Start from the honest premise: you cannot prevent cheating on an unproctored remote exam. A student with a second device, a group chat, and an AI assistant can defeat every technical control available to a normal instructor. Webcam proctoring doesn’t close the gap either — it’s expensive, legally awkward under the Data Privacy Act, unreliable on Philippine home connections, and demonstrably beatable.
That premise sounds defeatist. It isn’t. It just relocates the problem from surveillance to assessment design — and design is the part you actually control.
The reframe: stop asking “how do I stop them from looking things up?” and start asking “what can I ask that looking it up doesn’t answer?” The second question has good answers. The first doesn’t.
Why the surveillance approach fails
Worth being specific, because a lot of money gets spent here:
- Lockdown browsers control one device. They see nothing of the phone beside the laptop.
- Webcam proctoring produces flags a human must review, at volumes nobody reviews. False positives punish students with poor lighting, shared rooms, or unstable connections — which is a large share of Filipino students.
- Connection instability turns a technical fault into an integrity accusation, and you cannot tell them apart from the log.
- Data privacy. Recording a student’s home and requiring room scans is personal data processing. It needs a lawful basis, a retention policy, and a consent path — a real institutional decision, not an instructor’s toggle.
- AI assistance answers a recall question instantly, invisibly, on a second screen.
None of that means online exams are worthless. It means the recall-based exam is the format that broke.
Design tactic 1 — make the exam open-book on purpose
If students will have their notes regardless, declaring it open-book costs you nothing and buys you two things: the anxiety-driven cheating disappears, and you’re now obliged to write items that a notes search doesn’t answer.
That obligation is the actual benefit. An open-book exam forces items at Applying and above — which is where your course outcomes probably sit anyway. Our Bloom’s taxonomy stems are the practical starting point.
| Closed-book item (searchable) | Open-book item (not) |
|---|---|
| Define construct validity. | Here is a study’s instrument and method. Identify the strongest threat to its construct validity and justify your choice. |
| List the assumptions of linear regression. | This residual plot violates one assumption. Name it and state what it does to the standard errors. |
| What is the formula for KR-20? | This 20-item exam returned KR-20 = 0.58. Give two plausible causes and how you’d distinguish them. |
The right-hand items survive a Google search, a textbook, and an AI assistant giving a generic answer, because each requires judgement applied to this stimulus.
Design tactic 2 — give everyone a different exam
Two mechanisms, and the difference matters:
Shuffling reorders items and options within the same exam. It stops shoulder-surfing and casual “what did you put for #4” coordination. It’s a five-second setting and you should always turn it on.
Item pools draw each student’s exam from a larger bank — 60 items in the pool, 40 delivered. Now two students genuinely have different exams, and answer sharing collapses.
The catch is fairness: if the pool’s items vary in difficulty, students face unequal exams. The fix is to pool within a competency and cognitive level — item 12 always draws from “Applying, hypothesis testing,” never from the whole bank. That’s the same discipline as building parallel forms, applied per item.
Design tactic 3 — use time deliberately
Time pressure is the one control that genuinely constrains an unproctored student, because looking something up costs seconds.
But blunt time limits punish slow readers and weak connections more than they punish cheating. Two refinements:
- Per-item timers rather than one exam-wide limit. A student can’t bank time on easy items to research the hard ones, and nobody loses the whole exam to one stuck question.
- Set the limit from a real timing. Take the exam yourself and multiply by about three. Guessing at “one minute per item” produces a speed test.
Use time as friction, not as a weapon. An exam nobody can finish measures typing speed.
Design tactic 4 — assess what can’t be transferred
Some assessments are structurally resistant, because the answer is specific to the student:
- Tasks built on the student’s own data — their fieldwork, their dataset, their chosen case.
- Reflective analysis tied to something only that section experienced.
- Iterative work where drafts are submitted along the way, so the final product has a visible history.
- Brief oral defence of a submitted piece — five minutes on video call for a random sample of students changes the incentives for everyone, without proctoring anyone.
These don’t scale to a 200-student service course. For a 35-student major course they’re the strongest option available, and rubric-scored — see our rubric guide.
Detection after the fact
You won’t catch it live. You can notice it afterwards, and the signals are ordinary item analysis output read with a different question in mind:
- An item where difficulty jumps far above the rest of the exam — especially an item you expected to be hard. An answer that circulated shows up as an anomalously easy item.
- Negative or near-zero point-biserial on a hard item. Weak students getting a hard item right while strong students miss it is either a miskey or shared answers. Rule out the miskey first — it’s usually the miskey.
- Identical wrong answers on open-ended items, particularly identical phrasing. Wrong answers converging is far more diagnostic than right answers converging.
- Submission timestamps clustering in a way the item timing can’t explain.
Treat every one of these as a question, not a finding. Each has an innocent explanation, and acting on a statistical pattern alone is how institutions lose appeals. What the anomalies are genuinely good for is telling you which items to retire.
Publish the rules before the exam, not after
Most academic-integrity disputes are really disputes about what was permitted, and they’re won or lost in the syllabus. State plainly, in writing, before the first assessment:
- What resources are allowed — notes, textbook, calculator, the internet, AI assistants. Say it explicitly for each. “Open-book but no collaboration” is clear; silence is not.
- What collaboration means for this course, with an example of what crosses the line.
- The AI policy, specifically. Prohibited, permitted with citation, or permitted freely — students genuinely don’t know unless you say, and a blanket ban you can’t enforce is worse than a clear permission.
- The consequence, referencing your student handbook.
- The technical contingency — what a student does when their connection drops, and what you’ll do about it. Publishing this removes the incentive to fake a disconnection, because the honest path is already defined.
An exam where the rules were clear and generous is easier to defend than one that relied on a control that failed.
Frequently asked questions
Should I use webcam proctoring at all?
For high-stakes summative assessment where your institution has adopted it, with a privacy notice and a documented lawful basis, it has a role. As an individual instructor’s tool for a course exam, the cost, the false-positive burden, and the Data Privacy Act exposure usually outweigh what it buys.
Isn’t open-book just making the exam easier?
Only if the items stay at recall. Written properly, open-book exams are consistently harder, because “look it up” no longer produces the answer. Expect your class mean to drop the first time, and warn students in advance.
How do I handle students with unstable internet?
Design for it: per-item timers rather than a hard exam clock, answers saved as the student goes, and a published re-entry procedure. Then treat disconnections as an infrastructure problem rather than an integrity one — that’s both fairer and, in Philippine conditions, usually true.
Does randomizing items make grading unfair?
Not if items are pooled within the same competency and cognitive level. It becomes unfair when a pool mixes an easy recall item with a hard application item at the same position.
Can I tell if a student used AI on an essay?
Not reliably — AI-detection tools produce false positives at rates that make them unsafe for grading decisions, and they misfire disproportionately on non-native English writers. Design around it instead: require drafts, require course-specific and section-specific content, and state your AI policy clearly.
Is a short online quiz worth the setup?
Usually more than a big exam is. Frequent low-stakes quizzes reduce the payoff from cheating on any single one, give you far more data per outcome, and are where automated checking saves the most time.
Give the exam online, get the analysis automatically
Skoolari administers assessments through a student portal — QR-code access, shuffled items, per-item timers, and automatic scoring straight into your class record. Afterwards you get the full item report: difficulty, discrimination, distractor pull, and KR-20 for the whole exam, so the anomalies described above are visible without a spreadsheet. Building the items is covered too — see turning a PDF module into a quiz.