Sign up free

False positives in online proctoring: what the studies found

What studies, audits and reported cases found about honest students flagged by proctoring software, why the base rate matters, and how to read a flag.

The EdProcto editorial teamWritten and checked against the research it cites. Last checked . How the guides are made

Guides · Integrity and cheating · · 13 min read

How often does automated online proctoring flag honest students, and how should an instructor read a flag?

False positives in automated online proctoring are common wherever anyone has counted. In California's remotely proctored October 2020 bar exam, 3,190 of 8,920 candidates were flagged for review and 47 violation notices had been affirmed by January 2021. Vendors rarely publish error rates. Where cheating is uncommon, most flags land on honest students, so treat a flag as a reason to look, never a finding.

A false positive in online proctoring is a flag on a student who did nothing wrong. Where flags have been counted and then reviewed, many of them were that, and in the largest count nearly all.

This guide is for the lecturer who reads the flags and the administrator who buys the software. It covers what sets a flag off, what audits and reported cases counted, what vendors publish, and the arithmetic that makes most flags false when cheating is uncommon.

EdProcto publishes this guide; its own integrity settings come at the end. Each source was opened on 10 October 2026; where only an abstract could be read, the text says so. Counts of flags from single courses are in our guides to webcam proctoring and how much proctoring a quiz needs.

What is a false positive in online proctoring?

A false positive is a flag raised when no rule was broken, and it comes in two kinds.

  • The software misread the video. Respondus (n.d.) gives the plain case on its own help page: a student flagged as missing while the face is plain to see in the picture. Reducing such flags, it says, is "not a perfect science".
  • The software read the video correctly, and the behaviour was honest. The student did look away; a voice was heard. Coghlan et al. (2021) point out that these systems claim to detect behaviours, not misconduct, and that glancing around while thinking is quite normal.

Better cameras and models shrink the first kind. Nothing shrinks the second, because a student who is cheating and a student who is thinking can make the same movement.

What sets off a flag on an honest student?

Ordinary things: poor light, a look away, a restless body, a noisy home, a weak connection. The table sets each kind of flag beside an honest cause.

Kind of flag Honest cause Documented by
Face missing or not detected Backlighting, very low light, a cap, a hand on the face, reflective glasses The vendor's own list (Respondus, n.d.)
Eyes off the screen Thinking: eyes move about twice as often during a search of memory Ehrlichman and Micic (2012)
Movement Restlessness, as in adults with ADHD Teicher et al. (2012)
Noise A dog barking; other people in a shared home Meazure Learning (2021); Malhotra and Chhabra (2026)
Interrupted video A dropped or slow connection Respondus (n.d.)
A browser measure, such as "Leaving the tab" in EdProcto A tapped notification, a screen that locked Our guide to a fair process

Faces and light

Face detection fails more often in poor light, and one reported test found it failing more often on some faces. Respondus (n.d.) tells instructors that detection suffers from a cap, light behind the student, a very dim room, a hand on the face or glasses that reflect, and its list of flags includes repeated bad-lighting alerts.

Feathers (2021), reporting for Vice, described a test by a student software researcher. The researcher found that the face-detection files in Proctorio's browser extension had the same names as models published by the open-source library OpenCV, and performed identically to them. On nearly 11,000 faces from the FairFace image set, those models missed 57% of faces labelled Black and 40% of faces labelled white. A security researcher consulted by the reporter reproduced the analysis; the company did not tell the reporter which models it uses.

Proctorio disputes the test. Its chief executive writes that the image set holds no webcam pictures and "is not appropriate for the algorithm and its intended application", and that external auditors "found no measurable bias" in its face-detection models (Olsen, 2023).

An official body has examined one case. A Dutch student told the Netherlands Institute for Human Rights that the same software regularly answered "face not found" or "room too dark", so that she sat exams with a bright lamp shining in her face. Its final ruling found no discrimination in her case: her longer check-in and one restart had other causes, such as a poor internet connection or wearing glasses. It added that a university has a duty to check that such software works without discrimination before using it (College voor de Rechten van de Mens, 2023). More of the ruling is in remote membership exams.

