how-to
Secure Remote Exam Question Design: What It Can and Can't Stop
Question design is the cheapest exam security control you own. What it stops (lookup, AI chat, item sharing), what it can't, and what to layer with it.


Question design is the cheapest security control a remote exam has, and the only one no vendor can sell you. A well-built item makes lookup slow, makes sharing pointless and makes a chat window a poor shortcut. It also has a ceiling. Some attacks do not care what the question says, and an exam owner should know which.
This is a guide to getting the most out of item design, and to knowing where it stops.
What question design can and can't stop
Match each attack to what it needs from the question. Item design works against the attacks that need the question to be easy.
| Attack | What it needs | Does item design help? |
|---|---|---|
| Searching the question text | An answer that exists in public | Yes. Items built on supplied data and justification have nothing to find. |
| Pasting the question into a chat assistant | A question a model can answer from its text alone | Partly. It raises the effort, and it shrinks as models improve. |
| Harvesting and sharing items | The same items shown to many candidates | Yes. Bank depth, exposure caps and rotation remove the payoff. |
| An overlay that reads the screen | Whatever is on screen, however it is worded | Barely. If the full item is visible, so is the data it depends on. |
| A remote operator answering for the candidate | A live connection to the machine | No. A competent person answers whatever you ask. |
| A hidden camera or helper off-device | A view of the screen | Slightly. It changes how hard an item is to solve from a photograph. |
The last three rows are the point. Question design cannot protect an exam from a purchased expert, because the expert is the answer to a hard question. Those attacks need other layers, and the final section covers them.
A workflow for building items that resist copying
Most banks fill up with recall items that leak within one exam cycle, because teams write the question first. Work in three steps instead.
- Start from the objective. Write it as an observable action: "calculates a confidence interval from raw data", not "understands statistics". Then ask what evidence would prove it. An item that asks for a definition proves recall, not the skill.
- Choose a format that is expensive to copy. Recall and definition items are the easiest to leak and search. Scenario items, calculations with randomized inputs and short constructed responses are harder to share, because the method transfers and the answer does not.
- Pilot every item under real conditions. Run it on a small group the way the live exam runs. An item everyone answers correctly tells you nothing about skill, and one everyone fails usually has a wording problem. Remote conditions change how people read a stem, so a question that works on paper can fail on screen.
Banks, versions and exposure
Randomizing the order of five fixed questions does almost nothing. Drawing each candidate's paper from a deeper bank does a lot.
- Write sibling items. Same objective, same cognitive demand, different surface details. A calculation with different numbers is a sibling. Two candidates who see different items should face the same challenge.
- Cap exposure per item. Count how many candidates have seen each one, and rotate it out at the cap.
- Retire on a schedule, not when a candidate reports seeing it online. Keep retired items out of live draws. They still work as practice material and pilot comparisons.
- Log every exposure so you can trace a leak to a sitting.
- Set time limits from pilot data. Take the median completion time and add a buffer for slower readers and approved accommodations. A narrow exam window also limits how long an item is exposed.
Writing items a model answers badly
Banning tools does not work in a remote exam, because you cannot enforce the ban from the item. What you can do is make the default answer wrong, incomplete or slow to verify:
- Reasoning chains. The answer needs several steps, each depending on the last.
- Justification. The candidate must explain why, not just choose a letter.
- Conflicting evidence. Two plausible options, and the candidate weighs the trade-offs.
- Authentic tasks. A case analysis on a supplied dataset, a short incident triage that asks for the next three actions ranked with reasons, or a policy application where the candidate says which rule applies and where it runs out. Each needs a rubric with defined levels, or it turns into subjective grading and appeals.
Be honest about the shelf life. A capable model can answer most items whose full text is on the screen, and what resists one today may not next term. Plan on item design raising the cost of AI help, not removing it. Try your own items against the strongest assistant you can reach, on a schedule, and write the next version with what you learn.
One question at a time, and what it costs
Showing one question at a time with no back navigation slows harvesting. A candidate cannot capture the whole paper in one pass or return to copy a stem. It does not stop it: each screen can still be photographed or read by an overlay. The cost is candidate stress, which affects scores. Test both modes with your population, and say plainly in the instructions that navigation is locked, because surprise restrictions produce disputes, not security.
What to layer with it
Item design answers one question: what does this exam give away? Two more questions need other controls.
- Who has a live path to the machine? Remote operators and AI services need a connection. Aiseptor applies a default-deny network policy for the session, so only the assessment platform and resources you allow resolve. Remote-access tools are cut off from the network rather than terminated, and AI endpoints are blocked at the DNS and SNI level. It detects the screen-capture-exclusion flag an overlay must set to stay hidden, and the GPU memory changes a locally loaded model leaves behind. There is no kernel driver, webcam or microphone. The services that sell passes mostly need that connection, which is why it matters.
- Who is in the room, and what is in it? Cameras, earpieces, a helper out of frame and a stand-in for the candidate are outside what any software on the device can see. Invigilation, room checks and identity verification cover them. Aiseptor does not.
Run all three: items that give little away, a session with no path out, and a room you can trust. For the wider programme, securing high-stakes remote exams covers the process, and how to prevent AI cheating in online assessments covers the controls in more depth.
Frequently Asked Questions
How do you design secure questions for a remote exam?
Start from an observable learning objective, pick a format that is expensive to copy (scenarios, calculations with randomized inputs, short constructed responses), and pilot each item under real conditions before it enters the live bank. Then draw each candidate's paper from a bank of equivalent sibling items, cap how often each item is shown, and retire items on a schedule.
Can question design stop AI cheating?
It raises the cost but does not remove it. Items that need reasoning chains, written justification or conflicting evidence are harder for a model to answer well, but a capable assistant can answer most items whose full text is on screen, and that improves over time. Test your own items against the strongest assistant you can reach, regularly.
How do you create multiple versions of a remote exam?
Build a bank with more items per objective than one sitting needs, tag each item by objective and difficulty, and sample one item per tag for each paper. Keep versions equivalent by matching difficulty and blueprint coverage rather than reordering the same questions. Cap exposure per item, retire exposed items, and review response patterns after each sitting.
Does good question design replace proctoring or network controls?
No. It limits what an exam gives away, but it cannot stop a remote operator answering live, an overlay reading the screen, or a helper in the room. Those need a session with no path out for remote tools and AI services, and room and identity controls for what software on the device cannot see.