What counts as interview cheating, and how common is it?
Interview cheating is help the candidate isn't allowed: AI-written answers, notes or searches, someone prompting off camera, copied code, or a stand-in taking the interview. Nobody has shown how common it is with a method you can check, so this guide quotes no prevalence figure.
The figures that circulate come from vendors who sell detection, and they disagree with each other. One vendor blog gives a range of adoption rates and a detection accuracy with no method behind either. Other pages repeat survey numbers that we could not find on the publisher's own page. A number you can't check is a poor basis for a hiring rule, so the rest of this guide works from what an interview can actually show.
What matters more than the rate is the form. Each kind of cheating leaves different traces, and an online interview sees some of them and misses others.
| Form | What it looks like | Where it can show | What a tool can't see |
|---|---|---|---|
| AI-written answers | A chatbot open in another tab; the answer pasted in or retyped | The answer itself; a paste from outside the page; time away from the page | Retyped text; a chatbot on a phone |
| An AI assistant that listens | An app hears the question and shows an answer on screen or on a phone | The answer; eyes reading off to one side in webcam stills or video | Anything shown outside the shared screen; a phone below the camera |
| Notes, search or a script | A prepared script read aloud, or a search during the question | The answer's delivery; time away from the page | Paper on the desk; notes on other screens |
| Someone prompting | A friend off camera, or a voice over an earpiece | A second face in a webcam still; a voice on continuous video, for a person to hear | Anyone out of frame; a quiet earpiece |
| Copied code | A solution pasted from a chatbot or a solutions site | A paste from outside the page; code the candidate can't explain | Code retyped by hand from a phone or a separate display |
| A stand-in (proxy) | Someone else answers in the candidate's name | A Selfie that doesn't match the ID photo, or a webcam still that doesn't match the Selfie | Whether the ID document itself is genuine |
Source: browser specs and MDN as cited below; this guide's reading of what each check looks at, 17 Nov 2026
Where your rules allow notes or AI, using them isn't cheating. You set the rule, and you tell candidates before they start. If you are new to one-way interviews, start with how AI interviews are scored, then come back here.
How to spot a stand-in, a swap after the ID check or a deepfake candidate, and what a face match can tell you, is in the guide to proxy interviews.
How do candidates use AI to cheat in interviews?
Candidates use AI in three main ways: a chatbot in another tab whose answer they paste or retype, an assistant that listens to the question and shows an answer, and a written script they read aloud. Typed one-way interviews invite the first; video and live rounds invite the second and third.
Paste or retype from a chatbot
The simplest route is to copy the question into a chatbot and bring the answer back. In a typed interview, that answer often arrives as one long paste, with no pauses or corrections in between. It tends to be well organised and general: it covers the topic, but it names no project, number or decision from the candidate's own work.
A careful candidate retypes the answer instead of pasting it. Retyping leaves no paste for the browser to see. What remains is the time spent away from the interview page while the chatbot was open, and the answer itself.
Assistants that listen and answer
A newer kind of tool listens to the interviewer or the recorded question and puts a suggested answer on the candidate's screen, sometimes as an overlay, sometimes on a phone propped beside the laptop. The candidate reads it out in their own voice.
A screen recording shows only the screen the candidate chose to share. It doesn't show what's on other screens or on a phone. The traces such a tool leaves are in the delivery: a steady reading cadence, eyes moving across text off to one side, and pauses that fall before the answer rather than inside it.
Coding rounds
In a coding round the help is an AI-written solution, pasted or retyped into the editor. The answer may compile and look clean. What gives it away is the follow-up: the candidate can't say why they chose one approach over another, or what happens when an input changes.
The one published experiment with a method, from interviewing.io, is about exactly this kind of round. It is set out in the next section, because it says more about question design than about catching anyone.
How can you detect ChatGPT use in an interview?
You can't prove ChatGPT use from one answer; you can collect signals and then ask. Look at the answer, at the session (pastes from outside, time away while a question is open) and at the follow-up, where the candidate explains their own answer. Decide after you have asked.
The experiment with a method
interviewing.io ran 37 mock technical interviews in which candidates were told to use ChatGPT, and it was able to use 32 of them. Not one interviewer raised a concern that any candidate was cheating (interviewing.io's blog, undated, read 17 Nov 2026).
The questions told a different story. Candidates using ChatGPT passed 73% of the interviews with verbatim LeetCode questions, 67% with modified LeetCode questions and 25% with custom questions, against a 53% pass rate in interviewing.io's ordinary mock interviews.
The lesson is plain. The interviewers' eyes did not separate the candidates; the questions did. A question the chatbot has seen many times is easy to answer with help. A question about something specific, with a follow-up, is not.
