How to spot candidates using AI to cheat in interviews
Updated Nov 2, 2026 · 15 min read
Candidates using ChatGPT in interviews is now a real concern for hiring teams. There is no single reliable sign of it, and a hunch is not evidence. The more dependable approach is to design the interview so that outside help is less useful, and to be open with candidates about what is expected.
How can you tell if a candidate is using AI in an interview?
You usually cannot tell from one signal whether an interviewee is using AI, so look for repeated patterns across the whole interview and then test them with job related questions and a short live task. Each pattern below can also have an honest cause, such as nerves, a poor connection or the candidate's own notes. Treat a pattern as a reason to ask another question, never as proof, use the same process for every candidate, and let a person make the call.
- Polished, generic answers. The answer is smooth but could fit any company, with no names, dates or numbers in it.
- Odd pauses. A long gap before an answer that then arrives complete, repeated for each question.
- A reading cadence. An even pace with no restarts or filler words, as if the words are being read from a screen.
- Off screen glances. Repeated looks toward the same spot away from the camera.
- Weak follow up explanations. The candidate cannot say why, or gives a different story when you ask for detail.
To check a pattern, ask about a specific project, change one detail in the middle of an answer, ask for a second example, or set a short live task such as talking through a small problem. If the candidate explains it clearly, the concern is settled. If not, you have evidence from the interview itself, and a person on your team decides what to do next.
How do candidates get outside help in interviews?
Candidates get outside help in a few common ways. Knowing them helps you design around them, not hunt for them. None of these is proof of anything on its own, because some of them look the same as a person simply thinking.
- Reading from a second screen or a notes document while answering.
- Using a chat assistant to generate an answer from the question, either by typing it in or by having it listen.
- Using an interview copilot tool that listens to the call and suggests answers on screen.
- Having another person in the room who passes notes or messages.
- Preparing scripted answers from a memorised list of common questions, with or without AI help.
The last item is not new. Candidates have always prepared for common questions. What has changed is the speed and polish of the material available, so generic answers now sound better than they used to.
Can you reliably tell when a candidate is using an AI copilot?
No sign is reliable enough to act on by itself. Pauses, reading eye movements, very smooth phrasing and generic content can all be caused by nerves, a poor connection, careful preparation or a candidate reading their own legitimate notes. Treating any one of them as proof will sooner or later lead you to accuse an honest person.
A better question is whether the candidate can show real understanding of the answer they gave. That can be tested in the conversation, and it does not require you to guess what is happening on the other side of the screen.
What interview design makes outside help less useful?
Questions that depend on the candidate's own specific experience, and follow-ups that go deeper into what they just said, are hard for outside help to answer well. Generic questions are easy to answer generically. Specific questions need specific knowledge. The table shows techniques and why they work.
| Technique | What to do | Why it helps |
|---|---|---|
| Ask about specifics | After any answer, ask which project, which month, which team and what the candidate personally did | Generated or memorised answers tend to stay general. Real experience has details |
| Ask for their own example | Ask for a situation from the candidate's own work and let them choose it | A suggested answer cannot know their real history |
| Reason aloud | Give a small problem and ask them to talk through how they would approach it | You hear the thinking, and not only a finished answer |
| Change the detail mid-answer | Add a constraint such as less time, a smaller budget or a missing tool, and ask what changes | A prepared answer breaks easily, understanding adapts |
| Ask why | Ask why they chose that option over another one | Reasons are harder to fake than conclusions |
These are good interview practice anyway. They improve the quality of what you learn from every candidate, including those who are using no outside help at all.
What follow-up questions work best?
The best follow-ups are short, specific and based on the candidate's own words. Pick a detail from the answer and ask for more about it. If the candidate says they improved a process, ask what the process was before, what exactly they changed and how they knew it worked.
- What was the hardest part of that, and what did you try first?
- Who else was involved, and what did you do that they did not?
- What would you do differently if you did it again?
- Which part of that was your decision, and which part was someone else's?
- If the situation had gone the other way, what would have been your next step?
Do not use follow-ups as traps. The aim is to understand the candidate's experience, and a candidate who has done the work will usually find these questions easy.
How do you stop candidates from using hidden AI tools in a technical interview?
You cannot block every hidden tool, so set clear rules on AI and outside help, and design the interview so that outside help is not enough. A hidden tool can produce a final answer, but it helps much less when you keep changing the problem and asking why.
- Set clear rules. Say before the interview what AI, notes and outside help are allowed, and apply the same rules to everyone.
- Check identity. Confirm that the person who joins is the person who applied, using the method your company already uses.
- Use screen share or a secure platform when it suits the role, and tell candidates about it in advance.
- Make it interactive. Ask follow up questions while the candidate works, so you hear the reasoning and not only the result.
