Skip to main content

Exam Scoring and Proctoring

This page covers AI used to score assessments, detect cheating, or influence pass/fail and progression outcomes.

Why this is usually high-risk​

Education and certification decisions can affect long-term academic and career opportunities. AI errors in this domain can create large and hard-to-reverse harm, so these uses are usually treated as high-risk candidates.

Prohibited proctoring practices (screen before high-risk)

Proctoring that infers emotional state in an education context is prohibited (Art 5(1)(f)). Biometric categorisation inferring sensitive traits — race, political opinions, religion, sexual orientation, etc. — is also prohibited (Art 5(1)(g)). If your proctoring does either, stop and escalate; do not treat it as merely high-risk.

Required obligations​

WHEN​

  • AI contributes to exam scoring, integrity flags, or progression decisions.
  • A model change can alter scoring or suspicious-behavior outcomes.

MUST​

  • Classify as high-risk candidate and escalate for formal determination.
  • Prevent release until required high-risk controls and evidence are approved.
  • Check the Art 6(3) derogation: it may be lower-risk if it only performs a narrow procedural or preparatory task — but profiling of students makes it always high-risk. Self-classifying as not-high-risk still requires a documented assessment (Art 6(4)) and EU-database registration (Art 49(2)).
  • Support the affected student's right to explanation for scoring/progression decisions (Art 86).
  • Implement transparency notices where the feature also triggers user disclosure duties.

OUTPUT artifacts​

  • High-risk classification record and role assignment.
  • Release gate evidence showing required controls were completed.
  • Disclosure evidence (if applicable), such as UI copy and localization checks.
Real-world illustration

Pre-dates the AI Act and was resolved outside it (a policy reversal, not a court ruling), but illustrative of education-scoring harm (Annex III area 3):

  • Ofqual A-level algorithm (UK, 2020): an automated grading model downgraded roughly 40% of results and was withdrawn after public backlash. (overview)