AI proctoring in a voice interview: what to watch

On 2 November 2026 I pulled Google’s top ten results for AI proctoring in India and in the US, along with the AI answer that sits above them. Nearly every result on both pages was about exams: university tests, certification centres, online assessment platforms. Google’s AI answer in the US described it as software that monitors “online test-takers”. The questions Google suggested underneath were things like whether a remote proctor can see your screen and how to prove you are not using AI.

Nobody on those pages was talking about a job interview. Yet most of the hiring teams I speak to now ask some version of the same question during a demo: “Is your AI interview proctored?” They mean exam proctoring, because that is the only kind anyone has written about. This post is about the other kind. It opens a five-part series on how tab-switch and camera-off detection keep a voice interview honest, and it starts with the design question underneath all of it: what should proctoring watch in an interview, and what should it deliberately leave alone?

What is AI proctoring for interviews?

AI proctoring for interviews is software that watches for signs of outside help while a candidate is answering questions in a remote interview. It tracks events such as the interview tab being hidden, the camera being switched off, or answer timing that points to a second screen, then either stops the call or flags it for a recruiter to check in the recording.

The important word in that definition is “or”. Every proctoring system has to decide, signal by signal, whether an event is serious and unambiguous enough to act on automatically, or ambiguous enough that a human should look first. Exam proctoring leans hard towards acting. Interview proctoring, done properly, leans the other way. The rest of this post explains why, and where we drew the line.

Why exam proctoring does not fit a screening interview

Exam proctoring grew out of a specific problem: a test with a fixed answer key, taken by people who have every reason to look up the answers. If the answers exist somewhere, the only defence is to cut the candidate off from every other source. So exam tools lock the browser, block copy and paste, check for second monitors, scan the room with the webcam and check ID at the door.

A screening interview is a different object. The best questions in it have no answer key. “Walk me through the last customer complaint you handled” or “Why did you pick that database for your final-year project?” have answers that only the candidate has, because they are about the candidate’s own work. The strongest anti-cheating tool in an interview is not a lock. It is the follow-up question that a scripted answer cannot survive.

There is also a cost exam proctoring can afford and screening cannot. A student sitting a board exam will install whatever software the university demands, because the exam is not optional. A candidate who receives your screening link is choosing between you and the next employer. Every extra download, permission prompt or room scan loses some honest people before the first question. In high-volume hiring that loss is invisible on the day and very visible in the funnel a month later.

And some exam-style checks carry legal weight of their own. In August 2022 a federal court in Ohio held, in Ogletree v. Cleveland State University, that a webcam scan of a student’s bedroom before a remote test was an unreasonable search, noting that “room scans go where people otherwise would not”. That case was about a public university, not a private employer, but the instinct behind it travels: a hiring step that films a stranger’s home is a hard thing to defend to a candidate, a works council or a regulator.

The proctoring ladder: stop, flag, record, skip

When we designed proctoring for HireQwik’s browser interviews, we sorted every possible signal into four rungs. I find this the most useful way to evaluate any vendor’s proctoring, including ours.

The test for the top rung is strict. An event belongs in Stop only if a reasonable candidate would agree, in advance, that it breaks the rules, and if the software can see it reliably without guessing. Everything that needs interpretation goes lower.

The two hard stops in a HireQwik interview

Only two events made the Stop rung.

The interview tab going hidden. Browsers fire a standard event when a page becomes “not visible at all on any screen”, in the words of the W3C Page Visibility specification: a background tab, a minimised window. When that happens during a HireQwik interview, the candidate sees a warning that counts 1 of 3, 2 of 3, 3 of 3. The fourth time the page goes hidden, the call ends. The full mechanics, including the five-second timer on the warning that we would design differently today, are in what Tab Switch Detected means.

The camera being switched off. When the candidate turns the camera off, the microphone is muted and a blocking screen stays up until they press Turn On Camera. The interview is genuinely paused, not merely warned. Again there are three warnings, and the fourth camera shutoff ends the call. It is a separate counter from the tab rule, so a candidate who hides the tab twice and toggles the camera twice is still in the interview. When the camera is turned off walks through each screen.

Both rules pass the strict test. Both are visible to the candidate the moment they happen. Both are announced in the lobby before the interview starts, in a notice that says the interview is monitored with video recording, that the camera must stay on, and that switching tabs or turning off the camera will trigger warnings and may lead to termination. And both are mechanical: the browser reports the tab hidden or the camera track switched off, so the software is not guessing.

Notice what the camera rule does not do. It counts the camera control. It does not look at the picture. A hand over the lens with the camera still enabled does not pause anything, and we say so in our own help pages rather than pretend otherwise.

What a voice interview should only flag

The second rung is where most of the real detection lives, and where most of the mistakes in this category get made.

HireQwik computes a set of AI-assist signals after every completed browser interview, from data the call already produced. They include tab switches, long answer latency, a “slow but fluent” pattern (a long wait followed by an unusually smooth answer, the fingerprint of someone reading a chatbot’s output), a reading-aloud pattern, essay-style structure, zero self-correction, and camera-off events. Since 15 October 2026 they also include app switching: the interview window losing focus to another application for at least two seconds while the tab itself stays visible, which is how a side-by-side window or a second monitor shows up. The candidate is not shown anything when this happens. It is recorded for the recruiter.

