The Proctored Interview Flag Reference, and What Trips an Honest Candidate

You have an AI-monitored or auto-proctored interview scheduled, and you want to know what the system watches so your normal, honest behavior does not get you flagged as a cheater. This is a lookup document for the person being watched, not the person watching. Jump to the signal you are worried about, read what it claims to prove and what it looks like when it lies, and set your room, disclosures, and answering habits before you log in.

Every page that ranks for AI interview detection is written for the employer buying the tool. This one inverts it: each monitored signal maps to the honest behavior that trips it, the concrete countermeasure, and the recourse no vendor page publishes.

What signals does an AI-monitored interview actually collect?

An AI-monitored interview layers several distinct signals, and each claims to prove a different thing. Knowing which is which is the whole game, because passing one does not clear you on another.

The common signals are gaze and eye-line tracking, response-latency and answer-unfolding analysis, voice biometrics or speaker recognition, second-screen and tab or audio detection, speech-uniformity or coherence analysis, and liveness or identity checks. Vendors are explicit that these observe correlates, not the act itself.

Two clarifications matter most. Gaze is a coarse orientation signal, not identity: the camera is used for gaze direction analysis only, a coarse read on whether you are looking at your screen, and it is distinct from facial recognition, which characterises individuals from biometric features. Voice biometrics targets substitution, not coaching: voice authentication verifies that the person who registered is the person taking the interview, and it says nothing about whether the person on the call is being coached.

The signal you are most likely to be falsely accused by is not on the video layer at all. It is the text-detector layer that scores your written answers, and it is where an honest candidate carries the most exposure.

Which signal is which, and what trips each one?

The table below is the core of this reference. Read the row for the signal you are worried about: what it claims to prove, the honest behavior that trips it, and the fix.

Every one of these signals turns a correlate into an accusation at a threshold. None of them observe cheating directly. They observe correlates of cheating, and then a threshold turns a correlate into an accusation. That single fact is what your setup and your recourse both hang on.

Where do false positives actually come from?

The load-bearing false-positive risk is textual, not visual. The reproducible number is a detector gap for non-native English writers, and it dwarfs anything documented on the video side.

A Stanford study published in the Cell Press journal Patterns found a 61.22% false-positive rate on essays from non-native English speakers versus about 5.1% for native speakers. The mechanism is measurable and fixable in a way that indicts the tool: enriching the vocabulary of those essays cut the average false-positive rate to 11.77%, and afterward only 1 of 91 essays was unanimously flagged as AI-written. That proves the detector measures vocabulary range, not authorship.

Non-native writers are flagged roughly 12x more often than native writers (61.22 divided by 5.1), a figure derived from the two rows above. This is why take-homes and written assessments are where an honest candidate is most exposed. For context, OpenAI retired its own AI-text classifier at 26% accuracy, and seven detectors together flagged non-native writing as AI 61% of the time.

On the behavioral side, the documented false-positive triggers are all innocent: a nervous candidate who looks away to think, a non-native speaker who pauses, a neurodivergent applicant who does not hold eye contact. Note-taking is a specific trap. A candidate looking down to read a physical notebook they took for themselves, in a country where taking notes during an interview is normal, will trigger the same flag as a candidate reading from a phone.

The detector measures vocabulary range, not authorship, so a careful answer in a second language reads as a machine.

Why is your baseline suspicion higher than it used to be?

Your baseline suspicion is higher because real cheating rose and thresholds tightened around it, sweeping in innocent correlates. The behavioral flag rate more than doubled in six months.

Fabric's tracking of more than 50,000 candidates shows the share flagged for AI-assisted cheating rose from 15% in June 2025 to 35% in December 2025. On technical roles the rate hit 48% across 19,368 interviews. As the genuine fraud signal climbs, platforms lower their thresholds, and a pause or a glance that would have passed before now clears the bar for a flag.

The flag share more than doubled, 35 divided by 15 is 2.33x, a figure derived from the rows above. This is why disclosing your setup up front is now defensive rather than optional. At the same time, adoption of detection tooling is low precisely because accuracy is low: only about 11% of FAANG interviewers report their company uses cheating-detection software, and HireVue found that fewer than 1% of candidates produced highly similar shared responses. Many flags come from platforms whose own vendors call the output probabilistic.

Context on scale: 63% of active job seekers have now been interviewed by an AI, up 13 percentage points in six months, from a survey of 2,950 people. Only 21% believe most employers use AI responsibly, and 38% have walked away from a process because it included an AI interview.

How do I set up my room so honest behavior reads as clean?

Set up a front-facing camera with your full face in frame, light in front of you, a single screen, your phone in another room, the door closed, and the desk cleared. Framing and camera angle cause more false flags than any single behavior.

Framing matters because the proctor reads what it can see. If you look down and the view is primarily of your forehead, or if you turn far enough that only one eye is visible, the system reads it as looking at unauthorized resources. Camera angle compounds this: a front-facing view allows accurate face and gaze tracking, while a side-facing camera may misread normal movements as suspicious.

