Ensuring Academic Integrity through Automated Online Exam Proctoring: A Decade-Long Systematic Review

Synthesis: Malhotra & Chhabra (2026) synthesize 80 peer-reviewed articles (2014–2024) on AI-based automated proctoring systems (AIPS) for online examinations in higher education. They find that advanced machine- and deep-learning techniques (CNNs, RNNs, LSTMs) detect cheating more reliably than traditional methods by analyzing visual cues — eye movements, head posture, facial expressions — yet the field is marked by dataset limitations, limited generalizability, reproducibility gaps, and persistent privacy and fairness concerns. The review advocates integrating IoT and biometric technologies and building hybrid, privacy-preserving, context-aware frameworks.

Key Findings

• ML/DL techniques outperform traditional monitoring. CNNs and RNNs detect cheating by analyzing eye movements, head posture, facial expressions, and body language; RNNs/LSTMs monitor changes over time for detailed behavioral analysis. CNNs offer high visual accuracy but demand extensive training data; RNNs are stronger temporally but prone to vanishing-gradient and scalability issues.
• Four core AIPS features. Effective systems combine (1) authentication (e.g., camera face verification), (2) browsing tolerance/restrictions, (3) remote authorization and control (start/stop/resume exams, flag live suspicious activity), and (4) report generation from recorded exam sessions.
• Systematic quality-assessment gaps. Across the 80 studies: 35% did not fully disclose their dataset; 40% evaluated only a single model; 30% could not be fully reproduced; only 25% explicitly addressed ethical issues; 20% did not report standard metrics (precision, recall, F1, specificity, AUC).
• False positives/negatives undermine trust. Systems may flag normal behavior (looking away, adjusting posture) as suspicious while missing subtle cheating — a key reliability and acceptance barrier.
• Privacy and equity are persistent barriers. Continuous audiovisual surveillance, facial/voice/gaze/keystroke data, GDPR/PDP-Bill compliance, device dependency, and unstable internet disproportionately disadvantage rural and low-bandwidth students.
• Recommended directions. Hybrid AI models combining deep learning with rule-based logic to cut false positives, diverse geographically-inclusive datasets, privacy-preserving architectures (edge processing, anonymization, on-device handling), lightweight models for low-resource environments, and Multimodal AI behavioral input integration.

Study Design & Method

A systematic review following inclusion/exclusion criteria, synthesizing 80 peer-reviewed articles published 2014–2024 on AI-based proctoring systems in higher education. It applied a dual approach: quantitative evaluation of model performance (metrics such as precision, recall, F1, specificity, sensitivity, AUC) and thematic mapping of application domains, plus a quality assessment across technical competency, dataset description, ethical clarity, and methodological completeness.

What this means for practice

• Administrators. Deploy automated proctoring only where detection accuracy is balanced against Privacy, fairness, and student Trust: it can preserve the integrity of online summative assessment where in-person invigilation is unfeasible, but unvalidated, single-model systems risk false accusations and inequitable outcomes.
• Administrators. Require hybrid, privacy-preserving, context-aware design coupled with transparent consent and regulatory compliance before rollout, rather than adopting a vendor system on accuracy claims alone.
• Institutions. Plan for the equity cost of surveillance: continuous audiovisual monitoring, facial and gaze data, device dependency, and unstable internet disproportionately disadvantage rural and low-bandwidth students.
• Assessment professionals. Judge a proctoring system by its false-positive and false-negative behavior on your own student population, since flagging normal behavior such as looking away can penalize students without any misconduct.

Limitations

• Some included studies rely on small-scale or simulated datasets, which limits the real-world generalizability of the accuracy figures they report.
• The review synthesizes 80 peer-reviewed articles published 2014–2024 under fixed inclusion/exclusion criteria, so non-peer-reviewed and more recent work is excluded.
• The rapid evolution of AI methods means newer or hybrid frameworks may already surpass the performance of the systems reviewed.
• Ethical evaluation remains underexplored in the included literature, particularly long-term data storage and enforcement of user consent.

Citation

Malhotra, M., & Chhabra, I. (2026). Ensuring academic integrity through automated online exam proctoring: a decade long systematic review. Discover Education, 5, 207.