Exam proctoring & academic integrity

Generative AI has made classic plagiarism checks insufficient, and remote exams moved cheating and identity fraud into the home. Institutions combine similarity detection, writing-process analytics, identity verification and (more and more cautiously) proctoring software to protect the value of their qualifications. In the EU this is a high-risk AI domain: tools are aids that raise flags; sanctions must remain a human, reasoned, appealable decision.

Problem & value

• Today: lecturers check suspicious texts by hand or with similarity tools; remote exams use lockdown browsers, webcam recording and human invigilators watching many candidates; oral re-examinations settle doubtful cases.
• Gap: text-matching tools find copied sources but not freshly generated text, ghost-writing or paraphrase tools. Automated AI-text detectors give probabilistic scores that vendors themselves caution against using as sole evidence.
• AI value: triage at scale (prioritise the few submissions needing human attention), process evidence (drafts, edit history) instead of guesses, cheaper identity checks for remote and certification exams, and fairer, documented procedures.
• Caution: no reliable, independently validated benefit figures exist for AI-based behaviour flagging; false positives hit non-native speakers and neurodivergent or disabled candidates disproportionately, so pilots must measure this before rollout.

Variants

• Text and code similarity: classic matching against web/publications/prior work (Turnitin, Ouriginal, PlagScan, Compilatio, Urkund heritage), code similarity (MOSS, JPlag).
• AI-writing indicators: classifier-based detectors (Turnitin AI writing indicator, GPTZero, Copyleaks) – indicators only, never proof.
• Process-based integrity: keystroke/version-history evidence, e.g. Google Docs/Word history, Cadmus, Turnitin Clarity, plus authentic assessment and oral defences.
• Remote proctoring: lockdown browser, identity check and recorded or live-supervised sessions (ProctorExam, Proctorio, Honorlock, Respondus, Examity, Mettl, Inspera, Wiseflow, TestReach).
• Certification and entrance tests: identity verification at test centres or at home.
• Not supported: emotion recognition of students (see Risks – prohibited).

How it works

1. submission / exam session
2. identity & access (enrolment check, ID document + selfie match, SSO)
3. evidence capture (text, version history, browser/screen events; webcam only if justified)
4. analysis (similarity search | AI-writing classifier | style-consistency vs. earlier work | event anomalies)
5. triage (score + highlighted evidence; thresholds set low-recall for sanctions, high for review)
6. human review (lecturer / invigilator / examination board)
7. conversation with student (viva, right to be heard)
8. decision
9. appeal path

• Similarity: n-gram/fingerprint matching and embedding search over web and institutional corpora.
• AI-text indicators: fine-tuned encoder classifiers or perplexity-style features; weak on edited, translated or short texts and biased against non-native writers.
• Stylometry: compare a submission with the student's own earlier work; useful as a prompt for a conversation, not as evidence.
• Proctoring analytics: event rules (tab switches, second screen) and optionally vision models (face presence, multiple persons). Flags are time-stamped for a human; no automatic penalty.
• LLM role: summarise evidence into a case file, generate individual viva questions from the student's own submission to test understanding.

Data preparation

Less data preparation than in other use cases, but the data is very sensitive. Needed are student submissions with version history, reference corpora (own previous work, journals, web), enrolment photos and ID documents for identity checks, and exam-session telemetry. They live in the LMS or exam platform, student information system, file shares and sometimes paper archives.

Typical gaps: previous theses not available as searchable text (scans), no consent or retention rules for webcam recordings, missing baseline writing samples per student, inconsistent file formats, and unclear access rights to recordings.

Preparation steps: (1) digitise and OCR older theses to build an own reference corpus (PREP-006); (2) convert formats to plain text (PREP-007); (3) define data classification and retention for recordings (PREP-024); (4) check purpose and legal basis (PREP-025); (5) build a small evaluation set of known cases to measure false positives (PREP-018); (6) restrict access to recordings to the examination board. Tools: OCR such as Paperless-ngx, the LMS, a text-extraction script.

