How to tell if a candidate is using AI in an interview

Most interviewers have now had the feeling: the answers are fluent and well organised, but something is off. This guide sets out the signals worth noticing, the ones that mislead, the questions that bring outside help into the open, and how to turn a hunch into evidence you can review fairly.

Why this got harder in 2025 and 2026

Two things changed. First, real-time “interview copilot” tools became cheap and easy to use. They listen to the call, write an answer and show it in a window the vendors say is hidden from screen sharing. Final Round AI, for example, says its copilot windows are “hidden from screen sharing and recording in Zoom, Google Meet, Teams and more”.

Second, hiring teams started to measure the problem. Fabric, a vendor of AI-led interviews, analysed 19,368 interviews on its own platform between July 2025 and January 2026. It flagged 38.5% of candidates for cheating behaviour, using its own model and a threshold of more than 40% probability. That is one vendor’s data and one vendor’s definition, but it is not a small number. In a Checkr survey of 3,000 US hiring managers published in September 2025, 59% said they had suspected a candidate of using AI to misrepresent themselves. Gartner, as reported by HR Dive, predicts that by 2028 one in four candidate profiles worldwide could be fake.

You do not need to believe any single figure to accept the practical point. You will meet AI-assisted answers, and you need a consistent way to deal with them.

Timing signals: the pause-then-paragraph pattern

A copilot has to hear the question, generate an answer and display it before the candidate can start reading. That creates a recognisable rhythm.

• A pause of roughly the same length before every answer, whether the question is easy or hard.
• The question repeated back, or “That’s a great question”, to fill the gap.
• Then a long, complete answer delivered without hesitation.

Real thinking is uneven. People answer simple questions quickly and pause on hard ones, then often start, stop and restart. Ask a mix of easy and hard questions early on, so you have a sense of the candidate’s natural pace to compare against.

Eye and body signals, and why you shouldn’t over-read them

Reading looks different from remembering. Watch for:

• Eyes moving steadily left to right, line by line, near the camera or to one side.
• Gaze returning to the same spot before each answer.
• Typing or clicking noises when no typing is needed.
• A face lit by a second screen that you can’t see.

Be careful here. Many honest candidates show some of these. People with ADHD or autism may avoid eye contact. Nervous candidates look away to think. Some people put the video window on a second monitor, so they always appear to look aside. Poor lighting and a low camera angle distort gaze. Treat eye and body cues as a prompt to ask a follow-up question, never as a finding on their own.

Content signals: generic answers that collapse under follow-up

Generated answers have a style. They are well structured, balanced and general. They often open with a framework (“There are three key factors”) and close with a neat summary.

The clearer test is what happens next. Ask about something they just said and see if the detail holds:

• Ask for a specific example from their own work, with names, dates and numbers.
• Ask what went wrong, and what they would do differently.
• Ask them to explain a term they used in plain words.

A candidate who really did the work can go deeper on every turn. A candidate reading answers tends to restate the general point, or to pause again before each new detail.

Coding signals: pasted blocks and code they can’t explain

Technical rounds have their own signs.

• Code that appears in large, finished blocks rather than being built up line by line.
• No exploration: no print statements, no trial runs, no false starts.
• A correct solution, but a muddled explanation of why it works.
• Variable names or a style that changes abruptly partway through.
• Trouble adapting the solution when you change one requirement.

A shared code pad where you watch the candidate work is far more revealing than a submitted answer. If your setup flags pastes, a large paste followed by a weak explanation is worth a careful look.

Questions that expose outside help

Keep a short list of follow-ups ready. They are fair to every candidate and hard to answer from a script.

1. “Walk me through that line. Why is it there?”
2. “What happens if the input is ten times larger?”
3. “Before you run it, what output do you expect?”
4. “You mentioned [their own phrase]. Tell me more about how you used that.”
5. “What did your team lead push back on in that project?”
6. “If you could only keep one part of that design, which one, and why?”
7. “Change of plan: the data now arrives as a stream. What changes?”

Ask them in quick succession. The point is not to trip people up but to hear how they think when there is no time to read.

For a closer look at specific tools such as Cluely, Interview Coder and Final Round AI, see our guide on how to detect Cluely in interviews. If you suspect a different person altogether, read our guide to proxy interviews.

What Zoom, Meet and Teams can’t show you

General meeting apps were built for meetings. They show you a face and whatever the candidate chooses to share. They do not lock the candidate’s screen, and they do not tell you when the candidate switches tabs, leaves the window or pastes text. Overlay copilots are built specifically to stay out of their screen share.

Recording helps, and Zoom, Meet and Teams can all record on the right plans. But a recording of a call doesn’t show what happened outside the shared window. We compare the options on our page on a secure alternative to Zoom for interviews.

Make it provable: lockdown, live flags and recording

Your instincts are more useful when the session gives you hard facts to set them against.

• Fullscreen lockdown. The candidate’s interview runs in fullscreen, so leaving it is an event you can see rather than guess at.
• Live flags. You see a flag as it happens when the candidate leaves fullscreen, switches tabs, moves the cursor to another screen, or copies, cuts or pastes text. In a live interview, copying inside the shared code pad or chat is not flagged.
• Camera and recording. The session is recorded so you can review it with a colleague instead of relying on memory. Each flag is timestamped: click it in the report and you jump to that moment in the recording.
• An integrity score. Every live interview and assessment attempt gets a score out of 100. It starts at 100 and drops by a fixed amount for each flag: 15 for leaving fullscreen, 12 for the cursor leaving for another screen or app, 10 for a tab switch or the window losing focus, 8 for a paste and 5 for a copy or cut. It is rule-based, not AI, so you can always see why it fell. Scores of 80 or more show green, 50 to 79 amber and below 50 red, and in a live interview the interviewer watches it change as the call goes on.

Eagle Vision provides all of this in live interviews in fullscreen lockdown and in secure assessments. It flags behaviour, not specific apps. It will not name the tool a candidate used, and a copilot that never makes the candidate step out of the interview may not raise a flag. That is why the questions above still matter.

Treat a flag as a reason to look, not as proof

A fair process protects honest candidates and your employer brand.

• Tell candidates in advance what is monitored, what is recorded and whether any AI use is allowed.
• Review before deciding. Watch the recording around the flagged moments, ideally with a second person.
• Weigh everything together: flags, recording, follow-up answers and the rest of the process.
• Offer a second conversation when the evidence is unclear, with a different interviewer and fresh questions.
• Write down the reason for any decision based on integrity concerns.

Our guide on how to prevent cheating in online interviews covers setting the rules up before the interview starts.