Proctoring bias: where it comes from, and how we keep it out of decisions

Where proctoring bias comes from

Bias in proctoring rarely has a single cause. It usually builds up from four sources.

Facial recognition bias

Face detection and face matching systems have been shown to perform differently across demographic groups. Independent research, including a 2019 evaluation by the US National Institute of Standards and Technology (NIST), found that many face recognition algorithms had higher error rates for some groups than for others, and that the differences between algorithms were large. In practice, facial recognition bias can mean one candidate’s face goes undetected or unmatched more often than another’s.

In proctoring, that can lead to more “face not visible” or “identity mismatch” flags for some candidates, through no fault of their own.

Conditions that are not the candidate’s fault

Low light, a dim or backlit room, an old webcam, a slow connection or a shared household all affect what a camera and microphone pick up. These conditions aren’t spread evenly. Candidates with less money, less space or weaker internet are more likely to face them.

Behavior that looks unusual but is not misconduct

Some candidates read questions aloud, look away to think, move because of a disability, use assistive software or sit in a noisy home. Any rule that treats these as suspicious will flag some groups far more than others.

Automation bias

People tend to trust what a machine tells them, especially when they are busy. A reviewer who receives a flag labeled “suspicious” may see what the label tells them to see, and the bias in the system becomes bias in the decision.

Our rule: AI flags, people decide

The best protection against proctoring bias is to make sure no algorithm can decide an outcome. In Proctorly:

• No automated flag fails a candidate or imposes a sanction. A trained reviewer looks at each flag and records a decision with a reason, and your institution makes the finding.
• A failed identity check goes to a person, never to an automatic fail. A face match is a signal, not a verdict.
• Confirmed incidents go to a recorded committee decision, and no penalty is applied before it.
• Unusual audio and movement are raised as signals for a person to review in context, never as findings. Reading aloud, assistive tools, household noise and disability-related movement are not misconduct.
• Gentle reminders come first. Before anything is flagged, on-screen nudges remind a candidate of the rules, such as staying in view. Every nudge is logged.

Our human review proctoring page explains how reviewers work and what an appeal panel receives.

How we test and monitor fairness

We measure our detection threat by threat instead of with one overall number. For each threat, we count false referrals, meaning cases where a flag sent an innocent candidate to review. We then check the results for fairness across devices, networks, lighting and accommodations, so we can see whether candidates on older laptops, on weaker connections, in darker rooms or with accommodations are flagged more often. Fairness monitoring is in place.

Why we publish no accuracy figure

You won’t find an accuracy percentage on this page. A single figure hides which threats it covers, which candidates it was tested on and how many innocent people it flagged. Until we can publish per-threat results with the test set and its date, we publish no figure at all. We would rather explain how we measure than give you a number you can’t check.

Scores are review priorities, not verdicts

Each candidate’s review priority score shows the signals behind it and is labeled a priority for review, never a finding. Reviewers can see why a case was ranked where it was, and they can disagree with the ranking. A high score means “look at this first”. It does not mean “this candidate cheated”.

What we will not build

• No emotion, personality or “suspicion” scoring of candidates, ever.
• No inference of intent from a candidate’s face or voice.
• No claim to catch all cheating. We publish what each setup can and cannot observe. See what proctoring can and cannot see.

More safeguards in the review process

• Five evidence labels (Observed, Suspected, Blocked, Unobservable, Unavailable) are planned, so that a technical failure stays in its own category and out of the integrity score.
• A plan for candidates who cannot use face verification. If some of your candidates can’t use face verification, talk to us and we will plan the exam with you.
• Accommodations that travel with the candidate and stop the flags they would otherwise cause are planned. See proctoring accommodations.
• A notice to the candidate when a case is opened, and a reconsideration step, are planned.
• Blind double review to check how consistent reviewers are is planned.
• An AI case summary that helps the reviewer, and never decides, is planned.

What institutions can do

Technology is only part of the answer, and your policies matter just as much.

1. Put in writing that a flag is a question and not evidence of misconduct, and brief every reviewer and committee member on it.
2. Set up accommodations before exam day, so candidates aren’t flagged for what you have already agreed.
3. Give candidates guidance on lighting and setup, and a way to report technical problems straight away.
4. Where your law allows, track flag and outcome rates by exam and setting, and look into any pattern you find.
5. Make sure every candidate can see the evidence and appeal a decision.

Frequently asked questions