Abstract
Marking is among the most time-consuming activities in education, and artificial intelligence (AI) is widely proposed as a remedy. Attention has centred on generative large language models, yet regulators raise concerns about their bias, consistency, and explainability. This paper contrasts generative grading with per-question classification models trained only on an educator’s own marking and deployed in a human-in-the-loop workflow. The architectures are compared against criteria drawn from regulatory guidance: transparency, construct-relevant features, determinism, and specifiable training data. A narrower role for generative models is also identified, in composing and restructuring marker-approved feedback rather than in marking. Deployment analytics from one platform are then reported, covering approximately 166,000 marked questions, 39,500 essays, and six institutions: 66.2% of questions were marked in under one minute, 77.6% to 89.5% of AI suggestions were accepted, and essay models reached a quadratic weighted kappa of 0.93 for high-confidence predictions. Questionnaire responses from earlier pilots (33 markers, 30 students) are also reported: markers described marking as faster and consistent feedback as easier to provide; most students rated feedback more consistent and higher in quality. Implications for institutions selecting or building such tools are discussed, alongside the limitations of vendor-collected data.
Keywords
Automated marking, feedback, classification AI, generative AI, large language models, human-in-the-loop, quantitative platform analytics.