Eyes, movement and neurodivergence

Looking away and moving about are what thinking looks like. Ehrlichman and Micic (2012) review evidence that people move their eyes about twice as often when searching long-term memory as in tasks that need no such search, even in the dark (abstract read).

Teicher et al. (2012) tracked 40 adults with ADHD and 60 without through a 20-minute computerised attention task. Those with ADHD moved their heads through 2.2 times the area. A rule that counts head movement would count ADHD: our inference, not a test of any product.

Coghlan et al. (2021) make the general point: people who are disabled or neurodivergent may not behave as a model of "normal" exam behaviour expects, and may be flagged for it.

Noise, shared homes and poor connections

A microphone hears the household, and a video stream depends on the line. ProctorU's own analysis found that proctoring by software alone "could flag a test session inappropriately" if a student repeatedly rubbed their eyes or a dog barked (Meazure Learning, 2021). A review of 80 papers on such systems calls microphone false alarms common in shared or noisy living spaces (Malhotra and Chhabra, 2026).

In a Jisc survey of 21,697 UK higher education students in late 2020, 62% reported a poor wifi connection and 19% had no safe or private space to study (Knight, 2021).

How often are proctoring flags wrong when someone counts?

Often, and in the largest count nearly always. Neither count below is a general rate.

  • A bar exam. In California's remotely proctored bar exam of October 2020, 8,920 people completed the exam and 3,190 had their sittings flagged for review (Ward, 2021): by the proctoring software, on the Electronic Frontier Foundation's account of the bar's meeting (Kelley, 2020). The bar sent 432 violation notices. By 8 January 2021 it had affirmed 47, on a standard of clear and convincing evidence, and six were pending. So at most 53 of 3,190 flags, under 2%, ended in a finding or a pending one. Lawyers for candidates said notices cited eyes out of the webcam's view and audio not working. The software's maker told the journal that it never sets the rules on what counts as a violation: the jurisdiction does.
  • Fourteen courses. Tweissi et al. (2022) set one automatic proctoring tool against five human proctors over 244 exam attempts at a Middle Eastern university. The tool's average decision on violations was 35.61% against the humans' 25.95%, a significant difference, and the tool decided wrongly in 74 attempts, about 30% (abstract read; it does not say how many of the 74 were false alarms).

Many flags are never read at all. ProctorU found that schools and testing bodies reviewed only about 11% of the sessions its software had tagged, and it stopped selling proctoring without a human proctor (Meazure Learning, 2021). An unread flag harms nobody until someone treats the count itself as evidence.

Do vendors publish their error rates?

Rarely. Coghlan et al. (2021) write that proctoring companies' websites "rarely if ever cite rigorous studies" to justify their accuracy claims and settle concerns about false positives.

We read three vendors' pages on the day of writing.

  • Respondus (n.d.) defines a false positive and lists what causes one. It gives no rate.
  • Olsen (2023) says Proctorio has had external audits since 2019 and makes the reports available to customers through its trust centre. The post itself prints no figure.
  • Meazure Learning (2021) published the share of flagged sessions that were reviewed, and nothing on how many flags were wrong.

These are single pages. A rate may sit in a document we did not see, so ask.

What research prototypes report

Figures from small collections. Among the 80 papers Malhotra and Chhabra (2026) reviewed, one video model had a true positive rate of 0.72 and a false positive rate of 0.31, and one accuracy of 100% rested on 50 reference images. Most systems, they write, are trained on narrow data, and performance drops in real rooms with varied lighting.

What to ask before you buy

  1. The false flag rate for each kind of flag, and who measured it.
  2. Who it was measured on: which skin tones, disabilities, rooms and connections.
  3. The share of flagged sessions that customers reviewed, and the share they upheld.
  4. Whether each kind of flag can be switched off by itself.

Choosing a quiz tool for a school has more questions for any vendor.

Why are most flags false when cheating is uncommon?

Because honest students far outnumber cheating ones, so a small error rate on the many outweighs a high catch rate on the few. Forgetting this is the base rate fallacy.

An illustration, on figures we chose and no study measured: 1,000 sittings, and software that flags four in five of the cheating ones.

