A Statistical Analysis on ChatGPT's Evaluation Bias on Handwritten Diagram Based Exam Scripts

AI has revolutionized almost all fields including education ensuring superior efficiency by streamlining processes. This study aims to investigate possibility of utilizing Large Language Models (LLMs) for assessment/exam scripts evaluation by the educators especially for handwritten diagrambased scripts. 120 handwritten answer scripts which were expected to provide a diagram-based response were utilized. Each script was graded for its concept awareness and clarity of the response by both human evaluators and ChatGPT 5.2 model. ChatGPT was prompted 120 times with the question, marking rubric and the scanned handwritten student's answer script. Statistical analysis including descriptive statistics, score difference analysis, paired t-tests, effect size measurement and correlation analysis bought very meaningful insight into ChatGPT's evaluations. The results highlight the need for a hybrid framework that encourage human-in -the-loop. Because results revealed that ChatGPT significantly ($\mathbf{p}