# How should exam programs handle AI assistants a candidate can run on a second screen? | ExamFence

Source: https://examfence.com/blog/2026-10-29-ai-assistants-second-screen-exams.html

[Home](../index.md) / [Blog](index.md)


October 29, 2026 - ExamFence

**Short answer:** You cannot confiscate a screen you cannot see. The workable response is layered: a camera view that covers the workspace, items written so a pasted question returns a plausible but wrong answer, and response time patterns reviewed before any accusation is made. No single layer holds on its own.

## The tool changed, the channel did not

Second screens are an old problem. Candidates have read from phones, notes under keyboards, and a helpful person in the next chair for as long as exams have run remotely. What AI changed is the quality of what the second screen says. A photographed question now gets a fluent, mostly right answer in seconds, for subjects where a friend would have needed to be an expert.

Lockdown browsers still do their job on the exam machine. They were never built to govern the room, and treating a lockdown report as proof of a clean sitting is the first error programs make with this threat.

## Camera coverage is a policy decision before it is a technical one

A side or room view that shows the workspace, the candidate hands, and the space beside the keyboard makes second screen use visible rather than invisible. It also collects far more of a candidate home than a face cam does, which is why the decision belongs in policy, with notice, consent wording, and retention limits settled before the first candidate is asked.

Programs that bolt a room scan on quietly, then argue about the footage later, end up defending the surveillance instead of the result. Programs that publish what is captured and why, and keep it only as long as appeals run, rarely have that fight.

## Write items the assistant gets wrong

An assistant answers the question it is given, from patterns in its training. Items that hinge on a case detail buried mid stem, a locally defined term from the course material, or a chart the candidate must read off the screen degrade that answer fast. The correct response stops being the generic one.

This is not an argument for trick questions. It is the older craft of writing items that test application over recall, given a new reason. Recall items were always the weakest measure in the bank. Now they are also the easiest to outsource, which settles the priority for the next item writing workshop.

Response data helps too. A candidate whose answers arrive in uniform blocks of reading time followed by instant selection, on items of very different difficulty, is showing a rhythm worth a human review. Worth a review, not worth an automatic fail. The review step is what separates a defensible programme from a dragnet.

## Say the rule in words a candidate can repeat back

Policies that ban unauthorised aids in legal language get tested by candidates who genuinely did not understand that the assistant on their phone counted. State it plainly in the pre exam briefing: no AI tools, no second device, no help, on any screen. A rule stated plainly is easier to enforce fairly, and fairness is the asset the whole programme trades on.

Keep reading: [Remote proctoring flags: reviewing alerts before accusing a candidate](https://examfence.com/blog/2026-10-29-remote-proctoring-flags-reviewing-alerts.md), [Exam rescheduling rules: fair cutoffs that reduce no shows](https://examfence.com/blog/2026-10-29-exam-rescheduling-rules-fair-cutoffs.md), and [Exam window rules: how long a sitting should stay open](https://examfence.com/blog/2026-10-27-exam-window-rules-sitting-open.md) cover the nearby cases in this series.

## Frequently asked questions

### Can proctoring software detect an AI assistant on a second device?

Not directly. The assistant runs outside the exam machine, so detection is indirect: camera views of the workspace, gaze and head movement patterns, and answer timing that suggests reading. Each signal is suggestive, which is why human review sits between the flag and any decision.

### Are lockdown browsers useless against AI assistance?

No. They still close the exam machine itself: other applications, browsing, and screen sharing on that device. They simply do not govern the rest of the room. Treating them as one layer among several is the correct expectation.

### How can item design reduce the value of an AI answer?

By testing application to specific material: course defined terms, case details that change the answer, data the candidate must interpret from the item itself. Assistants answer generic recall well. They stumble when the right answer depends on the local and the specific.

### What should happen when a candidate is flagged for AI use?

A documented human review of the recording and response data, a chance for the candidate to respond, and a decision under the published policy and appeals route. Automatic sanctions from a software flag alone will not survive a serious challenge, and they should not.

## See your own numbers.

A free bot-traffic audit shows the human-automated split in your live traffic - no code changes, no commitment.

[Get a free bot-traffic audit](mailto:hello@examfence.com?subject=Free%20free%20bot-traffic%20audit)