All of these are flags. They are combined into a low, elevated or high band, shown to the recruiter with the evidence, and labelled as signals, not proof. They never change a score, a verdict or an automatic decision, and our test suite fails if anyone wires them into one.

Why so cautious? Because every one of them has honest explanations. A long pause is also a nervous fresher, a slow connection or someone translating in their head. A window losing focus is also a notification, a password manager or a permission prompt. When we set the thresholds we measured them on 1,447 completed production interviews rather than guessing. The median answer started 6.9 seconds after the question. A naive “eight seconds means cheating” rule would have flagged about 30 percent of honest candidates. Our thresholds sit at the tails instead: the long-latency signal needs a median of 15 seconds or more. The ChatGPT tell explains each of those signals in detail.

The app-switch flag is newer and I want to be straight about it: its thresholds (two trips away of at least two seconds each, or 20 seconds away in total) were set from first principles, not yet calibrated on production data the way the others were. That is exactly why it sits on the Flag rung and not the Stop rung.

What we chose not to collect

The Skip rung matters as much as the others, and it is where we differ most from exam proctoring.

• No lockdown browser or download. The interview runs in a normal browser tab. A forced install costs completion, and a determined cheater runs the helper on a second device anyway.
• No room scan. For the reasons the Ohio court gave, and because it tells you very little in a conversation-based interview.
• No screen recording or screen sharing. HireQwik never captures the candidate’s screen. It knows whether its own tab is visible and whether its window has focus. It does not know what else is open. The last post in this series answers whether an AI interview can see your screen in plain language for candidates.
• No gaze tracking or face analysis. The camera feed is recorded for a human reviewer. It is not scored, and nothing in the score depends on where someone’s eyes go.
• No identity verification. There is no face match and no voiceprint. That is a deliberate limit with a real cost, covered in the fourth post on why a camera being on is not proof of identity.

Each of those is a trade. Some of them we may revisit as on-device techniques improve and as customers ask for them on specific roles. But a vendor who collects all of it by default for a 10-20 minute screening conversation is solving an exam problem, not a hiring one.

How rare the events actually are

One reason to keep proctoring light is that the events it watches for are uncommon. In the same 1,447-interview sample, 3.4 percent of interviews showed any tab switch at all. When we stamped the AI-assist band onto 2,507 past interviews in August 2026, 97.4 percent came out low, 45 elevated and 2 high.

Those numbers cut both ways. They mean heavy-handed proctoring would inconvenience thousands of honest candidates to catch a small minority. They also mean the minority is real, and that when a band comes back elevated or high it deserves ten minutes with the recording, not a shrug.

They also say nothing about the cheating these signals cannot see. A helper tool that never takes focus, a friend reading answers from across the room, or a stand-in taking the whole call will not move a tab counter. The second post in this series maps six interview cheating methods to what catches each, including the ones nothing in the browser catches, and the third covers invisible AI helpers like Cluely.

Questions to ask any vendor about interview proctoring

If you are evaluating an AI interviewer, these questions separate interview proctoring from repackaged exam proctoring:

1. Which events end the interview automatically, and are candidates told in advance? You want a short, specific list and a lobby notice that names it.
2. Which signals only flag, and can a flag ever change a score or reject someone? The right answer is that flags never decide on their own.
3. How were the thresholds set? Ask for the sample size and the false-positive rate on honest candidates, not a list of signal names.
4. What do you collect that you do not need? Screen recordings, room scans and biometric templates all carry consent and retention duties.
5. What can your proctoring not see? A vendor who cannot answer this has not tested their own product.
6. What does the recruiter see after the call? Counts, a reason if the call was ended, the band with its evidence, and the recording.
7. What happens on a phone or a shared lab PC? Campus and blue-collar hiring run on cheap phones and shared machines. Ask how the rules behave there before a placement cell asks you.

What recruiters should do with proctoring results

Proctoring output is the start of a review, not the end of one. A practical routine:

• Read ended calls by reason. A call ended by the tab rule after twelve minutes of solid answers is a different case from one ended in the first minute. Look at how much was said first.
• Open the recording for every elevated or high band. Listen to the answers the signals point at. Most elevated bands are explained in a minute.
• Re-check, do not reject. Where doubt survives, offer a short live conversation. Honest candidates accept it easily.
• Write down the decision and why. If a candidate later complains, a note that says “band elevated, recording reviewed, answers specific to own project, advanced” is worth more than any score.
• Brief candidates properly. Most tab and camera events come from candidates who did not know the rules. What to tell candidates before a tab switch warning has copy you can paste into the invite.

The goal of proctoring in a screening interview is not to make cheating impossible. Nothing does that, in person or online. The goal is to make the honest path the easy one, to make the dishonest one visible to a human, and to never let a machine call someone a cheat on its own.

If you want to see the warnings, the camera pause and the AI-assist panel from the candidate’s side, take the live AI interview demo yourself. It is the same interviewer that screens real candidates, and you will get your own scorecard by email afterwards.