AI in Examinations: Institutions Detect Cheating Without Accusing Students?

Artificial intelligence is changing how universities conduct examinations, evaluate student performance, and protect academic integrity. Online proctoring systems can monitor examination sessions, detect unusual activity, compare submitted answers, and flag patterns that may indicate unauthorised assistance. These capabilities can help institutions investigate misconduct, particularly when thousands of students take examinations simultaneously across different locations.

However, the use of AI in examination monitoring introduces a serious challenge: detecting suspicious activity does not necessarily mean proving that a student has cheated. A student who looks away from a screen may be reading a question carefully, someone who briefly leaves the camera frame may be responding to an emergency, and an answer that resembles AI-generated text may reflect ordinary academic writing. If an automated system treats such behaviour as proof of misconduct, innocent students may face unnecessary investigations, damaged reputations, or unfair academic penalties.

Universities therefore need an approach that protects the credibility of examinations without treating algorithmic predictions as established facts. AI should help identify situations that deserve further examination, while trained human reviewers assess the available evidence, consider alternative explanations, and give students a fair opportunity to respond.

The central question is not simply whether AI can detect cheating. It is whether universities can use AI accurately, transparently, and proportionately while protecting students from false accusations. Achieving that balance requires better technology, carefully designed assessments, clear rules, independent review, and reliable disciplinary procedures.

1. How AI Detects Potential Examination Misconduct

AI-powered examination systems use different techniques depending on the type of examination and the suspected misconduct. Online proctoring platforms may analyse webcam footage, microphone input, screen activity, browser behaviour, and examination-session records. Other systems examine submitted answers for unusual similarities, possible plagiarism, or patterns that warrant further investigation.

Computer vision can identify whether a face is visible, detect additional people in a camera frame, or flag repeated departures from the examination screen. Some systems analyse head movements or changes in gaze direction, while others identify attempts to open prohibited applications or switch between browser windows. These signals can help a human reviewer understand what happened during an examination, but they do not automatically establish that a rule was violated.

Text-analysis systems present a different challenge. They may compare student submissions with existing sources, identify similarities between answers, or estimate whether text resembles material generated by an AI model. Yet similarities can arise from legitimate causes, including shared course terminology, standard definitions, common references, or answers to questions with limited possible responses.

Research published in PLOS ONE in 2024 demonstrated that AI-generated submissions could pass through a university assessment process without being identified. The study found that 94% of the deliberately inserted AI-written submissions in its experimental setting went undetected. This result illustrates the limitations of relying on ordinary marking processes to identify unauthorised AI use, although it should not be interpreted as a universal detection rate for every university or examination format.

The practical lesson is that AI detection tools can contribute useful evidence, but no single signal or detector should be treated as a complete account of a student’s conduct.

2. Why False Accusations Can Occur

False accusations can arise when a system interprets ordinary behaviour as suspicious or assigns too much importance to an unreliable indicator. In online examinations, students may look away from their screens to think, read handwritten notes permitted by the examination rules, adjust their seating position, or respond to environmental interruptions. Poor lighting, unstable internet connections, camera placement, and shared living spaces can also affect what a monitoring system records.

A student may be flagged because their face was not visible for several seconds, even though the interruption resulted from a technical problem. Another student may repeatedly look towards the side of the room because their examination setup places a second monitor or necessary accessibility equipment outside the camera’s central field of view. Without context, the system may struggle to distinguish these situations from genuine misconduct.

Text-based detection also carries risks. AI-writing detectors estimate whether a passage resembles patterns associated with machine-generated language. They do not provide direct access to the author’s writing process or prove who created the text. A formal writing style, predictable sentence structure, limited vocabulary, or heavily edited draft can complicate interpretation.

A 2025 study in the British Journal of Educational Technology examined the use of generative AI in higher education assessments and found that experienced markers could struggle to distinguish AI-assisted submissions from other work. The researchers also highlighted the possibility of both false positives and false negatives in academic judgement.

Universities must therefore distinguish between an indicator that justifies checking the evidence and evidence sufficient to support a formal finding of misconduct. Confusing these two stages can turn a useful monitoring tool into an unfair disciplinary mechanism.