Cheating sittings Honest sittings flagged Cheating sittings flagged All flags Flags on honest students
10 (1%) 99 (10% of 990) 8 107 93%
50 (5%) 95 (10% of 950) 40 135 70%
50 (5%) 19 (2% of 950) 40 59 32%
200 (20%) 80 (10% of 800) 160 240 33%

Read the second row. The software sounds good: it catches 80% of cheating and is right about 90% of honest students. Yet seven flags in ten point at someone who did nothing. A flag lifts the chance that a sitting involved cheating from 5% to about 30%: worth a look, and far from a finding.

A 10% false flag rate is kind beside California's bar exam, where over a third of sittings were flagged. And nobody knows the true share of cheating in a given exam: the surveys in what research since 2023 says disagree widely.

The example is set out as counts of sittings on purpose. Hoffrage et al. (2015) found that medical students given such figures as counts of people made correct inferences more often than those given probabilities, by 37 percentage points on average. The same sum for essay detectors is in AI detectors are not evidence.

How should an instructor read a proctoring flag?

As a reason to look, never as a finding. Put five questions to the flag before you put one to the student.

  1. What was measured? A face not found, a sound, a window left. Name the observation, not the suspicion. In EdProcto each measure has a plain name, such as "Leaving the tab" or "Full screen".
  2. What ordinary thing causes it? Go down the first table: light, glasses, thinking, restlessness, the household, the line.
  3. How many others have it? If a third of the room is flagged, the flag describes the tool or the exam. Compare the student with the class, not with zero.
  4. Does anything independent agree? The answers, their timing, a script that matches a neighbour's. A flag with nothing beside it is the second row of the table.
  5. What would settle it? Watch the recording yourself, then ask the student, with the record in front of you both.

Then follow your institution's procedure. Ours is in a fair process when a student is flagged, and the student's side is in your attempt was flagged.

What not to do with a flag

  • Do not add flags into a score. Ten bad-lighting alerts are one dim room.
  • Do not let software take marks off for what a model judged.

Where does EdProcto stand on false flags?

EdProcto is the best tool for a quiz at low or middle stakes when you want few false alarms, because its Light and Medium levels are plain switches in the browser, and the false flags documented above are mostly judgements about faces, eyes, movement and sound.

Among them are "Leaving the tab", "Full screen", "Copy and paste off", "Right-click off", "Keyboard shortcuts off" and "Printing off", each switched on by itself. EdProcto names five integrity levels: None, Light, Medium, Strict and Maximum. Strict and Maximum add the camera, the microphone and the face and gaze measures, and the arithmetic above applies to them as to any vendor's: switch them on only where the stakes need them.

A browser measure can catch an honest student too, since a tapped notification leaves the tab. Put the same five questions to it.

Free carries Light integrity and thirty students a quiz. Tutor carries Medium and 100 students. Educator takes 500 students a quiz and carries Strict; Maximum comes with a licence. See pricing and the page for universities. No setting makes cheating impossible, a flag is never proof, and a score is never a grade by itself.

Quick questions

How accurate is automated online proctoring?

No general figure exists. In one comparison a tool decided wrongly in 74 of 244 attempts (Tweissi et al., 2022), and in California's 2020 bar exam under 2% of 3,190 flags ended in an affirmed or pending notice (Ward, 2021).

Is a proctoring flag evidence of cheating?

It is a reason to look. In our illustration, with cheating in 5% of sittings and a 10% false flag rate, seven flags in ten fall on honest students.

Do proctoring vendors publish false positive rates?

Rarely. Coghlan et al. (2021) write that companies seldom cite rigorous studies, and the three vendor pages we read on 10 October 2026 printed no rate.

Which honest students are most at risk of a flag?

On the evidence above: those in poor light or wearing reflective glasses, those who look away to think, restless students including some with ADHD, and those in noisy homes or on weak connections.

Does EdProcto use a webcam or face detection?

Only where you switch them on. Its Light and Medium measures act in the browser; the camera and the face measures belong to its Strict and Maximum levels, and the faces are read in the student's browser: see integrity.

References

Integrity

Five levels, from none to a full proctoring stack, each measure switched on by itself with its own consequence.

Light on Free, Strict on Educator

See the five levels

Read next