how recruiters spot a proxy interview and in the fourth post of this series on deepfake interviews.

Question design beats detection

Look again at methods 5 and 6. The two hardest forms of cheating are invisible to every browser signal, ours included. What exposes them is the interview itself.

A cheating aid is good at one thing: producing a plausible first answer to a question it can hear. It is much worse at the second and third follow-up about the candidate’s own life. So design questions where the first answer is cheap and the follow-up is expensive:

• Anchor questions in their own record. “Your resume says you cut ticket backlog by a third. What was the backlog when you started, and what did you change first?” A chatbot does not know their backlog. HireQwik jobs can be set to generate three questions from each candidate’s resume, scored on a rubric written for those questions.
• Ask for the second level. “Why that tool and not the obvious alternative?” “What went wrong the first time?” Real experience has texture. Generated answers stay at the level of the job description.
• Ask for numbers they would only know from doing it. How many calls a day, how many accounts, how long the cycle took. Wrong numbers are not cheating, but no numbers at all is informative.
• Ask what they would do differently. Hindsight is personal and hard to fake fluently.
• Use a scenario with a twist. Give a short situation, then change one fact and ask how the answer changes. Rehearsed and generated answers rarely adapt well.

None of this is a trick. Honest candidates find these questions easier than generic ones, because they are about something they actually did.

How to review a flagged interview in ten minutes

When a call ends early or a band comes back elevated or high, this is the routine I would ask any recruiter to follow:

1. Read the reason first. A call ended by the tab rule, a call ended by the camera rule, and a completed call with an elevated band are three different situations.
2. Check how much was said. Twelve minutes of specific, detailed answers before an ended call is a different case from a call that ended in the first minute.
3. Open the evidence behind the band. Each signal shows what it saw: how many tab switches, the median answer delay, how many times the window lost focus. One weak signal is rarely meaningful. Several pointing the same way deserve attention.
4. Listen to two answers, not the whole call. Pick the answer with the longest delay and the one the evaluation scored highest. Ask whether they sound like this person talking about their own work.
5. Compare with the resume. Specific answers that match the resume’s details are reassuring. Generic answers to resume-specific questions are not.
6. Decide between three outcomes. Advance, offer a short live re-check, or reject on the content of the answers. Do not reject on the flag alone.
7. Write one line on the record. What you saw, what you decided, and why. That line is what you will need if the candidate asks.

What not to do

• Do not auto-reject on a signal. Every flag has honest explanations: a nervous fresher, a bad network, a notification, someone translating in their head. Our own flags are built so they cannot change a score or a decision, and our team member’s test call is the reason I am glad they are.
• Do not accuse in writing. “We would like a short follow-up conversation” gets you the information without making a claim you cannot prove.
• Do not invent stricter rules mid-drive. If the lobby said three warnings, the fourth event ends the call. Adding a private rule that one tab switch means rejection is unfair to everyone who read the stated one.
• Do not hide the rules. Candidates who know the tab and camera rules mostly follow them. What to tell candidates before a tab switch warning has copy for the invitation.
• Do not treat a clean band as proof of honesty. Methods 5 and 6 leave no trace in the band. The questions and the later rounds carry that weight.

Put the line in the invitation

The single cheapest anti-cheating step is one sentence in the interview invitation, along the lines of: “Please answer from your own experience. Notes you prepared beforehand are fine to glance at before you start; looking things up, using AI tools or having someone help you during the interview is not. The interview will warn you if the tab is hidden or the camera is turned off.”

It does three things. It tells honest candidates exactly what is allowed, it removes the “nobody told me” defence, and it gives you a clear standard to review against later.

Interview cheating will not go away, and no software ends it. What you can do is make the honest path easy, make the common shortcuts visible, and keep a human in charge of every judgement about a person. If you would like to see how the warnings and the AI-assist panel look on a real scorecard, book a short demo and we will walk you through one with sample data.