AI-text checkers
Tools that judge whether a piece of text was written by AI give a likelihood, not a verdict. No vendor we read for this guide publishes how often its checker is right with a method a buyer could check, so this guide quotes no figure. Treat a high likelihood as one more reason to ask, never as the answer.
Signals worth a question
These work with any tool, and with none:
- The answer has no detail from the candidate's own work: no project name, no number, no decision they made.
- The answer and the CV disagree, on dates, tools or the size of a team.
- A long paste arrives from outside the interview page.
- The candidate leaves the page during one question and comes back with a finished answer.
- On voice or video, the answer has the steady rhythm of reading rather than thinking.
- Asked to explain one step, the candidate can't.
None of these on its own means cheating. Two or three together, on the same question, are worth a specific question in the next round.
What is tab switch detection, and what does it tell you?
Tab switch detection is the browser telling the interview page that it was hidden or lost focus, and for how long. It shows the candidate left the page; it can't show why, or what they looked at. A single switch means little, and a pattern is worth a question.
The browser has two signals for this. The page's visibility changes to hidden when the candidate switches tab, minimises the window or, on a phone, switches app; MDN describes that change as the last event a page can reliably observe (MDN, visibilitychange, read 17 Nov 2026). The window also loses focus when the candidate clicks into another app while the page stays visible (WHATWG HTML, focus).
Some things leave no trace. On most browsers, a window placed over the interview page fires nothing at all. It doesn't see what's on other screens unless the candidate clicks away from the page. It never sees a phone out of the camera's view, or paper on the desk.
Each signal in more depth, with the browser feature behind it and which browsers support it, is in what browsers can detect in an interview.
False alarms are common. An operating system notification, a password manager, a language input window or the screen locking can all take focus for a moment. That is why a well-built tool records how long the candidate was away as well as the switch, and why a single short switch should carry almost no weight.
| Signal | What the browser sees | What it misses |
|---|---|---|
| Leaving the page (tab or focus) | That the page was hidden or lost focus, and for how long | A window covering the page on most browsers; other screens; other devices |
| Copy and paste | Pastes into the page, with their text and length | Retyping, dictation, text dragged in |
| Developer tools | No browser feature reports it; only the keyboard shortcut can be seen | Developer tools opened from the browser menu |
| Locking the browser to one window | A page can't do it: browsers keep a way out by design (WHATWG spec), and iPhone Safari can't do it at all | Everything outside that one window |
| Webcam stills | At the moment of the still: whether a face is there, how many, which way the head is turned | Everything between stills; anything out of frame; all audio |
| Screen recording | The screen the candidate shared, on a desktop or laptop only; the candidate can stop sharing | Other screens, phones, anything spoken |
| Face match (ID photo against a selfie) | Not a browser feature: a server compares the two faces | Whether the ID document is genuine |
Source: MDN, WHATWG and W3C specs (Clipboard API, Media Capture, Screen Capture) and the AWS Rekognition docs, as linked; checked 17 Nov 2026
What the time away tells you
The useful facts are when the candidate left, for how long, and what came next. Ten seconds away at the start of an interview, before the first answer, is usually a candidate closing something they forgot. A minute away in the middle of a hard question, followed by a finished answer typed in one go, is a different picture.
Read the time away against the question that was open. If the candidate left during the one question that needed a specific fact or a piece of code, and came back with it, that is the question to ask about in the next round. If the time away is spread evenly across easy and hard questions alike, it more likely says something about the candidate's room than about their answers.
Pastes, developer tools and the rest
A paste is the clearest trace a browser gets. When text is pasted into the page, the page can read that text and its length without asking permission (W3C Clipboard API). It can also tell a paste of text copied inside the interview from one brought in from outside.
The gaps are just as clear. Text dragged into an answer box arrives as a drop, not a paste, and a tool that listens only for pastes never sees it. Retyped text and dictation leave no clipboard trace at all. And a candidate pasting their own draft from a notes app looks exactly like a candidate pasting a chatbot's answer.
Developer tools have no browser feature that reports them. A tool can notice a developer-tools keyboard shortcut pressed while the interview page has focus, and nothing more. Treat any vendor claim beyond that with care.
Most interview tools record tab switches in some form. Hyring, for example, says tab changes are one of 20 integrity signals its Hyring Meet tool checks, and calls a flag "a pointer into the recording, not a conclusion" (Hyring help desk, dated 3 Mar 2026, read 17 Nov 2026). It doesn't publish what turns tab changes into a concern. That rule is the part worth asking any vendor for.
How do you design an interview that's hard to cheat?
Ask about the candidate's own work rather than textbook questions, set a time limit per question, follow up on one first-round answer in the live round, and tell candidates the rules before they start. In interviewing.io's experiment, candidates using ChatGPT passed far fewer custom questions than textbook ones.
