Can Interviewers Detect an Interview Copilot?

Ninety-seven of 587 timed LastRound sessions between 21 March and 1 August 2026 ran longer than fifteen minutes. That’s 17%, and it’s the group that carries almost all of the detection risk in this category.

A tool open for forty seconds gives an interviewer very little to notice. A tool open for forty minutes gives them a lot. Most writing about detection focuses on the software. The variable that actually matters is how long you keep it running.

What can technically be detected

Start with what’s true rather than what’s marketed.

A video call platform sees what you share. If you share a single browser tab, it captures that tab. If you share your whole desktop, it captures everything drawn on it, including overlays that aren’t drawn at the operating system level.

Interview platforms cannot see your other applications, read your clipboard, or inspect your second monitor unless you install their software and grant it those permissions. Some proctored assessment tools do exactly that, and they tell you so during installation.

So the technical question splits cleanly. During an ordinary Zoom or Meet interview, detection is essentially about what’s inside your screen share. During a proctored assessment with installed software, the answer is different and you should assume much more is visible.

What actually gets people noticed

In practice it isn’t software fingerprinting. It’s behaviour, and it’s the same handful of things.

Eye movement. Reading from a second screen produces a distinctive pattern: your eyes track horizontally while you speak, and you stop looking at the camera at exactly the moments you’d normally look up to think.

Register shift. Your unassisted answers and your assisted answers sound different. Longer sentences, cleaner structure, no filler words, vocabulary a step above how you were talking a minute ago.

The pause pattern. Natural thinking pauses come before you start talking and mid-sentence as you search for a word. Reading pauses come at line breaks. They land in different places and interviewers hear the difference without necessarily identifying it.

The follow-up collapse. This is the big one. A polished answer followed by a vague, halting response to “can you go deeper on that” is the clearest signal there is, because it shows the depth wasn’t yours.

Risk by share mode

The only test worth trusting

Record yourself. Start a meeting with no one else in it, share your screen the exact way your round will require, run the tool, and record the call. Then watch the playback.

This takes ten minutes and it’s the only claim about your specific setup that’s worth anything. Vendor statements cover the vendor’s typical configuration, not your operating system, your version, or your monitor arrangement. Zoom’s documentation on screen sharing is worth reading alongside it, because the differences between share modes are where most surprises come from.

The part that isn’t technical

Whether you should use one is a separate question from whether you’d be caught, and it’s the one that actually determines your outcome if things go wrong.

Some employers state a rule in the invitation. Many say nothing, and silence isn’t permission. If a round is described as a closed-book assessment, using live assistance is a breach regardless of whether anyone notices, and the consequence is usually withdrawal of the application rather than an awkward conversation.

I don’t know how commonly employers actually detect this, and I’d be sceptical of anyone who claims a number. There’s no public dataset, and both sides have reasons not to publish one. The NIST AI Risk Management Framework gives the vocabulary being used to govern automated assessment, but it doesn’t answer this.

If you’re asked directly

It happens, and it happens more now than it did a year ago. Usually phrased as “are you using any AI assistance for this?” somewhere in the first two minutes.

Answering honestly costs less than most people assume. Interviewers asking that question are generally trying to set expectations rather than catch you, and a straight answer followed by “I can turn it off” usually ends the matter. Being caught in a denial ends the process.

What causes real damage is denying it and then producing an answer with a register shift the interviewer already noticed. At that point the issue stops being the tool and becomes whether anything you said was yours.

The technical background here is worth knowing too. Screen sharing in browsers runs through a documented API, and MDN’s write-up of the Screen Capture API sets out exactly what a page can and cannot request. A meeting site can ask for a tab, a window, or a display. It cannot silently widen that choice after you’ve made it.

Frequently asked questions

Can interviewers detect an AI interview copilot?

Not usually through software during an ordinary video call, unless the tool appears in your screen share. What interviewers do notice is behavioural: eye movement, a shift in speaking register, and weak answers to follow-up questions.

Can Zoom or Google Meet see my other applications?

No. They capture what you choose to share. Sharing a single tab captures that tab only. Sharing your entire desktop captures everything drawn on it, which is where browser-based tools become visible.

What about proctored assessments?

Different situation entirely. Proctoring software you install can monitor processes, restrict applications, and use your camera. Assume much more is visible, and read what the installer asks permission for.

What gives candidates away most often?

A strong opening answer followed by a vague response when asked to go deeper. It shows the depth wasn’t yours. No amount of stealth in the software addresses that.

How do I check whether my setup is visible?

Start a meeting alone, share your screen exactly as the round will require, run the tool, and record it. Watch the playback. It’s the only test that reflects your own operating system and configuration.

One sentence

The software is rarely what catches people. The follow-up question is.