Rough share of effort: around 10-20 percent, mostly the reference corpus and legal set-up. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (certification bodies, training providers): Path A, because certified proctoring and plagiarism products are mature and the process is standard.
• Public (universities, schools): Path C, because biometric and webcam data are very sensitive, and open-source tooling on sovereign EU hosting (Moodle plugins, local similarity search) avoids transferring student data to third parties.
• Developer: Path C, because a high-risk system needs full control of data and explainability for conformity work.

Path A — Buy (recommended for certification bodies and enterprise)

• Similarity/AI indicators: Turnitin (US, EU data residency options), Ouriginal/Turnitin, Compilatio (FR, EU), PlagScan/Ouriginal (DE heritage), Copyleaks, GPTZero.
• Assessment platforms with proctoring or lockdown: Inspera (NO, EU), Wiseflow/UNIwise (DK, EU), ProctorExam (NL, EU), TestReach (IE, EU), Respondus, Proctorio, Honorlock, Examity (US).
• Choose A when: the institution lacks ML staff; require EU hosting, a DPIA-ready processing agreement, accessibility features, and the vendor's AI Act conformity statement (Annex III 3(d)). Prefer tools that present evidence rather than only a score.

Path B — Build on cloud APIs

• Embeddings + search (Mistral, Azure OpenAI EU regions, Cohere) over an institutional corpus for internal similarity; LLM (Claude, Mistral, GPT, Gemini in EU regions) for case summaries and viva questions.
• Identity verification APIs: EU eIDAS-aligned providers (e.g. IDnow, Onfido/Entrust, Veriff) – biometric processing needs explicit legal basis (see Risks).
• Choose B when: you integrate into an existing LMS (Moodle, ILIAS, Canvas) and want control of thresholds and evidence format.

Path C — Self-host (open source; recommended for public institutions)

• Code similarity: JPlag (KIT, open source), MOSS-style tools; text: open sentence embeddings (BGE-M3, multilingual-e5) with OpenSearch/pgvector; fingerprinting via winnowing.
• LLM for case summaries: Mistral, Llama or Qwen models (~8–30B) via vLLM/Ollama on HW2.
• Detection of AI text: open detectors exist (e.g. Binoculars) but are research-grade.
• Choose C when: data must not leave the university or for code courses.

Path D — Custom training

Training own AI-text detectors or cheating-behaviour models is rarely justified: data on true misconduct is scarce, labels are contested, drift with every new model release is fast. Stylometric models per institution need consent and careful bias audits.

State of the art (2026)

• AI-text detection is an arms race: detectors lose accuracy against paraphrasing, mixed human/AI writing and newer models; vendors state scores are not for sole disciplinary use, and several universities disabled them. Peer-reviewed studies show bias against non-native English writers (Liang et al., 2023).
• Shift to process and assessment design: version histories, in-class writing, oral exams, authentic tasks and AI-permitted assessments with disclosure (e.g. traffic-light AI-use scales).
• Remote proctoring has lost ground: after legal challenges and accessibility complaints, many institutions replaced webcam monitoring with on-site or hybrid exams, lockdown browsers or open-book redesign. Where used, human-in-the-loop review is the norm.
• Deepfake/identity risks (synthetic faces, voice clones, remote-control tools) push toward ID checks with liveness detection; see ANOM-014, ANOM-015.
• Limits: false positives, bias, no ground truth, privacy intrusion in the home, easy circumvention (second device, paraphrasers).

Shared resources & combinations

• Same stack as EDU-002/EDU-003 (LLM, embeddings) and EDU-007 (analytics): shared corpus index.
• Identity: ANOM-014 (document forgery & liveness), VIS-016 (face recognition; stricter rules), SPEECH-010 (voice biometrics – avoid for exams unless strictly justified).
• Content provenance: ANOM-015 (AI-content marking, Art. 50 AI Act).
• Hardware: HW2 suffices for embeddings and an LLM summariser.

