Answers are grounded in this paper and cite their sources.
Introduction “Alt-text” is a textual substitute for visual information on digital platforms, primarily by users of screen readers. Effective image description is crucial for inclusive, accessible communication; unfortunately, alt-text creation is often limited by the time and skill constraints of the humans writing it. This baseline study aimed to develop an AI-driven method to semi-automate alt-text generation, with minimal expert moderation, for publicly released science images from NASA's Chandra X-ray Observatory. Methods We used pre-trained vision-language models and image metadata to generate alt-text; however, we found the results to be unreliable with this method alone. We iterated the process to include a two-tier fact-checking system of “exogenous” and “endogenous” checks to ensure the alt-text accurately represented the image and image caption. We then evaluated the alt-texts by using an automated “rubric” and investigated the tool's ability to discern “Good” from “Bad” alt-text using developed examples. Results We compared these scores with human evaluators and found that although human and AI scoring does not always agree, the AI rarely rates alt-text low when a human rates it high, indicating that the AI is proficient at noticing high-quality alt-text. Furthermore, our quantified multi-dimensional evaluation step enables adaptability to automated corrections or training improvements. Finally, we find that although the resulting alt-texts still need human editing, the time required for this (compared with writing the alt-text from scratch) reduces the human time burden by 96%. Discussion This multistage approach aligns with current AI research and demonstrates the role AI can play in creating alt-text for large scientific datasets, while our findings reinforce the importance of using these tools alongside human expertise. In addition to performing the research, we add our real-world results to the Chandra X-ray Observatory public database to improve access to the images.
No takes yet. Share an insight, caveat, or question.
Schonhut-Stasik et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: