What problem does it solve?
High stakes tests, from language proficiency exams to graduate admission tests and teacher licensing exams, such as the GRE, TOEFL and Praxis, increasingly happen at home rather than in a supervised test center, a shift ETS says the Covid 19 pandemic forced within weeks. Remote testing removes the travel and scheduling burden of a test center, but it also removes the one thing a physical test center offered by default: a proctor who can see the whole room.
A single human proctor watching a live video feed cannot reliably catch a hidden phone, a second screen, a coached answer or a subtle handoff, and cannot review what happened after the fact unless the session was recorded. At the same time, evidence has piled up that automated cheating detection is not reliable enough to stand alone: several institutions, including Yale, Vanderbilt and Johns Hopkins Universities, disabled Turnitin's AI writing detection feature, citing concerns about false positives. Current good practice pairs AI generated signals with a trained human decision, and treats an AI flag as a lead to investigate, not a verdict.
How does it work?
- Verify identity before the exam starts. Government ID checks, facial matching and, on some exams, voice biometrics confirm the person taking the test is who they registered as, and repeat through the session to catch a substitution.
- Analyze behavior throughout the session. AI systems track keystroke and typing rhythm, keep the same person confirmed in the seat and watch for a second person entering the room, and monitor application or browser activity for patterns associated with outside help or content that was memorized or copied rather than produced live.
- Look across sessions as well as within one. Some systems, such as the Duolingo English Test, analyze patterns across thousands of test sessions at once, for example near identical answers from test takers in the same region, which a single proctor watching one session could never see.
- Flag, do not decide. The AI signals are handed to a trained human reviewer as a diagnosis to investigate, with the full session recording available to pause, rewind and examine in context, rather than triggering an automatic finding of misconduct.
- Record for later review. The full session is recorded, so a flagged case, a dispute or an audit can be reviewed after the fact, which a live only human proctor in a test center cannot offer.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Mainstream
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has published a number for the KPIs of this use case yet.
Value drivers: Risk and loss reduction, Compliance quality, Inclusion and access.
Indicative value
A testing organization administering 500,000 remote exam sessions a year
USD 800,000 to USD 4 million
Proctoring cost avoided versus one to one live human proctoring per year
How this is calculated
Formula: sessions * costPerHumanOnlyProctor * costReductionShare. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Remote exam sessions per year sessions, sessions per year | 500,000 | 500,000 | The reference organization. |
| Cost of a fully live, one proctor per session model costPerHumanOnlyProctor, USD per session | 8 | 20 | Editorial assumption for a live, one to one remote human proctor; replace with your own proctoring contract rates. |
| Share of proctoring cost avoided by asynchronous, AI assisted review instead of live one to one proctoring costReductionShare, fraction of cost | 0.2 | 0.4 | Editorial assumption. No source in this page's evidence discloses a cost figure; ETS and Duolingo describe the security approach, not its cost, so this range is a hypothesis to test against your own proctoring contract, not a benchmark. |
What it leaves out: Entirely a modeled hypothesis with no disclosed cost benchmark behind it. It leaves out the cost of the AI systems themselves, the reviewer time still needed for every flagged session, and the cost of a false accusation or a missed cheating case, which for a high stakes exam can be far larger than the proctoring cost itself.
Who already uses it?
6 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Universität Erfurt
Germany · Education · 2025
From 2020 the University of Erfurt ran remote exams in WISEflow, which took a reference photo of each student and, during the exam, compared webcam photos against it with face recognition from Amazon Rekognition. A match score below 99 percent sent a notice to the supervising examiner, who then had to give the student a chance to respond. A former student, supported by the Society for Civil Rights (GFF), sued for damages.
The Landgericht Erfurt dismissed the claim in October 2024, but on appeal the Thuringian Higher Regional Court (OLG Jena) held in November 2025 that processing her biometric data without explicit consent was unlawful under Article 9 GDPR and awarded EUR 200. The university had already decided in July 2025 that remote exams with photo proctoring are no longer allowed from 1 April 2026, and it moved to supervised exams on campus.