- Change the requirements. Add a new constraint partway through and ask what changes.
- Ask for debugging. Give a broken step or a failing result and ask the candidate to find the cause out loud.
- Ask tradeoff questions. Ask why one approach was chosen over another and what it costs.
- Probe resume details. Ask about a specific project on the resume, such as what broke and how it was fixed.
- Look closer at answers that arrive very fast or very polished, and ask the candidate to explain how they got there.
Reasoning, explanation, debugging and tradeoffs tell you more than a final answer does. Use them for every candidate, not only for those you doubt.
How do you detect fake candidates in video interviews?
Combine an identity check, structured questions and a short list of video and audio red flags, and treat every flag as a reason to look closer, not as proof. A fake candidate may be a different person, a person helped by another, or a face or voice changed by software.
| Red flag | What you might notice | What to do next |
|---|---|---|
| Lip sync lag | The mouth and the voice do not line up | Ask the candidate to repeat the answer, then check identity by your usual method |
| Glitches | The picture freezes, jumps or warps when the candidate moves | Ask the candidate to move, and see whether the picture follows |
| Unnatural facial edges | Blurred or shifting edges around the hair, jaw or glasses | Ask for a short live task on camera and check identity |
| Robotic cadence | Flat, even speech with little natural change in tone | Ask an unscripted follow up and listen for a natural reply |
| Voice changes | The voice sounds different from earlier in the call or from an earlier screen | Ask a second interviewer to join, or check identity again |
| Resume details do not hold up | The candidate cannot describe a project on their own resume in specific terms | Ask spontaneous, role related follow ups about that project |
Many of these can also come from a weak connection or an old camera, so confirm with a second method before you act. Useful second methods are a live task on camera, a short screen share or whiteboard exercise, a liveness or ID check, and a second interviewer. Never accuse a candidate on a visual impression alone.
How do you catch a proxy candidate?
A proxy candidate is a different person who takes the interview in place of the applicant. To catch impersonation, confirm identity at the start of every interview with the method your company already uses, and compare the voice with any earlier call. Then ask a follow-up about a specific detail from the applicant's own resume. A stand-in usually cannot describe someone else's project in specific terms. Treat a mismatch as a reason to check again, not as proof.
What should you tell candidates up front?
Tell candidates clearly, before the interview, what is allowed. Many problems come from unclear expectations. If you do not want candidates to use notes, a chat assistant or a second person, say so in the invitation. If short notes are fine, say that too. Candidates then know the rules and can follow them.
A short statement is enough. For example: This interview is meant to hear about your own experience and how you think. Please do not use AI tools or other people to prepare answers during the interview. You can use a pen and paper. Also tell candidates how the interview will be recorded and who will see the recording.
Which AI tools help recruiters spot interview fraud?
Five kinds of tool help recruiters spot interview fraud: identity checks, interview behavior checks, risk scoring, AI detectors and automated interview fraud controls. None of them can promise that candidates cannot cheat, so use them to raise a flag for a person to review, never to reject anyone on their own.
| Tools that reduce interview fraud risk | What the tool is meant to do | Limits | Examples vendors describe |
|---|---|---|---|
| Identity checks | Check that the person on the call matches an ID or a live check | Show who joined, not who is doing the thinking | Alex AI lists identity checks on its site |
| Interview behavior checks | Watch the session for signs of a second person, a changed voice or other odd behavior | Signs can have honest causes, so a person must look | Alex AI lists interview proctoring on its site |
| Risk scoring | Cross check candidate and session data for things that do not add up, and give a risk level | A score shows where to look, it is not a verdict | Alex AI describes fraud flags in its reports, and Ribbon lists fraud flagging |
| AI detectors | Estimate whether an answer, a document or a voice was made by a machine | Can be wrong, and carry legal, bias, accommodation and transparency risks, so they should not be your main screening tool | Sold as separate tools, none named here |
| Automated interview fraud controls | Sit inside an AI interviewer: follow up questions, transcripts and flags for a person to review | A control only helps if a person reads the flags | Interviewer.ai lists fraud checks in its Professional Edition |
Vendor details come from each vendor's own site, as we read them in October 2026: Alex AI, Ribbon and Interviewer.ai. Buyers also ask about HireVue, Sherlock AI and HeyMilo. We have not checked their fraud controls, so we make no claim about them here. Ask any vendor to show in a live demo what it flags, what a person sees and what happens next. To compare HireVue with other tools on format and price, see HireVue alternatives.
Software cannot make cheating impossible, and the most useful control is still the interview itself. Follow up questions that depend on the answer, a transcript you can read, and a person who reviews every flag will catch more than any single feature.
Should companies use AI detectors when screening applicants?
