Online Exam Proctoring AI: Widely Deployed, Thinly Validated, Increasingly Disputed

The use case

On 17 February 2021, a chemistry student at Cleveland State University in Ohio was asked to show his bedroom to his webcam before a remote exam. The scan lasted less than a minute, and as little as ten to twenty seconds¹. In August 2022 a federal court held that the university's practice of conducting room scans was unreasonable under the Fourth Amendment¹. The case is a useful entry point because it shows what online exam proctoring consists of in practice, and how little of it is artificial intelligence (AI) alone.

The tools named in the ruling were Respondus and Honorlock. The court described Respondus LockDown Browser as blocking internet access, Respondus Monitor as recording the exam and using AI to flag activity, and Honorlock as using the camera and AI monitoring¹. One vendor describes the flagging simply: if no face, or more than one face, is detected, the AI flags the incident and a human proctor may intervene (vendor-reported)¹³.

Scope note. Here, AI means the machine-learning parts of these products: the models that detect faces and behaviour in webcam video and decide what to flag. The lockdown browser is a blocking tool, not a flagging model¹. Identity checks, room scans and the review of flags are done by people, either live proctors or the institution's own staff¹⁸¹³. Detection of AI-written text and plagiarism checking are out of scope.

Where it is running

The University of Erfurt in Germany also ran remote exams with photo proctoring and a lockdown browser before stopping them⁷. An older deployment exists in Canada: the University of British Columbia piloted Proctorio in four distance courses in 2018¹⁷. No current independent count of institutions using machine-learning proctoring was found, and every deployment with evidence is in the United States, the Netherlands, Germany or Canada. The Proctorio figure is cumulative since 2013, so it measures volume over time, not current use.

What the evidence shows

Detection of cheating. In a 2021 controlled experiment with 30 students, none of the six who were told to cheat were flagged by Proctorio. An independent human check caught one of the six, and the authors put the system's sensitivity at very close to zero (measured in a controlled experiment, April 2021)⁴. The students were computer science students whose technical competence was far above average, and the sample was small⁴. The result is older than three years and no newer independent detection test was found, so it may not describe current versions. A plausible reason is that flags are tied to observable events such as a missing face or a second face¹³, and cheating that triggers no such event passes unflagged; the paper's abstract does not state a mechanism. No study measuring cheating rates before and after deployment was found.

False alarms. The 2018 pilot at the University of British Columbia found that automated flagging produced many false alarms, which instructors described as "fidgeting not cheating", and no instructor found evidence of cheating in the pilot courses (interviews with 4 instructors and 88 student survey responses, 2018, no control group)¹⁷.

Uneven effects. A study of about 357 students at one US institution found that students with darker skin tones and Black students were significantly more likely to be marked as needing instructor review; women with the darkest skin tones were far more likely to be flagged, and there was no significant male-female difference overall (measured without a control group, 2022, abstract)⁵. In the Netherlands, a student complained that Proctorio did not work for her because of her skin colour. The Netherlands Institute for Human Rights cleared Vrije Universiteit Amsterdam of discrimination but said it should have handled the complaint better, and added that the judgment does not rule out discrimination from such software in other situations (October 2023, as reported)¹⁰.

Human review. An experiment by Duolingo researchers planted fake AI copy-typing alerts in 170 real sessions and gave them to proctors. Proctors rejected 50% to 71% of the fake alerts, depending on the guidelines used, and 9% to 13% of genuine ones (model-implied rates, operator-authored experiment, October 2025)⁶. Under the first set of guidelines, fake alerts were rejected for 68% of Western test-takers but 18% of Chinese test-takers⁶. The authors name confirmation bias and nationality differences as findings⁶. The mechanism is that reviewers who see an AI signal tend to accept it, so the person in the loop does not fully correct the model.

Review is also often not done. In 2021 ProctorU reported that only about 11% of test sessions tagged as suspicious were reviewed by faculty or administrators, and cited a University of Iowa audit in which faculty reviewed 14% of flagged sessions (operator-reported, 2021). It added that reviewing one 60-minute exam for 150 students can take nine hours⁸.

What it takes: cost, integration, governance

No price was found on an official vendor page, so cost per student or per exam is not reported here. Human time is a visible cost: the nine-hour review figure above is operator-reported⁸. One vendor sets a default retention period of 365 days for webcam video, which can include desktop activity and audio if the institution requests it (vendor-reported)¹³.