No outcome disclosed.
Vrije Universiteit Amsterdam
Netherlands · Education · 2021
During the COVID 19 pandemic, Vrije Universiteit Amsterdam used Proctorio to supervise online exams. The software checks by webcam that the right student sits the exam, and it flags irregularities that human supervisors review afterwards. A student alleged that the face detection worked worse for her because of her skin color. In an interim ruling in December 2022 the Netherlands Institute for Human Rights found enough indications of discrimination to shift the burden of proof to the university.
In its final ruling of 17 October 2023 it found that the university had not discriminated through the use of the software in this individual case, but that it had made a prohibited distinction on the ground of race in handling her complaint, which no university body took up until May 2022. The Institute also held that a university has its own duty to verify that such software does not discriminate.
No outcome disclosed.
Duolingo
United States · Education · 2025
The Duolingo English Test, a remote, at home English proficiency test, secures every session with a proprietary lockdown testing application, layered identity verification (government ID, facial matching, device fingerprinting), and AI assisted behavior analysis, including keystroke and typing rhythm monitoring and cross session pattern detection.
Duolingo states that AI flags suspicious sessions for review but that "decisions aren't automated": a trained human proctor individually reviews every test session, with the ability to pause, speed up or slow down the recording, before any action follows. In two 2025 experiments Duolingo's researchers embedded faked copy typing signals in completed, already certified sessions without telling proctors and measured how many they rejected: about half under the original guidelines and 71% under revised guidelines, with some still accepted.
No outcome disclosed.
Pearson VUE
United States · Education · 2025
Pearson VUE's OnVUE online proctoring service, used to deliver certification and licensure exams for its testing program customers, requires test takers to consent to monitoring by both human proctors and, in Pearson VUE's own words, "assistive AI tools" throughout a session. Pearson VUE's OnVUE product manager Aaron Woodward says AI analyzes behavior patterns and flags anything suspicious to a human for intervention, rather than the AI itself deciding a test taker did something wrong, and the company describes this human in the loop approach as central to how it uses AI in its testing operations more broadly.
No outcome disclosed.
Meazure Learning (ProctorU)
United States · Education · 2021
In May 2021 ProctorU, the academic division of Meazure Learning, announced that it would no longer offer proctoring that only records a session and uses AI or similar analytics tools to flag possible violations for review by the school or test provider. It moved those partners to sessions with trained human proctors. The company said its data showed that schools and testing authorities reviewed only about 11 percent of AI flagged sessions, that AI could flag harmless behavior, and that the review load on instructors was high. The record documents a provider that withdrew an AI only service.
No outcome disclosed.
Educational Testing Service (ETS)
United States · Education · 2020
ETS built a remote proctored testing option for the GRE General Test and TOEFL iBT test in six weeks in 2020, in collaboration with ProctorU, later extending it to the HiSET exam and Praxis tests. The at home option layers identity verification (ID matching and, for TOEFL, voice biometrics), a live remote proctor who conducts a 360 degree room scan and monitors the session, and artificial intelligence that continuously scans for irregularities such as a second person entering the room, working alongside process data analysis by ETS's research and psychometrics team to detect suspicious patterns over time.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A defined set of behaviors that count as a security signal, agreed with academic or legal stakeholders
- Recorded video, audio and process data for every session, held under a clear retention policy
- A calibration set of known good and known problematic sessions to test human reviewer judgment
- Documented accommodations for test takers who cannot meet standard camera, room or biometric requirements
Systems to integrate
- Identity verification and biometric matching provider
- Secure testing application or lockdown browser
- Session recording and case management for flagged reviews
- Test delivery platform and score reporting system
Complexity: High
Identity verification and basic recording are straightforward to buy from a proctoring vendor. What happens after a flag needs a human review process trained and calibrated so it does not simply rubber stamp or blindly reject AI signals, a defensible appeals process, and a bias and accessibility review of biometric checks, since accuracy is not guaranteed to be even across skin tones or for test takers with some disabilities and should be tested on your own population rather than assumed.
