AI in Moderation: Assessing AI-Generated Alt Text for Digital Collections
In anticipation of the latest Web Content Accessibility Guidelines (WCAG) standards, WCAG 2.2, the [redacted] Libraries are seeking to improve the accessibility of our digital collections for all users, including those with visual disabilities. To do so, one of our priorities over the next year is to add alternative text (alt text) for most of the images in our digital collections. However, with over 1.6 million images to address, our organization needs a way to speed up the process, so we’re conducting a study to see if artificial intelligence (AI) can be used to generate alt text without compromising the quality and inclusivity of that alt text. We also want to find ways to mitigate the environmental impact of our use of AI, to the greatest extent possible. This study will challenge small vision-language models to describe a wide range of image content, including abstract art works, to see how the AI-generated text compares with human-created alt text for the same images. Both sets of alt text will be compared to common standards for effective alt text, and to standards for inclusive and respectful descriptive metadata.