Remote exams have become routine at many cégeps, colleges and universities: continuing education programs for working adults, distance courses, a final for students who cannot come to campus. The question is rarely whether to supervise them. It is how to supervise them without handing the judgment to an algorithm.
If you lead continuing education, advise teachers on assessment or run IT, you have probably seen an AI proctoring demo and heard the objections that follow, from students, from teachers and from the privacy officer. This article explains how AI-based proctoring typically works, why institutions hesitate, and a human-first model that keeps supervision rigorous without facial recognition or biometric data.
How does AI-based proctoring work?
Most AI proctoring tools follow the same pattern. Software on the student's computer records the webcam, the microphone and the screen, and an algorithm analyzes the recording for behaviour associated with cheating:
- Gaze and head movement: the student looks away from the screen, down or to the side for too long.
- Faces in frame: no face, a second face, or a face that no longer matches the one captured at the start.
- Sound: voices, whispering or background noise.
- Activity: switching windows, copying and pasting, opening another application.
Each event is weighted and combined into a suspicion or risk score for the session. Reviewers sort sessions by score and watch the highest first. Identity is often checked at the start by matching a webcam photo against a photo ID or a reference image, which is facial recognition.
The appeal is real: a few reviewers can cover thousands of sessions, at any hour. Most vendors also stress that a flag is not a verdict and that a person should review it. The hesitation comes from what happens around that review.
Why do institutions hesitate?
False flags put students on the defensive. An algorithm that watches eyes and faces cannot tell thinking from cheating. A student who looks at the ceiling to recall a formula, reads a question aloud or shares a small apartment with a toddler can land in the high-risk pile. Even when a reviewer clears the flag, the student may have had to explain themselves, and the starting point has shifted from trust to suspicion.
Equity concerns are widely reported. Students and advocacy groups have widely reported that face detection struggles more often with darker skin and poor lighting, sometimes asking students to add light or move before they can even start. Students with disabilities have reported being flagged for tics, involuntary movements, atypical eye movement or the use of assistive technology. These behaviours are part of how they work, not signs of cheating.
Stress has nothing to do with the exam. Knowing that software is scoring your gaze changes how you sit an exam. Students commonly describe staring at the screen so as not to look away, or dreading that a roommate will walk in. That anxiety measures nothing the course is trying to assess.
Biometric data carries extra legal weight. A facial template is biometric data, which privacy laws treat as especially sensitive. In Quebec, verifying or confirming someone's identity through biometric characteristics brings additional obligations: the person's express consent is required, and the process must be declared to the Commission d'accès à l'information (CAI) before it is put into service. Since an exam is rarely optional, privacy officers will also ask what happens to the student who says no.
Any new proctoring system, AI or not, also calls for a privacy impact assessment, as covered in remote proctoring and Law 25. These are general obligations; confirm how they apply to your project with your privacy officer or legal counsel.
What doesn't work on its own
Institutions wary of AI proctoring often fall back on one of three options. Each has its place. None is enough on its own for a high-stakes remote exam.
- An honor code alone. It states the expectation and builds the right culture. But when a teacher has a doubt, there is nothing to look at: no record and no timeline, only a suspicion against the student's word.
- A lockdown browser with no record. Safe Exam Browser, the free and open-source lockdown browser that works with Moodle quizzes, does what it promises: it keeps the student inside the exam and blocks other sites and applications. It does not confirm who is at the keyboard, cannot see the room, and keeps no recording to review if a question comes up later.
- A video call with the teacher watching 30 tiles. A Microsoft Teams call with cameras on feels human, but no one can watch 30 small faces closely for three hours, and the screens stay invisible. Nothing is locked down, and a recording of the call shows the teacher's gallery view rather than each student's session.
The gap is the same in all three: either no one can see what happened, or no one can show it afterwards.
A human-first model for remote proctoring
The alternative to AI proctoring is not less supervision. It is supervision where software enforces the rules and documents events, and people make every judgment. Six principles apply whatever tool you choose.
1. Confirm identity with the accounts you already manage
Students sign in with their institutional account through single sign-on, so your existing password and multi-factor policies apply and there is no new account to create. A person then confirms visually that the right student is there, for example by comparing the camera image with the student's ID card. A waiting room lets the proctor validate attendance before the exam starts, and a fallback such as a one-time code keeps a sign-in failure from costing exam time.
2. Lock down the exam environment, exam by exam
Block what the exam does not allow: other websites, applications, browser extensions and AI tools. Then allow what it does, such as a PDF for an open-book exam, Excel for a lab, or a generative AI tool if the teacher permits it. Assistive technology authorized for a student must stay available. Generative AI deserves a plan of its own, discussed in how to stop ChatGPT use during online exams.
3. Record what the exam needs, and nothing more
Decide per exam which captures are justified: camera, microphone, screens, keyboard and mouse. A closed-book final may need all of them; a take-home lab may need only the screen. Tell students beforehand what is recorded, who will see it and when it will be deleted.
4. Detect events with rules, not scores
Rules are explainable: an unauthorized application opened, the exam window lost focus, the student switched screens. Place each event on a timestamped timeline that anyone can check against the recording. A rule describes what happened on the computer; it does not guess intent from eye movements.
