{"id":27705,"date":"2026-10-01T10:13:43","date_gmt":"2026-10-01T14:13:43","guid":{"rendered":"https:\/\/uxpertise.ca\/pourquoi-vous-ne-pouvez-pas-vous-fier-uniquement-a-lia-pour-la-surveillance-des-examens-en-ligne\/"},"modified":"2026-10-01T10:16:12","modified_gmt":"2026-10-01T14:16:12","slug":"why-you-cannot-rely-on-ai-alone-for-online-proctoring","status":"publish","type":"post","link":"https:\/\/uxpertise.ca\/en\/why-you-cannot-rely-on-ai-alone-for-online-proctoring\/","title":{"rendered":"Why You Cannot Rely on AI Alone for Online Proctoring"},"content":{"rendered":"
A candidate is taking an online exam at home when their young child briefly walks into the webcam\u2019s view. The proctoring system detects another person and flags the moment. <\/p>\n
Was the candidate cheating?<\/p>\n
A trained human reviewer can examine what happened: a child wandered into the room, no assistance was exchanged, and the candidate continued the exam. With that context and the exam\u2019s rules, the reviewer can distinguish an accidental interruption from a violation. <\/p>\n
AI can flag unusual activity. That does not make the activity proof of cheating. <\/strong><\/p>\n Automation helps monitor exams at scale and identify moments that deserve attention. The challenge comes when those signals are treated as conclusions. <\/p>\n Looking away, speaking aloud or briefly disappearing from view can have several explanations. A candidate might be thinking through a question, using an approved accommodation or adjusting their webcam. Deciding whether a rule was broken requires more than recognizing a pattern. <\/p>\n Technical limitations can also affect candidates differently. A 2022 University of Louisville study <\/a>involving 357 students found that the evaluated proctoring system detected faces for an average of 78% of exam time for students with darker skin tones, compared with 92% for those with lighter skin tones. The study examined one system using 2020 exam data, so these figures are not a current industry-wide error rate. They illustrate why automated outputs need scrutiny. <\/p>\n Unreviewed alerts also create work for administrators, who must investigate ambiguous events and explain them to candidates.<\/p>\n A hybrid model combines automated detection with trained human review.<\/p>\n Automation identifies potential concerns. Reviewers examine the available evidence, what happened before and after the event, and the applicable exam rules. Approved accommodations and permitted resources should inform that assessment. <\/p>\n The reviewer can then dismiss an unsupported concern or document evidence of a violation. This contextual evaluation is central to the human review process used by Integrity Advocate, the online proctoring solution integrated into uxpertise LMS and uxpertise XP. <\/p>\n Human review is not infallible. Its value depends on training, consistent standards and sufficient evidence. Those safeguards help turn an alert into a considered finding. <\/p>\n If a candidate challenges an integrity decision, an automated risk score provides limited answers.<\/p>\n A defensible decision is one the organization can explain: what occurred, which rule applied, what evidence was reviewed and why the conclusion was reached. Documented human review provides a basis for responding to questions and reconsidering a finding when appropriate. This connection between evidence and reasoning is central to Integrity Advocate\u2019s approach to defensible outcomes. <\/p>\n That finding concerns the integrity of the exam session. It is separate from the candidate\u2019s academic score. <\/p>\nWhere AI-only proctoring falls short<\/h2>\n
What human review adds<\/h2>\n
Reliable results need an explanation<\/h2>\n
A Canadian example: National Payroll Institute<\/h2>\n