Validating AI-Based Scoring of Divergent Thinking in Elementary School Children

Abstract

Divergent Thinking (DT), a core aspect of creativity, is commonly assessed using Guilford's Alternative Uses Test (AUT). This study examines the validity of an automated scoring approach based on a Large Language Model (LLM), applied to AUT responses from 106 third- and fourth-grade students. We focused on the flexibility and originality dimensions, evaluating the automated scores against human ratings using content and criterion-related validity. The model showed strong correlations with human ratings for flexibility, and outperformed the best-known benchmark in assessing originality, supporting its validity. These findings suggest that LLM-based scoring offers a scalable and objective alternative for DT assessment in elementary educational contexts.