What is AI proctoring?
AI proctoring (you’ll also hear automated proctoring or AI based proctoring) uses software models to monitor a candidate during an online exam and raise flags for review. The software watches the webcam feed, the screen, audio and device events, then compares what it sees against a set of known patterns.
In plain terms, the AI proctoring meaning is that a machine watches and a person judges. The AI has no idea whether a candidate cheated. What it knows is that a pattern happened, like a face leaving the frame or a remote-access tool starting up, so it marks that moment for a person to look at in context.
What is an AI proctored exam?
An AI proctored exam is an online exam where AI monitoring runs for the whole session. For the candidate, it usually goes like this:
- They open the exam link and pass a device check.
- Their identity is checked, often with a face match and a photo ID.
- The exam opens in a locked browser session.
- The software records the session and marks any events that match known risks.
- Flagged moments go to a reviewer, either during the exam or after it.
Candidates should know what’s captured before they begin. Proctorly tells each candidate what is recorded before the exam starts, and none of the monitoring is hidden.
AI proctoring, record and review, and live proctoring
These terms overlap, so it’s worth pulling the three models apart.
| Model | What happens | Who judges a flag |
|---|---|---|
| Fully automated online proctoring | AI monitors the session and produces flags or a score, with no human step | Often nobody, or the instructor reads a report |
| Record and review (human reviewed AI proctoring) | AI flags moments in a recording, and a trained reviewer checks each one | A reviewer records a decision, and the institution decides the outcome |
| Live proctoring | A person watches in real time with AI support and can step in | The proctor and reviewer, with the institution deciding the outcome |
Most institutions use more than one model over a year, with lighter monitoring for quizzes and closer watching for end-semester and licensing exams. Our live proctoring page covers real-time watching in detail.
How does AI proctoring work?
AI remote proctoring has five stages. Software carries most of the load in the first four, and people take over in the fifth.
1. Identity checks
Every session begins by confirming who is sitting the exam. Proctorly matches the candidate’s face at the start and keeps matching it throughout, and it rejects photo or video replays. You set photo ID capture for each exam, and high-stakes sittings can add a short active liveness challenge. A failed match prompts a person to check. It never fails the candidate automatically.
2. A locked exam session
The exam opens in a locked browser session, and the system records events such as leaving the exam window, opening new tabs or windows and connecting extra monitors. Because the exam runs inside Proctorly, the attempt, the answers and the evidence all stay in one place.
3. Monitoring signals
AI proctoring software watches several channels at the same time. Proctorly looks for:
- AI assistants and chatbots. Known generative-AI tools, taken from a list we maintain.
- Second devices. Phones and other devices the camera can see.
- Remote control and screen sharing. Remote-assistance tools, virtual machines and browser extensions.
- Overlays and agentic browsers. Known overlay apps and AI browsers.
- Screen recording and extra displays. Recording tools and additional monitors.
- Unusual audio and movement. These go to a person as signals to review in context. They are never treated as findings.
- The computer itself. When an exam calls for it, an optional Windows desktop agent checks for remote-access and recording tools, overlays, extra or mirrored displays, USB devices, Bluetooth earphones, virtual machines, VPNs, extra browser windows and open office apps. It records that an app is open and never reads the candidate’s documents.
On-screen reminders can prompt a candidate to stay in view or follow a rule before anything gets flagged. Each reminder is logged.
If you want detail on a single threat, such as whether proctored exams detect phones, read our guide on what proctored exams can detect.
4. Flags and a review priority score
When a signal fires, the software creates a flag with a clip or snapshot and places it on the session timeline. Proctorly also gives each candidate a review priority score, which lists the signals behind it and is labeled as a priority for review. The score is never a finding and never sets a penalty.
5. Human review and the institution’s decision
A trained reviewer opens the case at the flagged moment and sees webcam, screen and events on one timeline. They decide whether the flag shows an incident and write down why. Confirmed incidents then go to your institution, which makes the finding through its own process. This step is what makes AI flags into results you can defend. Our human review proctoring page explains how it works.
What AI proctoring does not do
Knowing the limits is as important as knowing the features. Responsible AI proctoring should not:
- Decide outcomes. An automated flag should never fail a candidate or impose a sanction. In Proctorly, a reviewer records a disposition and the institution makes the finding.
- Infer emotion, personality or intent. Proctorly doesn’t score a candidate’s emotions, personality or “suspicion” from their face or voice.
- Watch in secret. Candidates learn what is captured before the exam starts.
- See everything. A camera and a screen recording can’t see every way a candidate might get help. No proctoring product catches all cheating, and any vendor claiming otherwise is overstating what software can do. Proctorly publishes the principles of what it can and can’t observe on what we can and cannot see, and we plan to publish a full coverage summary for each setup.
- Treat a technical failure as cheating. A dropped camera or a failed upload is a technical problem, and Proctorly records the interruption as a technical gap. We plan to add a separate evidence category that keeps technical gaps out of the integrity score in every report.
Why fully automated proctoring is not enough
Software flags patterns, and plenty of patterns have innocent causes. Some candidates read questions aloud. In a shared home, someone may walk past. A candidate might be using permitted assistive software, the lighting might change, or a disability might affect how a person moves or where they look. None of this is misconduct, yet a model can read any of it as a pattern.
