How to stop ChatGPT use during online exams

Every online exam now comes with an uninvited assistant one shortcut away. Whether you teach, chair a department or handle academic integrity cases, the question is no longer whether students can reach ChatGPT, Copilot or Gemini during an exam. It is through which door, and what closes each one.

The pressure peaks at midterms and finals, when online exams multiply and there are not enough rooms or proctors to bring everyone on campus. It shows up as answers that are fluent but generic, the same structure across half a group, and integrity cases that rest on a hunch. This article maps where AI enters an exam, why the usual safeguards fall short on their own, and how to layer controls that hold.

Where does generative AI show up during an online exam?

There isn't one door. In an online exam taken on a student's own computer, there are at least five.

• Browser tabs. ChatGPT, Copilot, Gemini and the many websites built on the same models are one new tab away. This is the most obvious route and the easiest to block.
• Desktop applications. The major assistants also come as apps for Windows and macOS. They run outside the browser, so a browser setting never sees them.
• AI built into the operating system and office tools. Recent versions of Windows and macOS include assistants and AI writing tools, and office suites now offer AI help inside the very document a student is typing in.
• Browser extensions. Sidebar assistants and "answer this" extensions can read the page the exam is on and suggest answers in place, inside the same browser as the exam.
• A second device. A phone or tablet on the desk, or a laptop beside the screen: photograph the question, read the answer.

The last channel deserves a straight answer. No software installed on the exam computer can block a phone it doesn't run on, however strict the lockdown. What addresses the phone is a camera that records the student, a person who watches live or reviews afterwards, and an exam whose questions a photographed prompt answers poorly. A camera doesn't see everything either: it makes the phone visible, and good exam design makes it less useful.

What does unchecked AI use cost?

The cost is not a count of incidents. It is what happens to the exam as a measure.

• Grades stop meaning what they say. A pass tells the student, the next teacher and the program that a skill was shown. If a tool showed it, the mark misleads all three.
• Honest students pay for it. Where marks are compared, or admission to the next program depends on them, every undetected shortcut tilts the scale against the students who followed the rules.
• Integrity cases rest on suspicion. Without evidence, a teacher either lets it go or opens a case based on how an answer "sounds". That is hard on the student and hard to defend on appeal.
• Exams retreat to paper. Some departments respond by bringing every exam back to a supervised room, with the rooms, proctors and printing that requires, and less flexibility for students with accommodations or who study at a distance.

What doesn't work on its own

Most institutions already use at least one of these. Each one helps; none of them is a control by itself.

• Honor statements. An integrity pledge at the start of the exam sets expectations and makes the rule explicit, which matters if a case goes further. It does not stop a determined student, and it leaves the teacher nothing to review.
• Blocking a list of URLs in one browser. A browser policy or network filter that blocks the best-known AI sites is a reasonable start. It misses the next new site, the desktop apps, the assistant built into the operating system, extensions and every other browser on the computer. And a student at home is not on your network at all.
• AI-text detectors applied afterwards. Running answers through a detector after the exam sounds like a safety net. In practice, detectors estimate a probability rather than show what happened. They produce false positives on honest work, light rewording can get past them, and they have little to say about short answers, calculations or code. A score is not evidence a student can examine and answer, which makes it a weak basis for an integrity decision.

The common flaw is relying on one measure to cover every channel. Generative AI reaches the exam through several, so the controls have to be layered.

How to stop AI use during online exams: a layered method

Treat it like any other security problem: list the channels, put a control on each, and make sure a person looks at what the controls catch. The method is vendor-neutral. The controls can come from your exam design, a lockdown tool, a proctoring application or people.

1. Design the exam so AI helps less

Ask students to apply what they learned to a case, a data set or material from your own course, and to show their reasoning, not only the answer. Consider an open-book format where it makes sense: when notes are allowed, you can ask harder questions and the incentive to hide things drops. Plan a short oral follow-up for a sample of students, or for any answer that raises questions. A student who wrote an answer can explain it.

2. Lock down sites, applications and extensions

Use a lockdown that controls the computer, not one browser. Safe Exam Browser, which is free and open source, does this for browser-based exams; proctoring applications do it alongside recording. Block by default and allow by exception, so a new AI site is covered before anyone has heard of it.

3. Monitor processes, focus and windows

A block list cannot anticipate every tool. Watching what runs on the computer, and when the exam window loses focus or the screen changes, catches what the list missed and tells the reviewer exactly which minute to look at.

4. Record, so the second device is visible

Camera recording, or a proctor watching live, is the only layer that reaches the phone on the desk. Record what the exam needs and no more, and tell students in advance what is captured.

5. Have a person review what is flagged

A detection is a lead, not a verdict. The reviewer checks the timestamp, watches that part of the recording and decides, following the institution's academic integrity policy. Keep the evidence with the file in case of appeal.

What if the teacher wants to allow an AI tool?

Not every exam should ban AI. A programming course may teach with a coding assistant; a writing course may assess how well students check and revise an AI draft. In those exams, using the tool well is part of the skill being measured.

That only works if the rule is set exam by exam:

1. Write the rule into the instructions. Name the tool that is allowed and what it may be used for.
2. Configure exactly that tool. Allow the one AI tool and keep every other site, app and extension blocked.
3. Tell students before exam day. No one should discover the rule in the middle of the exam.
4. Keep recording and review on. Allowing one tool does not mean dropping the other layers.

A single institution-wide switch fails both ways: it either bans a tool a course has adopted or opens every exam to everything. Accommodations follow the same logic. Assistive technologies authorized for a student, such as a screen reader or writing-support software, must keep working under the lockdown, as the accessibility page explains.

How Examo handles AI during exams

Examo is a native Windows and macOS application that frames and records an exam, whatever tool the teacher built it with: a Moodle quiz, a Word, Excel or PDF document, or an external exam application. It treats AI tools like any other unauthorized resource. AI websites such as ChatGPT, Copilot and Gemini, installed AI applications and browser extensions are blocked. Examo locks down the exam environment, applications and services included, and monitors the processes running on the computer for the whole exam.

Loss of focus on the exam window and window or screen changes are detected. Each detection is timestamped on the exam's event timeline, an alert lets a proctor react during the exam, and the teacher receives the event with time-coded evidence. Continuous recording covers the student's camera and microphone, every connected screen, and keyboard and mouse, and live proctoring lets a proctor watch camera and screen streams in real time. That is how a phone on the desk becomes visible: a person sees it, live or in the recording. Examo does not analyze the video.

Blocking is set per exam or per security profile, one of the steps described in how Examo works, so a teacher can allow a specific AI tool when the exam permits it while everything else stays blocked. Assistive technologies authorized for a student are not blocked. In Moodle, a "Require the Examo app" option in the quiz settings brings all of this to a quiz, as explained in how to proctor a Moodle quiz remotely.

Every incident is submitted to human review: no suspicion score, no automatic decision or sanction. After the exam, an Excel report of events and alerts is exported with the matching video and audio recordings. Examo itself contains no AI feature, with no facial recognition, no automated behavior analysis and no biometric data, for the reasons set out in remote proctoring without facial recognition.

The full feature set is on the proctoring and lockdown page, and the FAQ answers common questions about AI tools during exams. Examo is built in Montréal by LS2 Innovation, whose other software for higher education is used by about fifty colleges and universities.