Question design does more than any check, because it changes what help is worth. A chatbot can write a good answer to "What is a REST API?" It can't know which API this candidate built, what broke in production, or why they picked one database over another.
- Ask for a specific project, number or decision from the CV.
- Vary the question set between candidates where the role allows.
- Set a time limit per question for typed answers.
- Prefer one webcam check over a lockdown install that loses candidates.
- Check ID before an offer, and compare the first-round face with the live round.
- Say in the invite what's allowed: notes, AI, or nothing.
Follow up in the live round
The follow-up is the cheapest check there is. Pick one answer from the first round, the one with the most specific claim in it, and ask the candidate to go one level deeper: why that index, what the query does when the table doubles, what they would change now. A candidate who wrote the answer can do this in a minute. A candidate who read it can't, and no browser check is needed to tell the difference.
Write the follow-up down before the live round, next to the first-round answer it comes from. That keeps it about the answer, not about the flag, and it gives every candidate the same chance to explain their own work.
Time limits and webcam checks
A time limit per question narrows the window for help. It won't stop a candidate who has a listening assistant on a phone, but it makes a chatbot round trip harder, and a typed answer that arrives fully formed just before the limit is worth a look. Set limits from how long the answer should take someone who knows it, with room to think.
A webcam still every so often shows who is at the desk and whether anyone else is in the frame. It is a lighter ask of the candidate than installing a secure browser, which some candidates can't do on a work laptop or a shared machine. Pick the lightest check that answers the question you actually have about the role.
The last item in the checklist is also a legal point in India. Webcam stills, recordings and ID photos are candidates' personal data under the Digital Personal Data Protection Act, 2023, and section 5(1) asks for a notice of what is collected and why, given with or before the request for consent. Those duties commence on or about 13 May 2027. This is a summary, not legal advice.
Questions to ask any interview vendor
Take these to any tool you are weighing, Recagn included:
- Does it publish the rules that turn a signal into a flag?
- Is a flag a level or a number, and does it enter the score or the ranking?
- Does your client see it?
- What does the candidate have to install?
- How is the ID checked: by face match, by a document check, or both?
- What do the checks add to the price of one interview, before GST?
A vendor that can answer all six in writing has thought about what its flags mean. A vendor that answers with an accuracy figure and no method has not shown you anything you can check.
What should you do when a candidate is flagged?
Read the finding, look at the moment it points to, and ask the candidate before you decide. One flag is not proof: a tab switch can be a notification and a mismatched face can be bad light. Weigh the pattern and the answers together, and keep the flag out of what the client sees.
- Read what happened, and when.
- Open the evidence at that moment: the question, the still, the paste.
- Compare it with the quality of the answer to that question.
- Ask the candidate a specific question about it in the next round.
- Record the reason for your decision.
Face matching needs the same care as everything else. Amazon's own documentation for its face comparison service says it uses machine learning, "which are probabilistic", and recommends that a person review the result before any decision that affects someone's rights or access to services (AWS Rekognition, CompareFaces, read 17 Nov 2026). An old ID photo, glare or poor light all make a match less likely.
What to write down
Write down the flag, the moment it points to, the question you asked, and what the candidate said. One line each is enough. If a colleague picks up the role later, or a candidate asks why they didn't move on, that note is what you will rely on.
Keep the note about what happened, not about the candidate's character. "Left the page for most of question 2 and could not explain the indexes in the live round" is a fact you can stand behind. "Seemed dishonest" is not.
What the client sees
The client hired you to send people who can do the job. Send them your decision and the evidence for it: the scores, the answers, your notes on fit. A raw flag, without the context of the question you asked and the answer you got, invites the client to make a call on evidence they can't weigh. Keep it in your own review.
The decision is yours. A flag is evidence about the session, and the candidate's answer to your question is evidence too. Record both.
When are cheating checks the wrong tool?
Browser-based checks are the wrong tool for a high-stakes exam that needs a live invigilator or a locked-down machine, for a senior search where a long live conversation is the real test, and where the ID document itself must be checked against a government record.
A certification exam or a campus test with thousands of candidates needs a different setup: an installed secure browser, a live invigilator watching each candidate, or both. A browser-based interview can't lock the machine, and it isn't built to.
In a senior or executive search, the live interviews do the work. A candidate who reaches the third conversation with a client's leadership team has been tested in ways no browser check adds to.
Where the document matters, such as a regulated role or a background-checked hire, use a KYC or background-check step. A face match compares two photos. It doesn't tell you the ID behind one of them is real.
How Recagn does this
Recagn is our product. Recagn records activity on every interview set up since Activity Tracking shipped, and adds webcam proctoring, an ID check by face match and screen recording where you switch them on. Its rules turn what happened into findings, one sentence each on the interview's timeline, summed up as Integrity: No, Some or Strong concerns, never a number.
