Short answer: You can make cheating in AI and remote interviews harder without collecting biometrics. Detect what the browser can reliably see (leaving the window, switching tabs, exiting full screen, a second screen, no face or more than one face), pause and warn rather than auto-reject, and send anything unusual to a person who reviews the recording. Accept that some things, like a phone held off camera, can't be detected, and design your questions and review process with that in mind.
#What counts as cheating in an AI interview?
In a structured interview, cheating means presenting answers that aren't the candidate's own: reading AI-generated answers from another screen, being coached by someone in the room, or having someone else take the interview altogether.
Three terms are worth keeping distinct:
- Integrity signal: an event the interview software records, such as "left the tab at 10:02:11". It's a fact about the session, not a judgement about the person.
- Face detection (presence): checking whether a face is in the camera frame, and how many. It answers "is someone there?", not "who is this?".
- Face recognition (identification or verification): comparing a face against a stored template to establish identity. This is biometric processing, and the law treats it very differently.
Most of the legal risk in interview integrity comes from blurring the second and third.
#Why interview integrity matters now
Remote hiring has made impersonation a real, documented problem rather than a hypothetical one. In 2022, the FBI's Internet Crime Complaint Center reported an increase in complaints about deepfakes and stolen personal information used to apply for remote work positions (opens in a new tab), noting that interviewees' lip movements didn't always match the audio. In January 2025, the FBI warned that North Korean IT workers had been observed using AI and face-swapping technology during video interviews (opens in a new tab), and recommended identity verification during interviewing and onboarding.
Most candidates aren't fraudsters, though. The goal is a process that keeps answers honest for everyone without treating every applicant as a suspect.
#What employers can detect in a browser
Modern browsers expose a handful of reliable signals. The Page Visibility API (opens in a new tab) tells a page when it becomes hidden or visible, and Chromium-based browsers report through screen.isExtended (opens in a new tab) whether more than one screen is connected.
| Signal | How it's detected | How much to trust it |
|---|---|---|
| Switched tabs or minimized | Page Visibility API | High: the page really was hidden |
| Clicked into another app or window | Window loses focus | Medium: assistive tools and system pop-ups also take focus |
| Left full screen | Fullscreen API events | High, but some exits are accidental (the Esc key) |
| Second screen connected | screen.isExtended (Chrome and Edge) |
High where supported; not available in every browser |
| No face, or more than one face | On-device face presence check | Medium: lighting, camera angle and children walking past cause false alarms |
| Paste into a text answer | Clipboard events | High, but only matters if typed answers are allowed |
| Second device or browser session | Server-side session binding | High |
Because these checks depend on browser features, it's reasonable to require a specific setup. MSZ Hire, for example, runs interviews only in Chrome or Edge on a computer, in full screen, because that's where these signals behave consistently.
#What you can't detect (and shouldn't pretend to)
Be explicit, internally and with candidates, about the limits:
- A phone or tablet held off camera. No browser can see it.
- Notes on paper, an earpiece, or a helper behind the screen. A reviewer may notice eyes reading off screen or a second voice in the recording; software won't reliably flag it.
- A modified browser or remote-desktop session. Client-side signals can be suppressed or forged by a determined person.
That's why the best defences against AI-written answers are in the interview design, not the surveillance. Ask for specific, first-person examples ("Tell me about a time you handled a customer who was charged twice"), score evidence rather than polish, and use a structured scoring rubric that rewards concrete detail. Generic answers score poorly whether a person or a chatbot wrote them.
#A fair approach: pause, warn, then human review
A proportionate integrity policy gives candidates a chance to correct honest mistakes and keeps every consequential decision with a person. Here's a model you can adapt:
- Tell candidates the rules before they consent. List what's checked (presence on camera, staying in the window, full screen) and what isn't (faces, expressions and voice aren't analyzed).
- Check the setup before the first question. Confirm the camera, microphone, one face in frame and a single screen, so problems surface before anything counts.
- Pause on departure. If the candidate leaves the window, switches tabs, exits full screen or connects a second screen, pause the interview and show a clear warning: "Warning 1 of 2".
- Count incidents, not raw events. One alt-tab can trigger three events (focus lost, tab hidden, full screen exited). Count it once.
- End, don't reject, on the third departure. Submit the answers so far for human review, labelled with what happened.
