Automating Accessibility: Describing Digital Collections Using LLMs

This article investigates the viability of using generative artificial intelligence (GenAI) to create alternative text (alt text) for digital collections in an effort to comply with accessibility requirements described in the Americans with Disabilities Act Title II Regulations. With a diverse test set of different collection features, including images with text, contemporary images, historic images, illustrations, and data visualizations, the authors test seven models using a single prompt. The authors developed an assessment rubric to streamline analysis of the results. The research finds that model size does not always correlate to the success of a model—a promising finding for institutions that have a limited infrastructure for resource-intensive computing. The research also finds that modern collections typically outperform historic images. Furthermore, the findings confirm the authors’ assumption that models generally fail to communicate the nuance of materials for which cultural considerations and sensitive topics are central to the significance of the image. Ultimately, the authors’ findings continue to underpin the necessity for a human-in-the-loop approach when using GenAI. Future work is necessary to identify and refine prompts for different types of collection features, further analyze the role of text recognition, and define possible and most fit workflows for using GenAI to create alt text.