3. Why One AI Alert Should Never Prove Cheating

A reliable examination policy should treat an AI alert as a starting point for review, not as a final verdict. The system might indicate that a student repeatedly switched windows, that an additional person appeared in the camera frame, or that two submitted answers are unusually similar. Each observation requires a different kind of verification.

For example, repeated window switching might indicate that a student accessed an unauthorised website. It could also reflect an approved accessibility tool, an examination-platform error, or an application required for the assessment. Similarly, a person briefly appearing in the background does not prove that the student received assistance.

Universities should require corroborating evidence before escalating an allegation. Depending on the situation, this could include reviewing the relevant recording, checking system logs, examining the examination rules, comparing answer similarities in context, and asking the student to explain what happened.

The review should consider whether the evidence supports an innocent explanation. If a system identifies a possible second person in a room, the reviewer should establish whether the person interacted with the student or influenced the examination before drawing conclusions. If an answer is suspected of being AI-generated, the reviewer should examine the permitted-use policy, the nature of the assessment, and any available evidence about the student’s work.

The threshold for opening a preliminary review may reasonably be lower than the threshold for imposing a penalty. A preliminary review asks whether there is enough information to investigate further. A disciplinary finding requires a fair assessment of the evidence under the institution’s applicable rules.

This separation protects students from automatic punishment while still allowing institutions to investigate credible allegations.

4. Combining AI With Human Review

Human oversight is essential because examination misconduct often depends on context that an algorithm cannot reliably interpret. A trained reviewer can consider the question being asked, the examination conditions, the student’s approved accommodations, the sequence of events, and explanations that may not be visible in the raw data.

A practical system would first use AI to identify a specific event and preserve the relevant evidence. Instead of producing a vague label such as “high cheating risk,” the platform should indicate what occurred, when it occurred, and which rule may be relevant. A reviewer could then examine the event in its original context rather than relying on a numerical risk score alone.

The reviewer should be trained to recognise the limits of automated detection and avoid confirmation bias. Once an allegation is raised, there is a risk that every subsequent behaviour will be interpreted as evidence of guilt. Reviewers should instead consider alternative explanations and distinguish direct observations from assumptions.

A second reviewer may be appropriate for serious cases, particularly when the evidence is ambiguous or the possible penalty could affect a student’s progression, scholarship, or graduation. Reviewers should also have a way to challenge the system’s interpretation and document why an alert was dismissed or escalated.

Research into online examination monitoring has explored how human reviewers respond to potentially misleading signals. A 2025 paper presented at the Artificial Intelligence in Measurement and Education Conference found that reviewers rejected many deliberately fabricated alerts, but some were still accepted. The findings suggest that human review is valuable yet not automatically sufficient; consistent guidelines and evidence-based decision-making are also necessary.

The goal should be a system in which AI improves the efficiency of investigation while humans remain responsible for interpreting evidence and making consequential decisions.

5. Detecting AI-Generated Answers Without Misjudging Students

Generative AI has created a new challenge for universities because students can use AI tools to produce essays, code, summaries, and examination responses. Whether such use constitutes misconduct depends on the rules of the particular assessment. Some examinations prohibit external assistance entirely, while other assignments permit AI for brainstorming, editing, coding support, or research when its use is properly acknowledged.

Institutions must therefore define the permitted and prohibited uses of AI before deploying detection systems. A detector cannot determine whether conduct is misconduct without knowing the rules that applied to the assessment.

Text-based AI detection should be treated cautiously. A detector may identify characteristics associated with generated text, but those characteristics are not unique to AI. Students who write in a formal style, use language-learning tools, receive writing support, or revise their work extensively may produce text that a detector misclassifies.

Universities should avoid using a detector score as the sole basis for a misconduct finding. Instead, reviewers can examine the submitted work alongside relevant drafts, version history, citations, task-specific reasoning, and the student’s explanation. Where appropriate, a short oral discussion or supervised follow-up task may help establish whether the student understands the work they submitted.

However, students should not be expected to prove their innocence through invasive access to personal devices or private accounts. Any request for additional evidence must be proportionate, consistent with institutional rules, and respectful of privacy.