- 1
Decide what a flag means before you launch
Write down, before go live, exactly what a flagged session leads to: a human review, a request for clarification, a retest, or a finding, and who can override the AI's signal at each step.
- 2
Calibrate human reviewers against automation bias
A reviewer can lean on an AI flag instead of independently checking the recording. Require guidance that asks for independent evidence in the recording, not the AI signal alone, and test reviewers against known false signals before launch and periodically after. Do not assume a human in the loop fixes accuracy on its own.
- 3
Keep the full recording, not just the flag
A flag without the underlying video or process data cannot be investigated or defended in a dispute. Store the full session so a reviewer, and if needed an appeals panel, can see the context.
- 4
Separate detection from adjudication
The team or system that raises a flag should not be the same one that issues a finding of misconduct against a student, mirroring how a fraud alert and a fraud decision are usually kept apart in other industries.
- 5
Publish the appeal path and use it
Test takers need a real, timely way to contest a flag or a finding, with a person who was not involved in the original flag reviewing it.
Guardrails
- No automated finding of misconduct without a trained human reviewer examining the recording
- A documented, tested appeal path, reviewed by someone not involved in the original flag
- Accommodations for test takers who cannot meet standard camera, room or biometric requirements
- Retention and access limits on biometric data, video and process data of identifiable people
KPIs to instrument
- False positive and false negative rate of the AI signal against reviewer decisions, on a calibration set
- Reviewer agreement rate and how often reviewers accept an AI signal without independent evidence
- Time from flag to human decision
- Appeals received, upheld and their outcomes
Human in the loop
A trained human reviewer examines every flagged session before a finding of misconduct is recorded, and can pause, rewind and review the recording rather than relying on a real time glance alone. ETS describes a live proctor watching every session in real time, with the ability to cancel a session immediately if there is a clear attempt to cheat, pending review, while Duolingo describes a trained proctor reviewing every session's recording anonymously and asynchronously afterward. Both treat AI as an aid to that person's judgment, not a replacement for it.
Common failure modes
- Automation bias in the human reviewer
- A reviewer trusts the AI's flag instead of independently checking the recording, accepting it more often than an accurate signal alone would justify. Test and monitor reviewer behavior against known false signals, not only model accuracy.
- Detection tool disabled after real world false positives
- Several universities, including Yale, Vanderbilt and Johns Hopkins, turned off Turnitin's AI writing detector after it produced false positives that undermined confidence in it; a tool that is not trusted stops being used. Validate detection accuracy on your own population before relying on it for a consequence.
- Biometric and behavioral checks that fail some test takers more than others
- Facial matching and behavioral analysis are not guaranteed to perform evenly across skin tones, ages or for test takers with certain disabilities, which can produce more false flags for those groups if left untested. Test accuracy across demographic groups on your own population and provide a fallback for accommodations.
- A false accusation that is never corrected
- An incorrect finding of misconduct can end a student's exam, admission or licensure prospects. A slow or absent appeal path turns a detection error into a life changing one. Time bound the appeal process and staff it independently of the original decision.
What are the risks and rules?
EU AI Act
High risk
Annex III point 3(d) lists AI systems intended to be used for monitoring and detecting prohibited behavior of students during tests in the context of education and training institutions as high risk. One to one biometric identity verification, confirming a test taker is who they claim to be, falls outside the biometric identification use covered by Annex III point 1(a). The Article 5 restriction that does apply is different: Article 5(1)(f) prohibits AI systems that infer emotions in an education institution, so the behavioral analysis in this use case must stay limited to security signals such as identity and typing patterns, and must not be built or read as emotion inference.
Guidance
- Annex III, high risk AI systems (point 3, education and vocational training) (European Union, Europe). Lists monitoring and detecting prohibited student behavior during tests as a high risk education use.
- Protecting Student Privacy (US Department of Education, Student Privacy Policy Office, North America). Guidance on FERPA, relevant to video, audio and biometric data collected during a proctored exam.