Real test-takers report concrete tactics that reduce flags: laying out their materials, keeping their appearance presentable, staring intently at the screen, sitting up straight, staying in range of the camera, and using breathing techniques to stay still. A dropped feed is its own trap: a missing video or audio feed is a major flag, so if your connection stutters, contact the proctor immediately rather than continuing silently.

If your written history is what a text detector will score, the defensible move is to have your resume and answers assembled from your own record, with your own drafts on file. Refolk writes your resume from your own history and tailors it per posting, which leaves you a clear authorship trail rather than a polished document you cannot account for.

The setup-to-recourse procedure

Run these steps in order, from reading the policy through the appeal path. Each step has a clear done-state so you can tell when it is complete before the next one starts.

Sources disagree on whether human review comes before or after the outcome. Integrity Advocate places review before the decision issues; other workflows treat it as a post-decision appeal. Request review either way, because you cannot tell in advance which order your employer uses.

How this goes wrong: failure modes and false positives

The failure modes below are the most valuable part of this reference, because each is an honest behavior that the system cannot distinguish from cheating. Read the check column as your defense.

• Gaze flag on honest thinking. Looking up or away to recall triggers the same signal as reading a screen; there is no built-in distinction. Check: keep your eyes screen-ward and narrate your reasoning aloud so the thinking is audible.
• Note-glancing reads as phone-reading. Physical notes you wrote yourself flag identically to a phone. The false positive is your own scratch paper. Check: declare notes on camera at the start, or do not use them.
• Text detector penalizing careful English. A precise, low-perplexity answer reads as AI. The false positive is an ESL candidate or a formal, PhD-style writer. Check: keep your drafts, and hold that employers should never reject on a detector score alone.
• Side camera misreads motion. A profile view turns normal head movement into looking away. Check: front-facing camera with your face fully in frame.
• Accommodation behavior flagged as cheating. Screen-magnifier clicks or leaning in get flagged; one documented test was terminated at the flag limit. Check: approve accommodations in advance so approved behaviors are not treated as violations.
• Dropped audio or video read as evasion. A missing feed is a major flag. Check: contact the proctor immediately, and do not continue silently.
• Treating a probability as a verdict. A suspicion score is a correlate, not proof. Check: demand human review and the flag log rather than accepting the outcome.

The through-line across all seven is that voice biometrics and gaze answer different questions, and conflating them mislabels honesty. You can pass the identity check cleanly and still be flagged on gaze, and the reverse is equally true. They are not interchangeable safety nets.

What recourse do you have when you are flagged?

Your strongest lever is demanding the human review that best practice already prescribes, plus the timestamped flag log. Because a flag is probabilistic and cannot prove intent or authorship, a human reviewer gives context, examines edge cases, and may contact you for an explanation.

Documented redress includes advance human review and a formal appeal workflow: platforms are advised to tune thresholds to reduce false positives, enable human review, and provide an appeal workflow with human adjudication so flagged candidates can explain potential false positives. When you appeal, ask for the specific flagged moments rather than raw footage. Recruiters receive flagged moments, not hours of raw footage, and timestamped logs let a reviewer jump straight to the problem spots, so those same timestamps let you explain each one.

The stakes are real. Terminations happen at flag limits, where a test is ended midway once a candidate hits the maximum number of allowed flags. That is exactly why front-loading your disclosures and setup matters more than any argument after the fact.

Subject: Request for flag log and human review, [role] interview

Hi [name],

I understand my recent [assessment or interview] was flagged by the monitoring system. I want to resolve this and give you the context an automated score cannot.

Could you share the timestamped flag log so I can address the specific moments? I can already account for the likely triggers:
- I looked up several times to recall details; I was thinking, not reading.
- I used notes I wrote for myself, which I disclosed on camera at the start.
- My written answers are my own; I can share drafts and version history on request.

Since a suspicion score is a correlate rather than proof, I would appreciate a human review before any decision, and a redo if the log is inconclusive. I am happy to walk through any moment live.

Thank you,
[your name]

Who sits on the other side of these decisions is a real and searchable population. In Refolk's index of professional profiles, there are about 4,479 US Trust and Safety, fraud, and identity-verification professionals, against about 3,559 in the same roles in India, a US pool roughly 1.26x the size (4,479 divided by 3,559, derived). Top employers in that set included Walmart in the US and Bank of America and Barclays in India.

If you want to understand how these systems are built and where they misfire, you can find the people who build them. Refolk lets you search that population directly.

Pre-interview checklist

Run this before you log in. Each item is a check you either pass or fix, not a topic to think about.

How to keep this reference current

The signals in this document are stable, but the thresholds are not, so re-check the numbers rather than the mechanism. The flag rate that doubled from 15% to 35% in six months tells you the moving part is the threshold, not the list of signals, and thresholds tighten as real fraud rises.

Two things are worth re-checking before each interview. First, the disclosure policy, because permission is disclosure-dependent and each employer sets its own rule; get it in writing every time. Second, whether the specific platform routes flags through human review before or after the decision, since sources genuinely disagree on that order. When in doubt, ask for the flag log and human adjudication regardless. The one claim you can rely on across every platform is that these tools observe correlates, not cheating, so your job is to make your honest behavior legible and your authorship provable before the session starts.