5. Have a person review and decide
The proctor or teacher opens the timeline, jumps to the timestamp and watches the recording. Many events turn out to be harmless, such as a stray click or a notification. The ones that are not go through the institution's academic integrity process, where the student is heard.
6. Never sanction automatically
No score triggers a zero, a failure or a report. Software can raise an alert; only a person can decide that a rule was broken.
Here is how the three approaches compare:
| Criterion | AI-scored proctoring | Lockdown only | Human-reviewed recording |
|---|---|---|---|
| Identity | Often automated face matching against a photo ID | Institutional sign-in; no one sees who is at the keyboard | Institutional sign-in, confirmed visually by a proctor |
| Evidence | Flagged clips and a suspicion score | The exam platform's logs, at most | Recording plus a timestamped event timeline |
| Who decides | A person in principle, guided by the score | No one: there is nothing to review | A person, from the evidence |
| Biometric data | Often, when faces are matched | None | None |
| Student experience | Worry about being flagged for looking away or for noise | Least intrusive, least protection | Rules known in advance, a person to talk to |
How Examo handles it
Examo is built on the human-first model. It is a native Windows and macOS application that frames and records an exam, whatever tool the teacher used to build it: a Moodle quiz, a Word, Excel or PDF document, or an external exam application. It contains no AI-based feature: no facial recognition, no automated behaviour analysis and no biometric data.
- Identity. Students sign in with their institution's Microsoft 365 account, and single sign-on is mandatory. In the waiting room, the proctor or teacher validates attendance, and for remote exams the proctor confirms identity visually from the camera feed. If sign-in fails, the proctor generates a one-time access code.
- Lockdown. Unauthorized websites, applications, browser extensions and AI tools such as ChatGPT, Copilot and Gemini are blocked, and running processes are monitored. Teachers allow specific resources per exam or security profile, and assistive technologies authorized for a student are not blocked, as the accessibility page explains.
- Recording. Video and audio of the student, every connected screen, and keyboard and mouse interactions, for the whole exam. Each capture can be switched off per exam or per security profile.
- Detection and review. Unauthorized applications, sites and AI tools, loss of focus and window or screen changes are logged on a timestamped timeline, and an alert goes out when something is detected. After the exam, an Excel report of events and alerts is exported with the matching video and audio recordings. Every incident goes to human review: no suspicion score, no automatic decision or sanction.
The full list of modes and settings is on the proctoring and lockdown page.
Data is hosted in Microsoft Azure's Canadian regions and in the institution's own Microsoft 365 environment, kept for the period the institution sets, up to seven years, and never used to train AI models. The security and privacy page lists what is collected and who can see it, and the student guide explains it all to students in plain language. Examo is a new product from LS2 Innovation, a Montréal company whose software for higher education is used by about fifty colleges and universities.
Key takeaways
- AI proctoring scores gaze, faces and sound; institutions hesitate over false flags, equity concerns, student stress and the legal weight of biometric data.
- An honor code, a lockdown browser with no record or a video call with 30 tiles each leaves a gap: no one sees what happened, or no one can show it afterwards.
- A human-first model confirms identity through institutional sign-in and a person, locks down the exam, records only what it needs and logs rule-based events on a timeline.
- Software enforces the rules and documents events; a person reviews the evidence and decides, and no score ever triggers a sanction.
- In Quebec, biometric identity verification requires express consent and a prior declaration to the CAI; confirm the details with your privacy officer.
See a proctored exam with no AI in the loop
In 30 minutes, we show sign-in, the waiting room, lockdown, a detected incident and the report a teacher reviews.
Book a demoFrequently asked questions
Is AI proctoring the same as facial recognition?
Not exactly. AI proctoring is the broader practice of having an algorithm analyze webcam, audio and screen recordings to flag behaviour such as looking away, a second face or background voices, often rolled into a suspicion score. Facial recognition is one technique within it: matching the student's face against a photo ID or reference image. Many AI proctoring tools use both, which is why biometric privacy rules often apply.
How can you verify a student's identity in a remote exam without biometrics?
Combine two checks. First, students sign in with their institutional account through single sign-on, so your existing password and multi-factor policies apply. Second, a proctor confirms visually that the right person is in front of the camera, for example against the student's ID card, ideally in a waiting room before the exam starts. A one-time access code covers sign-in failures without delaying the student.
Does facial recognition for exams require consent in Quebec?
In general terms, yes. Quebec law requires a person's express consent before their identity is verified or confirmed using biometric characteristics, and the process must be declared to the Commission d'accès à l'information before it is put into service. A privacy impact assessment is also expected for a new system that handles personal information. Requirements depend on the specifics, so confirm with your privacy officer or legal counsel.
Who decides whether a student cheated in a human-reviewed model?
A person, never the software. The proctoring tool enforces the exam's rules and documents events on a timestamped timeline with the matching recording. The teacher or proctor reviews that evidence, and anything that looks like a breach goes through the institution's academic integrity process, where the student can explain. No score triggers an automatic zero, failure or sanction.
Does Examo use artificial intelligence?
No. Examo contains no AI-based feature: no facial recognition, no automated behaviour analysis and no biometric data, and recordings are never used to train AI models. Its detections are rule-based, such as an unauthorized application or AI tool, loss of focus on the exam window, or a window or screen change. Each one is timestamped and submitted to human review, and data is hosted in Canada.