Online Exams: Proctoring and Anti-Cheating Methods

What actually works

No single method fully prevents cheating. What works is layered protection, with strictness matched to the stakes:

• low stakes (self-check, progress quiz): randomization and time limits;
• medium (course credit, internal staff assessment): plus browser lockdown and webcam recording;
• high (entrance or certification exam): plus live or hybrid proctoring and identity checks.

Start with the design of the exam itself: it is cheap and doesn't touch privacy.

Layer 1. Exam design

• A question bank larger than the exam needs; each student gets a random selection.
• Shuffle the order of questions and answer options.
• Parameterized problems: the same task with different numbers for each student.
• Time limits for the whole test or per question, leaving less time to search for answers.
• No going back to earlier questions where that makes sense.
• Open questions and applied tasks instead of definitions that are easy to search.

A minimal example of question selection:

import random

def build_exam(bank, per_topic, seed):
    rng = random.Random(seed)  # seed = student id + attempt id
    exam = []
    for topic, questions in bank.items():
        exam += rng.sample(questions, per_topic[topic])
    rng.shuffle(exam)
    return exam

A fixed seed lets you reproduce the exact variant during an appeal.

Layer 2. Browser lockdown

A lockdown browser or kiosk mode prevents opening other tabs, copying text, taking screenshots and launching other programs. Options:

• a dedicated browser app the student installs before the exam;
• a web mode that tracks focus loss and exits from full screen.

Limitations: the web mode only records events, it doesn't block them. A second device or a hint from someone in the room is not solved by lockdown.

Layer 3. Webcam and proctoring

AI typically tracks: no face in frame, a second face, looking away, other voices, a phone in view, tab switching.

How accurate AI proctoring is

Algorithms make mistakes in both directions:

• false positives: poor lighting, a weak camera, a habit of looking away while thinking, household noise, differences in appearance or health conditions;
• misses: hints from outside the frame, a second monitor, an earpiece.

So an AI flag is a signal to review, not proof. The right process: the algorithm raises a flag, a human watches the clip, and the student can contest the decision.

Privacy and ethics

Proctoring means access to the camera, microphone and sometimes the screen in a student's or employee's home. Before rolling it out:

• tell people in advance what is recorded, who watches and how long recordings are kept;
• get consent and offer an alternative (an in-person exam) where possible;
• collect the minimum, for example don't require a full room scan unless necessary;
• limit retention and access to recordings;
• check compliance with local personal data law and, for students in other countries, their requirements too.

For companies there is also trust: aggressive proctoring on an internal quiz can hurt employee relations more than cheating does.

How to choose the level of protection

1. Define the stakes: what does a person gain by passing?
2. Assess the risk: how easy is it to find answers online?
3. Start with exam design, which covers most of the risk.
4. Add technical measures only when their cost and privacy intrusion are justified.
5. Write down the appeal procedure before the first exam.