No, AI detectors should not be the main tool you use to screen applicants. A detector gives an estimate that a text, voice or video was made by a machine. It does not give proof, and a wrong flag can hurt an honest candidate.
- False positives. An honest candidate can be flagged, and has no easy way to show the flag is wrong.
- Legal risk. Rules on automated tools in hiring differ by country and region, so ask your legal advisor before you use one.
- Bias. A detector may flag some writing or speech styles more often than others, so check whether its results are uneven.
- Accommodations. A candidate who uses assistive tools, or who has an agreed adjustment, can look like a mismatch to a detector.
- Transparency. Candidates should know that a detector is used and have a way to respond.
If you use one at all, treat the result as a weak signal that calls for human review, never as a reason to reject. Write a clear AI use policy, tell candidates about it, test any tool on cases where you already know the answer, and give candidates a way to appeal. Our guide to fair first round interviews covers the checks that matter more.
What should you not do?
Do not accuse a candidate on a hunch, and do not monitor candidates in ways they have not been told about. Both approaches harm honest candidates and expose your company to complaints. Keep your response proportionate to what you actually know.
- Do not tell a candidate they cheated because an answer sounded polished.
- Do not use covert screen capture, hidden recording or software on the candidate's device without telling them first.
- Do not treat reading eye movement or a pause as evidence.
- Do not reject a candidate on the basis of a suspicion you have not tested with questions.
- Do not apply extra scrutiny to some candidates and not others. Use the same process for everyone.
If you are unsure about an answer, ask another follow-up about the detail. If the candidate can explain it clearly, you have your answer. If they cannot, you have evidence from the interview itself, and a person on your team can decide what to do with it.
What if an answer still seems too polished?
Ask the candidate to explain one part of it in their own words, using a different example. Say it in a neutral way, such as: Can you give me a second example of that from a different project? Candidates who understand their material can usually do this. Record what happened and score the answer on the evidence you heard, not on your suspicion.
It also helps to compare notes with another interviewer who heard the same answer. Two people reading the same transcript with the same scoring guide will often reach a clearer view than one person working from memory.
What do voice AI interview tools do?
Voice AI interview tools run a screening interview by software, by phone or by browser audio, instead of a person asking the questions. The system asks questions, transcribes the answers, and may summarize them, score them or match them against the requirements of the role. Employers use them for speed, consistency, scale and flexibility. The main concerns are bias, accommodations, transparency, privacy and consent.
Trust the results moderately, not blindly. Read the transcript and the evidence behind a score before you rely on it, and keep a person in charge of the decision.
Do AI interview tools only summarize answers, or do they interpret them and match them to the role?
It depends on the tool: some only summarize what the candidate said, and many go further by scoring answers or matching them against role requirements. A structured tool scores each answer against competencies or a rubric for the role. A screening tool may compare answers with knock out criteria, such as work authorization or availability.
How much should you trust that? Moderately, not blindly. Accuracy and bias vary between tools and between roles, so read a sample of reports against the transcripts, ask what each score is based on, and keep a person in the loop. Our guide to how AI interview scoring works lists the questions to ask.
How do you tell a good AI interviewer from a bad one when comparing platforms?
A good AI interviewer asks questions tied to the job, shows the evidence behind every score, and leaves the hiring decision to a person. Compare platforms on the points below, and ask each vendor to show them in a live session.
| What to compare | Good sign | Warning sign |
|---|---|---|
| Questions | Built from the role, with a structured rubric tied to competencies | A generic question bank used for every role |
| Scores | Transparent, with the transcript, the evidence and an audit trail | A number with no reason behind it |
| Decisions | A person reviews the output and decides | Candidates are rejected automatically |
| Candidates | Told that AI runs the interview, with accommodations on request | No notice, and no way to ask for another format |
| Fairness and privacy | Bias testing, privacy review and data handling described in writing | Vague claims about personality, culture fit or truthfulness |
Run the same pilot role through two platforms and compare the reports side by side. Our buyer's checklist lists the questions to ask.
Where AI Interview Agents fits
In an AI Interview Agents interview, the AI interviewer asks follow-up questions based on what the candidate said, so answers are explored in more depth than a fixed list would allow. Your team can read the transcript of every interview and watch the recording when one is available, which lets you check exactly what was asked and what the candidate answered.
The interview report gives scores with written reasoning tied to what the candidate said. Scoring works from the transcript of what the candidate said, and the product does not analyse faces or expressions. A person on your hiring team makes every hiring decision, and the system does not reject candidates on its own. Before a browser interview starts, the candidate confirms that they understand it is recorded and may be shared with the hiring team. The security page describes where data is stored and who can see it. The free interview kit generator on the site can help you draft specific follow-up questions for your role.