Legal basis. In July 2021 the Amsterdam Court of Appeal found, in summary proceedings, that the University of Amsterdam's use of Proctorio was in line with the General Data Protection Regulation. According to a law-firm summary, it noted that recordings were reviewed only if abnormal behaviour was flagged, and that the university had involved its data protection officer and carried out an impact assessment. It said necessity would depend on circumstances⁹. In the United States, the Ohio ruling concerned room scans at a public university and held them unreasonable¹. In 2022 the Dutch education minister said that safeguarding student privacy is the responsibility of institutions, and that since the software had turned out to be unsafe after a leak, he would not recommend it¹¹.

EU AI Act. Recital 56 of Regulation (EU) 2024/1689 says AI systems used for monitoring and detecting prohibited behaviour of students during tests should be classified as high-risk, and recital 44 says systems that detect emotional state in education should be prohibited (recitals are explanatory, not operative text)¹⁴. Utrecht University's guidance lists proctoring under the education category of Annex III and notes duties for institutions as deployers¹⁵. The original application date for these duties was 2 August 2026¹⁵. A law-firm alert of 27 May 2026 reports a provisional political agreement on 6 May 2026 to move stand-alone Annex III obligations to 2 December 2027, with formal adoption still expected¹⁶. Whether it was adopted could not be verified. The articles and Annex III themselves were not read.

Where it falls short

The challenge question is what would make this use case look less successful than it is presented. Four points emerge.

• Thin proof of detection. The one controlled test found has 30 students, one programme and a 2021 date⁴. Vendor volume figures measure how many exams were run, not how many cheating attempts were caught¹².
• No evidence that gains persist or reach those who need them. No before-and-after cheating data was found, and the only fairness measurement found shows higher flag rates for darker-skinned and Black students⁵.
• Dependence on people. The benefit rests on reviewers, and review rates were low in the operator-reported 2021 data⁸. ProctorU responded in 2021 by ending AI-only proctoring and requiring a trained employee to review flagged sessions⁸. Another vendor says it combines AI with live human proctors (vendor-reported)¹³.
• Student experience and disputes. A 2021 survey by a Dutch student association reported that hundreds of students in Amsterdam, Rotterdam and Wageningen were allegedly wrongly accused¹⁸. In the 2018 pilot about 85% of student comments were negative, mostly about privacy and technical problems¹⁷.

Universities have also withdrawn. The University of Erfurt's Senate decided on 2 July 2025 that remote exams with photo proctoring would end; its page states that from summer semester 2026 remote exams with facial recognition are no longer permitted or technically possible. Reasons given were the end of the pandemic exception, internal audit results, system evaluation feedback, and critical views from teachers and students on how effective the cheating prevention was⁷. The page's summary line gives winter semester 2025/26 as the start, while its title and body give summer semester 2026⁷. After the August 2022 ruling, Cleveland State said the matter remained in active litigation³; no appeal or later ruling was found in the pages opened. No independent audit of current products' error rates was found.

Phase: In wide use — why

The hypothesis, In wide use, is kept. Machine-learning flagging is confirmed in named products¹⁴, and deployments are documented at universities in the United States and the Netherlands¹¹¹. The evidence is not clearly stronger for another phase. Two cautions apply: scale figures are vendor-reported and cumulative¹², and most evidence of use dates from 2020-21. If current independent counts showed much smaller use, the phase would be In early use.

What to watch next

• 2 December 2027: the date reported for stand-alone Annex III obligations, if the amending act is formally adopted and published in the Official Journal¹⁶.
• Any appeal or further ruling in the Cleveland State case; none was found in the pages opened, and a court filing would be the observable trigger³.
• Other universities following Erfurt's decision to end remote proctored exams; a Senate or board decision would be the trigger⁷.
• Publication of a new independent test of detection accuracy or false-flag rates for current products; none was found after 2021⁴.

Questions to ask before adopting

• What share of cheating attempts does the tool detect in a controlled test, who ran the test, and when?
• Does a person review every flag, how long does that take, and what happens when no one has time?
• Are flag rates reported by skin tone, disability and nationality, and what is done when they differ?
• What data (video, audio, room scans, desktop activity) is kept, for how long, and who can view it?
• What legal basis and impact assessment cover the use, and would it count as high-risk under the EU AI Act?