Proctoring Face Detection Failures: What Happens When the Camera Cannot See You

The flag starts as a camera problem

Most online proctoring tools do not watch the student the way a person would. They run a face detection model over the webcam feed and report a number: the percentage of the exam during which a face was detected. When that number falls below a threshold, the system raises a low facial detection flag, and the instructor gets a report that marks the student as more in need of review. Nothing about that flag says what caused it. A dropped connection, a dim room, a student who leans out of frame, and a model that cannot see a darker face all produce the same output: the face was not detected, and the student is now a case to review.

The research that measured this is not a blog post or a vendor white paper. A study published in Frontiers in Education in 2022 examined instructor outputs from about 357 students across four courses at one institution, using a proctoring system deployed at more than 1,500 universities. Students with darker skin tones had their faces detected for an average of 78 percent of their assessment, against 87 percent for medium skin tones and 92 percent for lighter ones. Black students were distinctly less recognized than their white counterparts. And when the results were cut by both race and sex, women with the darkest skin tones were far more likely than any other group to be flagged for review. The study was the first to quantify the bias at the intersection of race and sex, and it found the bias in the flagging itself, not just in the detection.

The student who found it first

Before the study, a student found the failure by living it. Robin Pocornie, a student at Vrije Universiteit Amsterdam, filed a complaint after Proctorio's software repeatedly failed to recognize her face because of her dark skin, causing exam delays and stress. She was sometimes unable to access the exam environment at all, and was occasionally ejected from the exam with error messages reading face not found and room too dark. The Netherlands Institute for Human Rights issued an interim decision that did not accept the university's explanation, noting that it was beyond dispute that the student had problems logging in, and that scientific research shows facial recognition software generally works less well with dark-skinned people. The university had ten weeks to refute the suspicion of discrimination, or the ruling would go against it.

The numbers behind the complaint are worse than the study's averages suggest. A finding published by Motherboard, based on the student's own testing of Proctorio's OpenCV-based face detection, found the software failed to detect Black faces 57 percent of the time. The detection rate for white faces was around 60 percent, which means the software failed to detect a white person's face two times out of five. The problem is not only that the model is biased. It is that the model is unreliable for everyone, and the unreliability falls hardest on the students who are already most likely to be doubted.

What a fair process does with a face detection flag

The fix is not a better camera. It is a process that treats a detection failure as a technical event, not as evidence. A low facial detection flag should trigger a check of the recording, the lighting, and the connection log before it triggers a conversation with the student about integrity. The instructor should see the detection percentage alongside the reason it dropped, if the software can supply one, and the student should have a way to contest a flag without the contest itself becoming a second accusation. Programs that use these tools should also test them on their own students before adopting them, because the vendor's accuracy claims are measured on the vendor's data, not on the faces in your lecture hall.

The deeper question is whether face detection should gate an exam at all. A student who is present, answering, and visible to a human proctor on a video call is not cheating because a model lost track of their face for three minutes. The model is a convenience for the institution, and when its convenience becomes the student's problem, the institution has the order backwards. If the AI layer around the exam is the worry, the AI overlay your lockdown browser cannot see covers the tools that run on top of the exam itself, and answer patterns: how statistics catch cheating the camera misses covers the method that works after the camera has done its worst.