Assessing the Effectiveness of Easy-to-Read Principles in Different Generative AI Models: Enhancing Knowledge Acquisition for People with Disabilities

Abstract

The principles of easy-to-read (E2R) materials are crucial for individuals with disabilities, as they simplify language, provide clear structures, and use user-friendly designs to reduce the difficulty in understanding information. These principles help individuals overcome barriers to comprehending important content. Online friendships have become a common form of social interaction in today's society; providing easy-to-understand online safety knowledge to people with disabilities has become an important issue in public health and social participation. However, existing educational resources are often insufficient, and developing materials that adhere to E2R principles requires significant time and resources. To address the challenge of insufficient resources, the study introduces generative AI technology to rapidly and effectively create educational materials that comply with E2R principles to meet the learning needs of individuals with disabilities regarding online friendship safety. This study examines the performance of three generative AI models—GPT-4o, Claude-3.5-sonnet, and Gemini-1.5-pro—in developing educational materials based on E2R principles, specifically tailored to provide appropriate online safety knowledge for individuals with disabilities. The research was conducted in two stages: the first stage evaluated the AI models' ability to generate basic knowledge.In contrast, the second stage assessed their ability to integrate E2R principles into generating educational materials. The evaluation used statistical methods, including the Wilcox-on and Kruskal-Wallis tests, to determine each AI model's outcomes in both stages. The results showed that GPT-4o performed best across both stages, consistently generating complete and coherent knowledge and materials. Claude-3.5-sonnet followed closely, with some fluctuations but overall stable performance. Gemini-1.5-pro performed the worst, producing unstable content with significant variability. The findings of this study are significant, as they indicate that GPT-4o is the most suitable AI model for developing easy-to-read educational materials, whether for basic knowledge generation or for integrating external E2R resources to create online friendship safety materials. This research provides empirical evidence and design recommendations that can guide the future use of AI technology in creating educational materials tailored to vulnerable groups.

Ähnliche Arbeiten