Candidates see one consent screen before Start that lists each check the role switches on. Nothing is installed; the interview runs in the candidate's own browser.
Find the mark at 2:05: the finding opens at its moment, beside the Question that was on screen.
Farhan I. · Interview timeline
- Q1
- Q2
- Q3
- Q4
- Q5
- Q6
0:005:50
2:05 · Q2 · Text Strong
The webcam still at 2:05 didn't match the Selfie.
Given an orders table and a users table in PostgreSQL, write a query to find the top 5 users by total order value in the last 30 days, and explain which indexes would make it fast.
Recagn's rules list what happened; you decide what it means. Not used in scoring or ranking; the hiring company never sees it.
This is Farhan I.'s interview timeline, from a run on 17 Nov 2026 with a performed candidate. The role is Backend developer (Node.js), with webcam Snapshots and the ID check by face match switched on. Here are the steps from the last section, applied to it.
Read what happened and when. Recagn's finding reads: "The webcam still at 2:05 didn't match the Selfie." It is a Strong concern, and it is the one that sets his level.
Open the evidence at that moment. The mark sits in question 2, which asks for a PostgreSQL query and the indexes that would make it fast. Other stills in that question are marked as flagged too. The still at 2:05 is the one that didn't match.
Set it against the rest of the session. In the same question, a stretch away from the page began at 1:30 and was still running when the 2:05 still was taken. His Selfie matched his ID photo before Start. So the question is who sat at question 2, not who started the interview.
Ask before you decide. In the live round, ask him to walk through that query and its indexes. His answer settles it, not the mark.
Record the decision. It is yours. Integrity never rejects him, and the client never sees it. Recagn recommends. Only you can reject a candidate.
In the same run, a second candidate's timeline shows the ChatGPT-shaped pattern from earlier: pastes from outside the page and repeated time away.
AI-written answers are a separate check
Integrity flags pastes from outside the page. Recagn's Authenticity Detector reads the answers themselves for AI-written text, answers read aloud and reading off screen, and gives a reason for each finding. The two checks never feed each other.
CV claims
Recagn's Consistency Checker sets each claim on the CV against what the candidate said in the interview, so an answer that contradicts the CV shows up beside it.
Why you can't reliably spot an AI-written CV, and how to test its claims in the interview instead, is in the guide to AI-generated resumes.
Which check settles which kind of fake experience, from a date sum to a relieving letter, is in the guide to fake experience on a resume.
Where the rules and the costs are
The full rule table is on Recagn's Integrity rules, with the version and the date it applies from. Each check, what it records and what it adds to an interview's credits are on Recagn's AI proctoring page.
Who sees it
Only the recruiter sees Integrity. It never enters the score, the ranking or the client's Company Report. Under the level, the review carries one line: "Recagn's rules list what happened; you decide what it means. Not used in scoring or ranking; the hiring company never sees it."
Flagging what happened in the session comes off your plate. Whether it matters stays yours. Recagn does the screening. You make the calls.
Questions people ask
Isn't a tab switch just a candidate checking their notes?
Often, yes. The browser tells the interview page that it was hidden and for how long, not why, and a notification or a password manager leaves the same mark. One switch means little. Repeated time away during a single question is worth asking about in the next round.
Can a candidate be flagged without cheating?
Yes. A notification, a password manager, poor light or a dropped connection can each leave a mark on any tool's timeline. That's why a flag should start a conversation with the candidate, not end one.
Do tools like Hyring flag tab switches?
Yes. Hyring's help page says tab changes are one of 20 integrity signals its Hyring Meet tool checks, and tells recruiters to treat a flag as a pointer into the recording (Hyring help desk, dated 3 Mar 2026, read 17 Nov 2026). It doesn't publish how many tab changes raise a concern. Ask any vendor for its rules; Recagn publishes its own as Recagn's Integrity rules.
What do cheating checks add to the cost of an AI interview?
On Recagn, activity tracking, the ID check by face match and screen recording add nothing. Webcam Snapshots add one credit to an interview, at most ₹50 on Starter before GST, and Continuous webcam adds two credits for the first 60 minutes. Each check and its credits are on Recagn's AI proctoring page.
Should you tell candidates what the interview checks?
Yes, before they start. Say what's recorded (activity, webcam stills, an ID photo) and what's allowed, so no candidate can say they didn't know. India's DPDP Act will also ask for a notice of what's collected and why once its notice duties commence, on or about 13 May 2027.
Should one cheating flag disqualify a candidate?
No. A flag says what happened, not why. Read it, look at the moment it points to, ask the candidate about it, and then decide.
Is it cheating if a candidate uses AI when you didn't say not to?
Not by your rules yet. Say in the invite what's allowed (notes, AI or nothing) and judge each candidate against that. If your client is happy for candidates to use AI, assess how well they use it instead.