Assessment design also matters. Questions that require students to apply concepts to a specific dataset, explain their reasoning, interpret experimental results, or defend a design decision can provide richer evidence of learning than a generic essay alone. These methods are not foolproof, but they can make assessment more meaningful without depending entirely on automated detection.

6. Preventing Bias in Automated Proctoring

AI systems may perform differently across students because of differences in lighting, camera quality, facial appearance, disability, behaviour, and testing environments. These differences matter when automated alerts influence disciplinary decisions.

For example, facial-recognition systems can make errors when matching images, and error rates may vary across demographic groups. The US National Institute of Standards and Technology has documented demographic differences in face-recognition performance, including the effects of image quality and the representation of different groups in training data. These findings concern face-recognition technologies generally and should not be treated as a direct measurement of every examination-proctoring product. They nevertheless demonstrate why institutions must evaluate biometric technologies carefully.

Bias can also arise when monitoring software assumes that every student has the same physical environment. A student living in a crowded household may experience more background movement than someone taking an examination in a private room. Students with certain disabilities may move their heads, use assistive technology, or need breaks in ways that automated systems interpret incorrectly.

Universities should test examination tools using representative users and realistic conditions before relying on them in high-stakes settings. They should examine whether alert rates and error rates differ across relevant groups, including students using approved accommodations. Where the system cannot perform reliably for a particular use case, institutions should consider an alternative method rather than transferring the system’s limitations onto the student.

Accessibility must be built into the examination process from the beginning. Students should know how to request accommodations, how those accommodations will be recorded, and how reviewers will distinguish approved behaviour from possible rule violations.

Fairness is not achieved simply by applying the same algorithm to everyone. It requires examining whether the system produces reliable and proportionate results for the people who must use it.

7. Protecting Student Privacy and Personal Data

Online proctoring can collect highly sensitive information, including video recordings, audio, facial images, room backgrounds, screen activity, identity documents, and behavioural data. Universities must determine whether each category is genuinely necessary for the examination and whether a less intrusive method could achieve the same purpose.

A student taking an examination at home may inadvertently reveal family members, personal belongings, religious objects, or other details of private life. Continuous recording can therefore extend beyond the academic activity that the university intends to supervise.

Institutions should clearly explain what data are collected, why they are required, who can access them, how long they will be retained, and under what circumstances they may be shared. Retention periods should be limited to legitimate operational, review, and legal needs rather than keeping recordings indefinitely.

Security controls should include restricted access, secure storage, documented deletion procedures, and safeguards against unauthorised disclosure. Universities should also assess vendors before purchasing their services and establish clear responsibilities for handling student data.

Research published in the Journal of Computing in Higher Education in 2025 examined stakeholder perceptions of privacy in online proctoring. It identified concerns about sensitive data collection, access, and possible misuse, and emphasised the importance of involving students and educators in decisions about these systems.

In India, institutions should assess their data practices against applicable privacy law and relevant regulatory requirements. Transparency and data minimisation should be treated as essential elements of responsible examination design, not optional additions after the technology has been purchased.

8. Designing a Fair Investigation and Appeals Process

Even a carefully designed AI system will occasionally generate an incorrect alert. Universities therefore need a formal process through which students can understand allegations, respond to evidence, and challenge decisions.

A student should be informed of the specific conduct under investigation and the rule that may have been breached. Where disclosure is permitted, the institution should provide enough information about the relevant evidence for the student to respond meaningfully. A vague statement that an AI system detected suspicious behaviour is not an adequate explanation of the allegation.

The student should have a reasonable opportunity to provide context, submit relevant evidence, identify technical problems, or explain approved tools and accommodations. Investigators should consider that response impartially and document the reasons for their conclusion.

Serious penalties should be decided through the institution’s established disciplinary procedure, not imposed automatically by the software or by a single unreviewed alert. The process should include a clear route for appeal to an appropriate authority that can reconsider the evidence and the fairness of the original decision.

Penalties should be proportionate to the proven conduct and consistent with published regulations. A technical irregularity, an ambiguous monitoring event, and deliberate impersonation are not equivalent situations and should not be treated as though they were.

