glossaryPage.heroH1

Mixed and dependent on item type. On MCQ and fill-in-the-blank items where there is an answer key, AI grading agrees with a human marker 95%+ of the time (essentially matching the answer key with tolerance for minor spelling variation). On short-answer questions up to 200 words with a clear rubric, agreement is 80-88% with a human rubric, with lower-confidence responses flagged for teacher review. For long-form essays and subjective writing, NEPC research finds that current AI grading scores on surface features (length, vocabulary diversity, sentence structure) more reliably than on argument quality, making them unreliable for high-stakes summative essay grading without significant teacher review. Best practice: use AI grading as a teacher-time-saver for objective and well-scoped short-answer items; retain teacher-grader primacy for essays and high-stakes assessment.

Per UNESCO 2024 AI in Education guidance and the EU AI Act 2024 (which classifies education-assessment AI as high-risk), AI grading on high-stakes summative assessment without significant teacher review is not appropriate. The EU AI Act requires human-in-the-loop review for high-risk educational AI. Best practice: AI grading is appropriate for formative assessment (low-stakes weekly quizzes, in-class practice work) with full teacher review available, and inappropriate for summative assessment (final exams, certification, admissions) without explicit teacher review of every AI score. School AI-use policies typically specify which assessment types may use AI grading.

AI grading models can encode bias — NEPC research has documented automated-essay-scoring tools penalising linguistic patterns associated with English-language learners, even when content quality is equivalent. Per OECD AI Principles fairness guidance and EU AI Act 2024 bias-audit requirements, AI grading tools should be evaluated for per-protected-group performance variation. Schools running AI grading should: review per-demographic-group scoring patterns periodically, retain teacher-review-before-release workflow for all AI grades, publish AI-grading policy to students and families per transparency requirements, and provide students with the right to request human-only grading where school policy permits.

NEPC (National Education Policy Center) publishes research on automated essay scoring with critical-perspective analysis. UNESCO "Guidance for Generative AI in Education and Research" (2023, updated 2024) covers AI grading in the broader AI-in-education context. The EU AI Act (2024) provides the regulatory framework classifying education-assessment AI as high-risk. AACE (Association for the Advancement of Computing in Education) publishes peer-reviewed research on edtech ethics including AI grading. Educause publishes practitioner-level deployment patterns. ETS (Educational Testing Service) has published research on automated essay scoring including the e-rater system used in some standardised testing contexts.