University of North Carolina at Chapel Hill Libraries is evaluating AI-generated alt-text, full text image descriptions, and transcriptions to improve accessibility and discovery across repositories containing approximately 2.5 million images. Two pilots compared large language models and refined selected tools (ChatGPT and Gemini) using a five-point rubric informed by WCAG 2.2 AA, Cooper Hewitt Accessibility Guidelines, and UNC’s Conscious Editing Guidelines. This presentation shares examples from the two pilots, explains our evolving model selection, and introduces a workflow that generates descriptions and assesses outputs for inconsistencies, unsupported claims, and bias. AI-generated risk scores and review tags help prioritize staff review, while prompt and processing optimizations reduce token use, processing time, and cost.
No takes yet. Share an insight, caveat, or question.
Stukes et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: