Generative AI as Exam Evaluation Copilot

This study investigates whether large language models (LLMs) can help improve the reliability of grading open-ended exam responses by identifying potentially unstable grading decisions. Using real university exam data, it compares AI-generated grades with repeated assessments by the same instructor. Results show that AI–human grading discrepancies are significantly associated with intra-rater grading variability, suggesting that LLMs can act as evaluation copilots, flagging responses that may require reconsideration while preserving instructor control.