The emotional impact of an accusation also deserves attention. A misconduct allegation can threaten a student’s academic progress and reputation even if it is eventually dismissed. Universities should therefore protect confidentiality, avoid unnecessary public disclosure, and communicate decisions respectfully.

A fair process does not weaken academic integrity. It strengthens the legitimacy of disciplinary decisions by ensuring that genuine misconduct can be addressed without sacrificing the rights of innocent students.

9. Rethinking Examination Design in the Age of AI

Monitoring technology should not be the university’s only response to examination misconduct. Institutions can reduce opportunities for cheating by designing assessments that measure understanding, reasoning, and application rather than relying exclusively on easily reproduced answers.

In suitable subjects, universities can use case-based questions, practical demonstrations, supervised problem-solving, oral examinations, laboratory work, and project reviews. These approaches allow educators to observe how students apply knowledge and explain decisions. Their suitability depends on the discipline, class size, accessibility requirements, and available resources.

Universities should also distinguish between assessments in which AI use is prohibited and those in which responsible AI use forms part of the learning objective. In some courses, students may be expected to use AI tools critically, verify their outputs, identify errors, and explain how they reached a final result. Clear rules help prevent students from being accused of conduct that the institution has never adequately defined.

The assessment process can also preserve evidence of learning through staged submissions, feedback records, supervised components, and brief explanations of key decisions. Such evidence can help instructors understand a student’s progress without requiring intrusive surveillance throughout every examination.

UNESCO’s guidance on generative AI in education and research advocates a human-centred approach that considers privacy, institutional preparedness, and the educational purpose of the technology.

The most effective strategy combines thoughtful assessment design with proportionate monitoring. Technology can assist with identifying irregularities, but strong assessment practices reduce the need to infer misconduct from ambiguous behaviour alone.

10. A Responsible Framework for Universities

Universities should adopt AI examination systems only after defining their purpose, evaluating their reliability, and establishing safeguards for students. The process should begin with a clear policy explaining what counts as misconduct, which forms of AI assistance are permitted, and how automated alerts will be reviewed.

Before deployment, institutions should test the system in realistic examination conditions and examine the frequency of false alerts. They should also review accessibility, privacy, security, cost, and the quality of evidence the platform produces. If a vendor cannot explain how its alerts are generated or provide sufficient information for meaningful review, the university should question whether the tool is suitable for high-stakes decisions.

During examinations, AI should record specific events and identify potential concerns without declaring students guilty. After the examination, trained reviewers should examine relevant evidence, consider alternative explanations, and escalate only cases that justify further investigation.

Institutions should periodically audit outcomes to identify patterns of mistaken alerts, inconsistent decisions, and possible bias. Students and faculty should have channels to report problems with the technology, and policy changes should reflect the findings of these reviews.

Accountability must remain with the institution. A university cannot transfer responsibility for a disciplinary decision to an external software provider simply because an algorithm generated the original alert. The institution must be able to explain and defend its decisions under its own academic rules.

Conclusion: AI Should Support Fairness, Not Replace Judgement

Artificial intelligence can help universities detect potential examination misconduct by identifying unusual activity, examining submitted work, and organising evidence for review. These capabilities can support academic integrity, especially in large online examinations where manual monitoring alone may be difficult.

However, automated systems can misinterpret normal behaviour, struggle to identify unauthorised AI use reliably, and produce errors influenced by technical conditions or demographic differences. An alert is therefore not proof of cheating, and a detector score should never replace a fair investigation.

Universities can reduce the risk of wrongful accusations by combining AI with trained human review, clear examination policies, privacy safeguards, accessible assessment arrangements, and meaningful appeal procedures. They should also improve assessment design so that students demonstrate understanding through reasoning, application, and supervised work where appropriate.

The objective is not to eliminate technology from examination monitoring. It is to use it responsibly. Academic integrity requires institutions to address genuine misconduct, but justice requires them to distinguish evidence from suspicion. When AI supports rather than replaces human judgement, universities can protect the credibility of examinations while respecting the dignity, rights, and educational opportunities of every student.

Online Internship with Certificate