What AI cheating looks like in exams
Most AI cheating in exams falls into four forms. We describe each one only in terms of what it is.
A chatbot in another window or app
This is the simplest version. The candidate pastes a question into an AI chatbot running in another tab, window or app on the same computer, then copies the answer back. Most searches for “chatgpt cheating” are about this pattern.
Overlays and AI browsers
Some newer software sits on top of the exam window. Other tools are whole browsers built around an AI agent that can read the page and suggest or type answers. Both are designed to be harder to spot than a separate window.
AI on a second device
Once the exam computer is locked down, the AI moves to a phone or tablet, and the candidate photographs or types the question into it. At that point you’re dealing with a second-device problem more than a software one.
AI-written work in take-home and open-book tasks
Away from timed exams, AI can draft essays, reports and answers that a candidate hands in as their own work. It’s the hardest form to detect after the fact, and it’s where policy and design count for the most.
Can teachers detect ChatGPT?
Sometimes, though the text alone won’t tell you reliably. AI text detectors estimate the likelihood that something was machine-written. They can’t prove who wrote it, and they make mistakes in both directions. Many institutions treat a detector score as one signal among several, since on its own it is a weak basis for a misconduct finding.
Faculty often pick up on AI-written work in other ways. The answers may not match a candidate’s earlier work, they may contain confident errors, or the style may feel out of place for the course. Any of these is a reason to ask questions (a short follow-up conversation about the work, say). None of them is a finding by itself.
A proctored exam changes the question. You’re no longer asking who wrote the text. You’re asking whether an AI tool was running during the sitting, and software signals can help answer that.
AI cheating detection methods during a proctored exam
Chatgpt detection in online exams works best when it watches the session rather than the answers. This is what Proctorly does.
Detection of known AI tools
Proctorly checks for known generative-AI tools from a list we maintain. If one is running during the exam, the system raises a flag and attaches the evidence.
Overlays, agentic browsers and the desktop agent
Proctorly also detects known overlay apps and AI browsers. The exam runs in a locked browser session that records tab and window events. When an exam needs deeper checks, the Windows desktop agent looks at running apps, screen overlays, extra browser windows, tabs and extensions. It records that an app is open and never reads the candidate’s documents.
Second devices
Once the AI is on a phone, the camera does the detecting. Proctorly flags phones and other devices it sees in the camera’s view, and repeated glances down can raise a behavior signal for review. To understand what a camera can and can’t see, read can proctored exams detect phones.
Exam design that makes AI less useful
Detection catches AI use, and good design makes it less worth attempting. Proctorly’s online exam software gives you these controls:
- Randomized delivery. Questions and options are shuffled from item pools, so candidates don’t all see the same question text.
- Section timers and locked navigation. With less time on each question, there’s less time to consult a tool.
- Adaptive testing. Every candidate takes a different path through the paper.
- Questions that need working. Use numeric answers with tolerance, short answers and essays. Candidates can also photograph handwritten working on a phone and attach it to the question for the marker.
- Record-only sittings for practice. Low-stakes practice runs raise no flags, so candidates can get used to the format without pressure.
Many institutions add a short oral follow-up for high-stakes written work, in which the candidate talks through their answer. That’s a policy choice for your institution. Proctorly doesn’t provide it as a feature.
Writing an academic integrity AI policy
Your rules matter as much as your tools. A good academic integrity AI policy answers these questions in plain language:
- What is allowed. For example, students may use AI to study before the exam but not during it.
- What is banned, and where. Spell it out for each assessment type: timed exams, take-home exams and coursework.
- How permitted use is disclosed. If coursework allows AI, explain how students declare that they used it.
- What is monitored. Tell candidates before the exam what gets captured and checked.
- What happens after a flag. Say who reviews it, who decides and how a candidate can appeal.
- Accommodations. State clearly that using a permitted assistive tool is not misconduct.
Publish the policy before exam season, link to it from your exam instructions and brief faculty on what it says. Our guide to academic integrity covers the wider framework.
What AI cheating detection software cannot do
If an AI cheating detection tool claims to catch everything, be skeptical. No detection software can do the following:
- See a device outside the camera’s frame. The computer has no view of AI running on a hidden phone.
- Prove authorship from text alone. A detector score is only an estimate.
- Keep up with every new tool on day one. Detection lists are maintained, so new tools get added over time.
- Decide intent. Proctorly doesn’t infer emotion, personality or “suspicion” from a candidate’s face or voice, and an automated flag never fails anyone.
We’re open about these limits. A camera and a screen recording can’t see every way a candidate might get help. We publish the principles behind what each setup can and can’t observe, and a full coverage summary is planned. For what the Trust center shows today, see what we can and cannot see.
How human review handles AI flags
Detecting an AI tool is a strong signal, but it is still only a signal. The candidate may have left an app open before the exam started, the browser may come with a built-in assistant, or a permitted assistive tool may look like something else.
In Proctorly, a trained reviewer opens the case at the flagged moment, sees webcam, screen and events on one timeline, and records a decision along with a required reason. Confirmed incidents go to your committee through the malpractice workflow, and no penalty is applied before that. For any appeal, the evidence pack holds the clips, the timeline, the reviewer’s reasoning and a tamper-evident audit trail. This is how human review proctoring turns an AI flag into a fair decision.
A note for students reading this: the software is built to detect AI tools running during a proctored exam, and a person reviews every flag. If your coursework allows AI tools, follow your institution’s disclosure rules. And if you’re unsure what’s allowed, ask before the exam rather than after it.
Frequently asked questions
Can teachers detect ChatGPT in written work?
Not reliably from the text alone. AI text detectors produce estimates rather than proof, and they can be wrong either way. Teachers often spot work that doesn’t fit a student’s earlier submissions and follow up with questions. A proctored exam works differently: the session is checked for AI tools running at the time, and a person reviews any flag before a decision is made.
Can proctoring detect ChatGPT during an online exam?
Proctorly detects known generative-AI tools from a maintained list, plus known overlay apps and AI browsers, and the Windows desktop agent checks running apps. The computer can’t see AI used on a phone outside the camera’s view. That’s why second-device detection, exam design and human review all have a part to play. No tool catches every case.
What are the best AI cheating detection methods?
The approach that works best combines detection, design and policy. While the exam runs, look for AI tools, overlays, AI browsers and second devices. In the paper itself, use randomized item pools, timers and questions that need working. Your policy should say clearly which AI use is allowed and how students disclose it. Every flag then goes to a trained person before the institution decides anything.
Should we ban AI completely?
Your institution decides that. Many take a mixed approach, banning AI in timed exams and allowing it with disclosure in some coursework. Whatever you choose, the rule for each assessment type should be clear, shared with students in advance and enforced through a fair review process. In Proctorly you set the rules and the assurance level for each exam.
Does an AI-tool flag mean the student cheated?
No. A flag means the system detected an AI tool or overlay during the session, and there may be an innocent explanation, such as an app left open or a permitted assistive tool. In Proctorly, a trained reviewer examines the evidence and records a decision with a reason. Only your institution decides the outcome, after its own committee process.