Effort, cost & KPIs

• A: weeks (SSO/LMS integration, DPIA, policy); licences per student or submission.
• B: 2–4 months for integration, evidence UI and governance.
• C: 3–6 months incl. corpus building; low licence cost, staff cost dominates.
• KPIs: share of flags confirmed after review, false-positive rate by subgroup (language, disability), review time per case, appeals upheld, student-trust survey, cases resolved by viva.

People & change

Involve management and the examination board (process owner and decision maker), the legal office, IT security and the DPO from the start. Students and their representatives must be informed and heard early, since proctoring intrudes on privacy. A DPIA is required in practice (webcam, biometric ID, systematic evaluation of people; GUIDE-005), and an AVV with every provider. Under the AI Act, systems for monitoring prohibited behaviour during tests are high-risk (Annex III) with duties from 2 Dec 2027 (GUIDE-014); emotion inference in education is prohibited.

Works council: yes where examiners, invigilators or staff are logged or evaluated; the objective suitability to monitor behaviour triggers co-determination under §87(1) no. 6 BetrVG (staff council in the public sector), plus §90(1) no. 3 information (GUIDE-004).

Support AI literacy of examiners (Art. 4 AI Act, GUIDE-006): AI-writing detectors have error rates and must never be sole evidence. Typical concerns are false accusations and surveillance. Address them with a documented appeal route, oral defence as follow-up, human decisions on every case, and a transparent policy communicated before the exam period. Pilot in low-stakes exams first (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: Annex III 3(b) high-risk covers AI evaluating learning outcomes; 3(d) covers AI that monitors and detects prohibited behaviour of students during tests. Obligations (risk management, data governance, logging, human oversight, transparency to deployers, accuracy) apply from the Annex III application date; check the status of the 2025–26 "Digital Omnibus" proposal that may postpone it. Deployers must ensure human oversight and inform affected persons.
• Art. 5 prohibition: inferring emotions of students in educational institutions (Art. 5(1)(f)) is prohibited, except for medical/safety reasons. Do not use "stress", "suspicious emotion" or gaze-emotion scoring. Also avoid biometric categorisation inferring sensitive traits.
• GDPR: webcam video, room scans, ID and face templates are intrusive; biometric templates are special-category data (Art. 9). DPIA mandatory (Art. 35); necessity and less-intrusive alternatives must be shown; consent is rarely freely given in exams. Art. 22 limits solely automated decisions with legal effects; keep human review and right to contest. Dutch courts and several European data-protection authorities have scrutinised proctoring. Short retention.
• Education law: examination regulations of the state/university must authorise the method; document due process, burden of proof lies with the institution.
• Risks and mitigations: false accusations (flags only, viva, appeal), bias (subgroup audits), accessibility (alternatives on request), vendor data transfer to third countries (EU hosting, SCCs), security of recordings.

Getting started

1. Define policy first: allowed AI use per course, evidence standards, appeal procedure.
2. Redesign high-stakes assessments (oral component, process evidence, on-site) before buying tools.
3. Run a DPIA with the data-protection officer; document why proctoring is necessary and proportionate.
4. Pilot a similarity + process-evidence tool on one faculty; measure false positives by subgroup.
5. Train staff on interpreting indicators; ban score-only sanctions.

Sources

1. Regulation (EU) 2024/1689 (AI Act), Art. 5 and Annex III — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
2. European Commission — Guidelines on prohibited AI practices (Art. 5) — https://digital-strategy.ec.europa.eu/en/library/commission-guidelines-prohibited-artificial-intelligence-practices-established-ai-act
3. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
4. Liang et al., "GPT detectors are biased against non-native English writers" (2023) — https://arxiv.org/abs/2304.02819
5. Turnitin — AI writing detection guidance — https://guides.turnitin.com/hc/en/articles/28477544839821-AI-writing-detection-in-the-classic-report-view
6. JPlag source-code plagiarism detector — https://github.com/jplag/JPlag
7. Inspera — assessment platform — https://www.inspera.com/
8. ProctorExam — https://proctorexam.com/