glossaryPage.heroH1

Four main concerns per NEPC and AACE research. Privacy: webcam-and-microphone-based proctoring captures student-home environments, family members, religious-practice items, and personal information that institutions hold as exam-integrity evidence. Equity: AI-flagged suspicious-behaviour systems disproportionately flag students with disabilities (eye movement, fidgeting, head-position), students of colour (skin-tone bias in webcam face-detection), neurodivergent students, and students in non-traditional home-test environments. False positives: AI-content detection tools have published false-positive rates of 5-25% in independent research, meaning a substantial portion of flagged work was not actually AI-generated. Equipment access: bandwidth and webcam quality vary across students; technical proctoring failures disproportionately affect students with weaker home internet. Institutions weighing deployment should publish their policy, provide alternative-assessment options, and review per-demographic-group performance regularly.

Unreliable in practice. Per independent research from Stanford HAI (Human-Centered AI Institute) and others, AI-content detection tools have false-positive rates of 5-25% (work flagged as AI-generated that was not) and false-negative rates of 30-60% (AI-generated work not flagged). The reliability has not improved as much as some early vendor claims suggested because AI-generated text has become harder to distinguish from human writing as models improve. Per UNESCO 2024 guidance: institutions should not rely on AI-content detection as a primary academic-integrity tool. The more durable response is assessment design (in-class work, oral defence, portfolio with process-evidence) rather than detection.

Per UNESCO 2024 AI in Education guidance and AACE peer-reviewed research, the most resilient approach is assessment design rather than detection. Specific patterns: (1) Move high-stakes assessment to in-class formats (timed in-class essays, oral defence, lab practical, portfolio-presentation-and-questioning) where AI assistance is not available. (2) Use process-evidence in take-home work (draft history, research-log, source-annotation) so the work demonstrates the student's thinking process not just the final output. (3) Adopt explicit AI-use policies with student and faculty input — some institutions permit AI with disclosure, some restrict by assignment type, some integrate AI use into the curriculum. (4) Where proctoring is used, deploy with explicit privacy policy, accommodations workflow, and per-demographic-group equity review.

Plagiarism detection (Turnitin, SafeAssign, Unicheck) compares submitted text against a database of prior submissions, published academic work, and internet content, producing a similarity report. It detects text that matches existing sources — the long-standing academic-integrity concern of copying without citation. It does not detect AI-generated text per se (AI-generated text is novel, not copied). AI-content detection is a separate category attempting to identify AI-generated text by linguistic features. Turnitin and others have added AI-content-detection features to their core plagiarism tools, with mixed independent-evaluation results. The two tools address different academic-integrity concerns; institutions typically deploy plagiarism detection (mature, well-understood) and approach AI-content detection more cautiously (experimental, unreliable).

Per UNESCO 2024 guidance, NEPC research, and AACE peer-reviewed work: (1) Publish the institutional policy to students and faculty explicitly — what tools deploy, what data captures, what review process applies. (2) Provide accommodation workflow for students with disabilities, religious-practice requirements, or non-traditional test environments. (3) Review per-demographic-group flagging rates regularly and adjust if disparities surface. (4) Provide student-appeal workflow with human reviewer (not AI) for any academic-integrity finding. (5) Retain teacher / instructor primacy in academic-integrity decisions — the AI flag is advisory; the human decides. (6) Periodically re-evaluate whether the deployment is worth the privacy, equity, and false-positive costs versus alternative-assessment-design approaches.