Generative AI for Question Generation, Answer Scoring and Faculty-Supervised Assessment

Abstract

Writing question papers and marking descriptive answers remain the slowest, most subjective parts of examination. This review organises recent work on automating them into three themes: generating questions, scoring answers, and deciding how far automated scores can be trusted. Eight core studies on Bloom's-level question generation, semantic and multi-signal evaluation and LLM-based scoring are examined, together with two 2025 studies. The synthesis shows that quality falls at higher cognitive levels, that combined scoring signals outperform similarity alone, that prompt design strongly affects reliability, and that no surveyed work integrates the full examination workflow. We therefore outline a faculty-in-the-loop reference design in which every generated question and every low-confidence score is checked by a human before use, and set out a plan for evaluating it. A web prototype built with Flask, MySQL and a hosted LLM already implements question generation, timed delivery and objective scoring for multiple-choice examinations, while descriptive scoring and confidence routing remain future work; no evaluation results are reported yet