Treating an automated flag as a verdict causes real harm. Honest candidates face accusations they can’t easily answer, and faculty lose hours on reports they can’t interpret. Appeal panels end up with a score that has no reasoning behind it.
Human reviewed AI proctoring puts things in the right order. The AI cuts hours of recording down to the moments worth checking. A person looks at each of those moments and writes down why it is or isn’t an incident, and the institution then decides with the clips, timeline and reasoning in front of it. You waste fewer hours, and the record holds up when someone asks, “Who decided, and on what evidence?”
Fairness, privacy and accessibility
Candidates, faculty and regulators all have fair questions about AI proctoring, and buyers should expect straight answers.
- Fairness. Proctorly tests detection threat by threat, counts false referrals and checks results for fairness across devices, networks, lighting and accommodations. Fairness monitoring is in place. Proctorly does not publish accuracy figures. You can read more about proctoring bias and how we test for it.
- Privacy. Data is hosted in the client’s geographic location by default, and institutions can keep data in their own S3 bucket. Your institution decides how long recordings are kept. Candidate recordings and data are not used to train AI models unless the institution opts in.
- Accessibility and accommodations. We plan to let you record extra time, breaks, permitted assistive software and reading aloud as accommodations that stop the flags they would otherwise trigger, and to offer an identity route without face biometrics. For now, agree each candidate’s accommodations with our team when the exam is set up, and talk to us about any candidate who can’t use face verification.
How to evaluate AI proctoring software
Take these eight questions into any vendor conversation. They apply to Proctorly and to every other supplier you’re considering.
- Who reviews every flag? Is it your staff, the vendor’s reviewers or nobody at all?
- Can every score be explained? Can a reviewer see the signals behind each one?
- Which threats are covered? Ask for the list, and check it includes AI tools, second devices, remote access and virtual machines.
- What can the setup not see? A vendor ought to be able to tell you.
- How are technical failures recorded? They should never be counted as misconduct.
- What does an appeal panel receive? Look for clips, timeline, reasoning and an audit trail in a single export.
- Where is data stored, and for how long? Find out who sets retention and whether you can choose the region.
- Can you set the level per exam? A weekly quiz and a licensing exam call for different levels of care.
Our guide to choosing online proctoring software has a wider buyer checklist covering identity, LMS fit and service models.
How Proctorly approaches AI proctoring
Proctorly follows one rule: AI flags, people decide. For each exam, you pick an assurance level:
- Recorded. You get an identity check, recording and automated flags, and your staff review cases in the reviewer console.
- Every flag reviewed. A trained Proctorly reviewer checks every automated flag and records a decision. You only see reviewed cases.
- Live pop-in. This adds a proctor watching a group, who steps in by chat or audio when a signal fires.
- Live dedicated. A proctor watches a small group for the whole exam, with a live check-in and identity confirmation.
At the Every flag reviewed level, reviewers work 24×7 during exams, in English, with one reviewer per 40 candidates. They review 95 of every 100 flags within 5 minutes, and the remainder are still reviewed, only later. Each case keeps its clips, timeline, the reviewer’s reasoning and a tamper-evident audit trail, and you can export it as one evidence pack for an appeal or audit.
Frequently asked questions
What is AI proctoring?
AI proctoring is software that monitors a candidate during an online exam. It watches the webcam, screen, audio and device for known risk patterns (a second face, a phone in view or a known AI tool, for example) and raises flags with clips for review. The AI doesn’t decide whether anyone cheated. In Proctorly, a trained reviewer checks each flag and the institution makes the final decision.
How does AI proctoring work?
There are five stages. First the candidate’s identity is checked. The exam then opens in a locked browser session while the software monitors signals from the camera, screen and device. If a signal fires, the software creates a flag with evidence and a review priority. A person reviews each flag, records a reason and passes confirmed incidents to the institution, which decides the outcome through its own process.
Can AI proctoring fail a student on its own?
It shouldn’t, and in Proctorly it can’t: no automated flag fails a candidate or imposes a sanction. The review priority score lists the signals behind it and is labeled as a priority for review rather than a finding. A trained reviewer records a decision with a reason, and any finding of misconduct comes from your institution. If a case is opened, candidates should ask their institution about its appeal process.
Can AI proctoring detect ChatGPT and other AI tools?
Proctorly detects known generative-AI tools from a maintained list. It also picks up known overlay apps, AI browsers, remote-access tools and second devices in view. Since no software sees every way of getting help, detection is just one layer, working alongside exam design, identity checks, a locked session and human review of every flag. Proctorly publishes the principles of what it can and can’t observe, and a full coverage summary for each setup is planned.
Is AI proctoring fair to candidates?
That depends on how flags are handled. Reading aloud, household noise, assistive tools or disability-related movement can all trigger a pattern, so a person has to review each flag in context. Proctorly tests detection threat by threat, counts false referrals and monitors fairness across devices, networks, lighting and accommodations. It never infers emotion or personality, and a flag is never a finding on its own.