Controls to put in place
- Human review of the full session recording before any finding of misconduct
- A timed, independently staffed appeal process
- Documented accommodations for camera, room and biometric requirements
- Retention limits and access controls on video, audio and biometric data
- Accuracy testing of identity and behavior signals across demographic groups
When it went wrong elsewhere
- Universities disable Turnitin's AI writing detector over false positives. Several institutions, including Yale, Vanderbilt and Johns Hopkins Universities, have disabled Turnitin's AI detection feature, citing concerns about false positives, The Chronicle of Higher Education reported. The live article returns an HTTP 403 for automated access; this is the Wayback Machine copy, checked 2026-09-29.
- Thuringia Higher Regional Court awards a student damages over face recognition in Universität Erfurt remote exams without explicit consent. The Thüringer Oberlandesgericht (case 3 U 885/24, judgment of 13 October 2025, announced 17 November 2025) held that comparing a student's webcam images with a biometric reference picture during online exams, done by Amazon Rekognition inside the WISEflow exam platform, needed her explicit consent under GDPR Article 9, which she had not given. It overturned the lower court and awarded her 200 euros in non material damages, dismissing the rest of the claim.
- Netherlands Institute for Human Rights rules on a student's complaint that Proctorio face detection at Vrije Universiteit worked worse for her dark skin. The student reported repeated "face not found" and "room too dark" messages and argued the software detected people with dark skin less well. In an interim ruling of December 2022 (2022-146) the institute found enough indication to presume discrimination and asked the university to prove otherwise. In its final ruling (2023-111) it found the university had shown it did not discriminate against this student, whose problems matched those of other students and had causes unrelated to skin color, but that the university had not handled her discrimination complaint with sufficient care.
Frequently asked questions
- Can AI alone determine that a student cheated?
In the systems in this page's evidence, no. Duolingo and ETS both describe a trained human proctor reviewing every session, with AI as an aid to detect signals a person watching in real time would miss, not as the decision maker. Duolingo says in a blog post that decisions "aren't automated."
- Why are universities turning off AI cheating detectors?
Because the accuracy has not matched the stakes. Several institutions, including Yale, Vanderbilt and Johns Hopkins, disabled Turnitin's AI writing detection tool after it produced false positives. This is different from remote proctoring's identity and behavior signals, which are reviewed by a human before a finding is recorded, but the lesson, validate before you rely on it, applies to both.
- How does remote AI proctoring compare with a human proctor in a test center?
ETS says it built its first remote proctored option, with ProctorU, in six weeks in 2020, and has since run it for hundreds of thousands of test takers on the GRE, TOEFL iBT, HiSET and Praxis exams, layering identity checks, environmental scanning and process data analysis on top of live human proctoring. Duolingo argues this catches some cheating methods, such as collusion patterns across thousands of sessions, that a single test center proctor never could.
- Is exam proctoring AI regulated?
Under the EU AI Act, Annex III point 3(d) lists AI used to monitor and detect prohibited student behavior during tests as high risk, which brings obligations for risk management, human oversight, logging and conformity assessment. Facial recognition used for identity checks can raise separate biometric data questions under GDPR and similar laws.
How to cite this page
Blits.ai AI Use Case Library, "AI assisted proctoring and integrity monitoring for remote exams", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/exam-and-assessment-integrity. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 12 November 2026: Added deployments: Meazure Learning (ProctorU), Vrije Universiteit Amsterdam.
- 12 November 2026: Round F1 (failures and incidents): added the Thüringer Oberlandesgericht judgment on face recognition in Universität Erfurt remote exams and the Netherlands Institute for Human Rights ruling on Proctorio at Vrije Universiteit, both read with usecases:source.
- 12 November 2026: Classification corrected: autonomy changed from supervised agent to assist. Flags sessions for a trained reviewer, who decides every finding.
- 13 October 2026: Added a measured result from Duolingo.
- 29 September 2026: First published