#451: Rubric-Guided Generative AI for Scalable Creativity Assessment in Educational Games

Assessing creativity at scale remains challenging in learning sciences. This study examined whether GPT-4o could assess creativity in 421 student-designed levels from Physics Playground. Using a structured rubric, GPT-4o ratings were compared with human raters. Results showed strong correlations (r = .74) with multimodal inputs, and reliability analyses showed consistent performance across runs. Findings suggest that rubric-guided LLMs can approximate human creativity judgments efficiently. This approach offers a promising method for scalable automated creativity assessment.