How Does AI Proctoring Work? What the Camera Actually Checks

AI proctoring checks webcam footage for face presence, identity, head movement, visible objects, lighting, and distance. Some systems also estimate gaze. Oxagile’s experience developing proctoring platforms shows how these observations fit into a review process: software logs time-stamped events, and a human proctor examines the footage to establish what happened.

“AI is watching” is often where an explanation of automated exam monitoring begins. It leaves the more useful part untold, like what the camera is looking for, why an ordinary movement might attract attention, and what happens to that observation afterward.

We’ll answer “how does AI proctoring work?” by following an exam from the initial identity check to the final report. Drawing on Oxagile’s experience developing EdTech solutions and proctoring systems, we’ll look at how camera observations become recorded events and what a proctor needs to interpret them. Research and examples from other products help explain each check’s limitations.

Before the exam starts: Identity and room checks

Before the first question appears, the system confirms who is sitting down and what is visible in the room. The identity check needs a reference to compare with the face on camera. That reference is the enrolled profile, which is typically a photo or biometric record captured at registration or sign-in.

At sign-in, face recognition can be supported by another identity check, as in a WebRTC proctoring platform our team modernized for a U.S. provider. 

Case in point

Scaling a WebRTC proctoring platform to 5,000+ daily exams

Oxagile modernized a platform for a U.S. proctoring provider serving students from 129 countries. The project connected biometric identity checks, automated monitoring, recordings, and reporting within a system designed to support more simultaneous exams.

• The platform supports 1,000+ simultaneous connections and 5,000+ daily exams
• A 93% increase in platform capacity
• WebRTC recording of the test taker’s face and desktop activity
• Biometric identity validation supported by challenge questions
• Automated incident alerts and reporting with penalty-point scoring

See the case

The setup also extends to the workspace. Object recognition checks the room for books, phones, smart watches, and notes. The exam rules determine what those observations mean, for instance, a textbook may be allowed in an open-book assessment and prohibited in another test. Once the initial checks are complete, monitoring follows the session as it unfolds.

How does AI proctoring work during the exam?

Monitoring can bring several inputs into a single event log. For example, Oxagile’s platform flags light and audio changes, head movements and window resizes.

What the camera checks during an AI-proctored exam

Face in frame

The most basic check establishes whether a human face remains visible. Losing sight of it may generate an absence alert even when the test taker is still at the desk. Respondus Monitor’s documentation describes how poor lighting, slouching, or tilting a laptop screen can make the face difficult to assess or crop it out of view¹.

Same person

Presence leaves the identity question open. Recognition compares the visible face with the enrolled profile, a task that depends on a usable image. Respondus tests its models using images of people with different glasses, hairstyles, and head coverings to account for variations in appearance². An identity alert needs to be considered alongside the conditions in which the image was captured.

Extra people

Face detection is also able to identify more than one face in the image. The recording gives the proctor a sequence to examine: when the additional face appeared, how long it remained, and what happened during that interval. Those details establish how the observation relates to the exam rules.

Head and eyes

The recording also shows changes in head position, including ordinary posture adjustments. Some products supplement this with an estimate of gaze direction. A recent review of 80 studies found webcam gaze tracking struggles with movement and poor light³. An apparent look away can therefore reflect an unreliable estimate. Gaze availability and alert settings vary by product.

Objects

The other part of the camera’s view is the workspace. Object detection looks for visible items such as phones, books, notes, and smart watches. A proctor examines the item against the permitted materials and uses the recording to establish how it appeared and was used during the exam.

Light level

Every visual check depends on the image it receives, so recording conditions also enter the analysis. The system checks light level, distance from camera, head movement, and headgear. Lighting changes may be logged, while dim light or glare can obscure facial details. Respondus describes prompts to correct poor lighting and improve the footage available for review¹.

Distance

This check monitors how far the face appears to be from the webcam. Posture and laptop position also affect the framing: the same resource documents slouching and accidental screen adjustments among the reasons faces become cropped¹. A distance alert consequently needs to be examined together with what remains visible in the image.

Together, these checks provide real-time video analysis of the session. Other parts of the platform add observations that originate outside the camera image.

What the camera does not check

A microphone supplies the audio, while screen capture and browser monitoring record activity on the computer. A window-resize alert might therefore appear alongside a head-movement alert in the same exam record, although they come from different inputs.

For identity verification, a platform may draw on another source of evidence, specifically, the way a person types. Oxagile’s systems also learn typing patterns as a second check on identity. This form of behavioral biometrics uses keystroke data collected through the software.

The camera’s own output remains a visual observation. It’s not capable of establishing intent, and its face checks have distinct purposes: detection locates a face, while recognition compares that face with the enrolled profile. Poor visibility often disrupts detection before a meaningful identity comparison is possible.

Why some innocent students get flagged

That dependence on the recording explains how an uneventful exam sometimes still produces alerts:

The 2026 review found darker skin tones are more likely to be misclassified. The same review warns neurodivergent behaviors are more likely to trigger alerts³.

Pre-exam setup checks help identify lighting and framing problems. Respondus reports cutting false positive flags by 80% in a year¹. The 80% figure is Respondus’s own reported improvement in Respondus Monitor.

What happens after a flag

When a check triggers a flag, the system logs the event with a timestamp. For the proctor, the next requirement is a way to open the relevant footage. Oxagile’s systems produce video reels with time-stamped events for review.

The proctor then follows what happened before and after the event. For a missing-face alert, that may reveal how a change in camera angle cropped the face. For an object alert, the reviewer examines the item and the permitted materials. The reviewer now has the opportunity to assess the initial observation against what the recording shows and the exam allows.

The review then returns to the full session. The platform scores sessions with penalty points and reports incidents in real time. A post-exam report brings the event records together. Automated scores help prioritize sessions for review, but a person determines whether the evidence establishes a violation.

AI exam proctoring narrows hours of video to a few moments worth examining. Each flag needs to give the reviewer both a reason to look and enough footage to reach a decision.

Questions to ask before you build or buy

For a team selecting or developing an AI proctoring platform, the same sequence becomes a set of decisions about how the service will operate. The questions below follow those decisions from camera checks to the people and data behind them.

For a software development project on remote proctoring, a sample exam brings these requirements together. Following it through to the final report establishes what the student sees, what the software records, and what a human reviewer needs to make a decision.

Final thoughts on AI-based proctoring

Every camera check serves a purpose, but each one also introduces the risk of false alerts. For EdTech providers, the challenge is deciding which checks are worth running, how sensitive they need to be, and how much time proctors spend reviewing the results. Those choices matter even more as exam volumes grow and requirements change.