- Give reviewers context. Show each incident with its time, duration and a link to that moment in the recording, so a person can judge whether it was a notification or a pattern.
- Decide with a person. A reviewer can invite the candidate to a live follow-up, ask for identity verification at the next stage, or move on. The software never decides.
This is also what makes the approach defensible under laws that restrict decisions based solely on automated processing, such as Quebec's section 12.1. Our guide to Quebec Law 25 and AI interviews explains why a genuine human review matters.
#Legal guardrails: stay out of biometrics
This section is general information as of November 2026, not legal advice.
#Illinois BIPA
The Illinois Biometric Information Privacy Act (opens in a new tab) defines a biometric identifier as "a retina or iris scan, fingerprint, voiceprint, or scan of hand or face geometry", and expressly excludes photographs. Before collecting a biometric identifier, a private entity must inform the person in writing of the purpose and retention period and obtain a written release, and it must publish a retention and destruction policy. BIPA gives individuals a private right of action, with liquidated damages of $1,000 per negligent violation and $5,000 per intentional or reckless violation.
The practical line: a face recognition or "identity match" feature that derives face geometry almost certainly needs BIPA consent. A presence check that runs on the candidate's device and sends only an event name ("no face at 10:04") is designed not to create face geometry at all. Ask vendors precisely what is computed, where, and what leaves the device.
#Quebec biometrics and Law 25
Quebec is stricter than most jurisdictions. According to the Commission d'accès à l'information (opens in a new tab), biometric identity verification can't be required, needs express consent with a non-biometric alternative, and must be disclosed to the Commission; a biometric database must be disclosed at least 60 days before it goes into service. Separately, since 22 September 2023, Law 25 has required organizations that use technology able to identify, locate or profile a person (opens in a new tab) to tell them about it.
#Federal guidance in Canada
The Office of the Privacy Commissioner's 2025 guidance on biometrics (opens in a new tab) says consent to collect photos or video doesn't automatically allow you to extract biometric information from them, and warns that even a face detection system that assigns a unique numerical representation to a face can involve sensitive biometric information, however briefly it's kept. That's the reason to prefer detectors that output only "face present" or "how many", with no embeddings or templates.
#Maryland
Maryland has prohibited employers since 1 October 2020 from using certain facial recognition services during a job interview (opens in a new tab) unless the applicant consents by signing a waiver.
For a wider tour of US rules, including Illinois' separate AI Video Interview Act, see our guide to hiring US candidates with AI interviews.
#Accommodations: integrity rules must bend
Anti-cheating rules can quietly discriminate. A screen reader, magnifier, dictation tool or on-screen keyboard can take window focus without the candidate going anywhere. A candidate with a tremor may trigger extra events. Someone without a reliable camera can't pass a face check at all.
In Ontario, employers must tell individually selected applicants that accommodations are available on request (opens in a new tab) for the assessment or selection process, and consult with anyone who asks. Build this in:
- Offer an interview without video, typed answers or a human interview before the interview starts, with a visible link to ask.
- For approved accommodations, stop counting focus changes on their own while still counting clearer signs such as hidden tabs.
- Train reviewers to read integrity signals alongside accommodation context, and keep disability details out of what hiring managers see.
#A quick interview-integrity checklist
- Candidates see what's monitored, and what isn't, before they consent.
- Setup checks run before the first question.
- Departures pause the interview with numbered warnings.
- Nothing is rejected automatically; a person reviews every flagged interview.
- No face recognition, face templates, emotion analysis or voiceprints.
- Recordings have a fixed retention period, and viewing is logged.
- Accommodations, including a human interview, are offered up front.
- Questions ask for specific, personal evidence that's hard to outsource.
#How MSZ Hire approaches it
MSZ Hire's structured AI voice interviews run in Chrome or Edge on a computer, in full screen, with the camera recorded for human review. Face presence checks run on the candidate's device, with no biometric templates and no emotion or voice analysis. Leaving the window pauses the questions with a warning (the clock keeps running), and the third time ends it and sends the answers to the hiring team for review. Question order is randomized per candidate, recordings and transcripts are deleted after 180 days, and video and PDF downloads are audit-logged. You can see how an interview works, or start free and invite yourself to judge the experience first-hand.
Written by the MSZ Hire Editorial Team, MSZ Systems Inc., Toronto. Last reviewed 16 November 2026. General information, not legal advice.