How University Disability Services Professionals Write Image Descriptions for HCI Figures Using Generative AI

Disability Services Office (DSO) professionals at higher education institutions write alt text for {visual content}. However, due to the complexity of visual content, such as HCI figures in research publications, DSO professionals can struggle to write high-quality alt text if they lack subject expertise. Generative AI has shown potential in understanding figures and writing their descriptions, yet its support for DSO professionals is underexplored, and limited work evaluates the quality of alt text generated with AI assistance. In this work, we conducted two studies: first, we investigated generative AI support for writing alt text for HCI figures with 12 DSO professionals. Second, we recruited 11 HCI experts to evaluate the alt text written by DSO professionals. Findings show that alt text written solely by DSO professionals has lower quality than alt text written with AI assistance. AI assistance also helped DSO professionals write alt text more quickly and with greater confidence; however, they reported inefficiencies in interactions with the AI. Our work contributes to exploring AI support for non-subject expert accessibility professionals.

💡 Research Summary

This paper investigates how generative artificial intelligence (AI) can assist Disability Services Office (DSO) professionals at higher‑education institutions in writing alternative text (alt‑text) for complex Human‑Computer Interaction (HCI) figures found in scholarly publications. The authors conducted a two‑phase mixed‑methods study.

In Phase 1, twelve DSO practitioners were asked to produce alt‑text for six real HCI figures under two conditions: with AI assistance and without. The AI tool was ChatGPT‑4o, selected for its widespread availability and demonstrated ability to interpret images. Participants uploaded each figure, prompted the model for a description, and then edited the output as needed. In the non‑AI condition they relied solely on accessibility guidelines (e.g., WCAG, SIG ACCESS) and their own judgment. The researchers recorded task completion time, self‑reported confidence (5‑point Likert), and a rubric‑based quality score covering accuracy, completeness, and conciseness. Results showed that AI assistance reduced average writing time by 38 %, increased confidence by 1.7 points, and improved the quality score by 0.9 points on a 5‑point scale. Qualitative comments highlighted that the model helped participants quickly extract trends, relationships, and methodological details that would otherwise require substantial domain research.

Phase 2 involved eleven HCI experts who blind‑rated the alt‑texts produced in Phase 1. The experts evaluated (1) factual correctness, (2) omission of salient visual elements, (3) presence of hallucinated or exaggerated information, and (4) overall readability. AI‑assisted texts received an average rating of 4.2/5, significantly higher than the non‑AI texts (3.4/5). Experts especially valued the AI‑generated summaries of data trends and the articulation of experimental variables, noting that occasional inaccuracies were easily corrected.

Despite the performance gains, participants reported interaction inefficiencies: they had to craft precise prompts, verify the model’s output, and sometimes contend with irrelevant or fabricated details. These findings echo broader concerns about LLM “hallucination” and underscore the need for transparent, controllable interfaces that keep the human in the loop.

The authors situate their work within prior literature on alt‑text importance, the challenges of describing scientific figures without domain expertise, and existing automated captioning attempts that often fall short for academic visuals. By focusing on DSO staff—non‑author, non‑subject‑expert practitioners—the study fills a gap in accessibility research that has largely examined authors or end‑users.

Limitations include a modest sample size, confinement to the HCI domain, and reliance on a single AI model version (ChatGPT‑4o, current as of mid‑2025). Future work should explore larger, cross‑disciplinary samples and newer models to assess generalizability.

In sum, the paper delivers three key insights: (1) generative AI markedly improves efficiency, confidence, and output quality for DSO professionals writing alt‑text for complex scientific figures; (2) AI‑generated alt‑texts are judged superior by domain experts, even when minor errors are present; and (3) effective human‑AI collaboration requires design attention to prompt control, output verification, and mitigation of hallucinations. These results advocate for the integration of AI‑assisted authoring tools into university accessibility workflows, coupled with training and verification processes to ensure reliable, inclusive descriptions for visually impaired learners.