Download
- Paper (ACM Digital Library)
- Shorter version in Norwegian (Universell utforming av digitale læremidler)
Summary
This study examines whether AI can write useful image descriptions for Norwegian digital teaching materials. Three AI models (GPT-5.2, Claude Opus 4.5 and Gemini 3 Pro) were tested on 30 images from science and social studies textbooks. Descriptions generated from the image alone were compared with descriptions generated when the model also received the surrounding text, teacher guides and curriculum goals. A survey with 28 respondents and a workshop with Statped specialists showed that the descriptions were often too long and interpretive, and that the models tended to add context as extra content rather than use it to decide what to leave out. The descriptions were more reliable for diagrams than for creative images. Human quality control is still needed.
Citation
Kimogol, S. S., Nachtnebel, D., Safari, M. C., & Sanderson, N. C. (2026). Evaluating the Quality of Context-Aware AI-Generated Accessible Image Descriptions in Digital Teaching Materials. In Universal Access in Human-Computer Interaction (pp. 274-293). Lecture Notes in Computer Science. Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-29653-5_17
@inproceedings{kimogol2026evaluating,
title={Evaluating the Quality of Context-Aware AI-Generated Accessible Image Descriptions in Digital Teaching Materials},
author={Kimogol, Stephen Simei and Nachtnebel, Dan and Safari, Mugula Chris and Sanderson, Norun Christine},
booktitle={Universal Access in Human-Computer Interaction},
series={Lecture Notes in Computer Science},
pages={274--293},
year={2026},
publisher={Springer Nature Switzerland},
doi={10.1007/978-3-032-29653-5_17}
}