AI and Accessibility: Describing Graphs with Alternative Texts

The persistent global STEM knowledge gap disproportionately affects the visually impaired, primarily due to inadequate or absent alt text. This study investigates whether Visual Language Models (VLMs) can bridge this accessibility divide by generating accurate alt text for complex STEM imagery. Using VLMs, we conducted a systematic evaluation against a curated dataset of 533 expert-annotated images from education ebooks. Our analysis reveals a clear dichotomy: VLMs perform well on straightforward visual content such as labeled chemistry diagrams but falter significantly with images requiring quantitative reasoning, including unlabeled charts and intricate 3D figures. These limitations highlight risks of misinformation when relying solely on automated descriptions for critical scientific visuals. These findings underscore the necessity of integrating human expertise with VLM efficiency to ensure accuracy and equity. This research proposes a hybrid intelligence approach combining rapid model output with expert validation as the only ethically sound solution to enhance STEM accessibility and foster inclusive knowledge dissemination.