AI-based grading is increasingly used in education, yet its effectiveness in evaluating handwritten responses for cloze-form questions remains underexplored. This study investigates the reliability of state-of-the-art (SOTA) AI models in assessing 5,000 handwritten responses from 500 cloze-form questions, collected from students with diverse writing styles. We systematically group responses based on handwriting characteristics and compare AI-based grading with human evaluation under a strict grading policy. Our findings reveal a 29.40% average error rate for the best-performing AI model, which is 4.06% points higher than the human grader’s average error rate. Additionally, AI performance varies widely across handwriting groups, indicating potential biases in automated grading systems. These results highlight the challenges of AI-based assessment in real-world educational settings and the need for robust grading models.