Examining o3-mini's Capabilitiesin Generating Websites for Individuals with Low Vision Through Different Prompting Strategies

Sammanfattning

This study investigates the accessibility performance of o3-mini, an AI code generationmodel developed by OpenAI, focusing on how well it addresses the needs of users withlow vision on the web. The evaluation starts by detecting accessibility issues inAI-generated code using the automatic tool WAVE, which maps accessibility errors tothe Web Content Accessibility Guidelines (WCAG). Furthermore, to assess the model’sability to meet accessibility requirements, selected success criteria from the WCAG 2.2were applied in a manual review of the output. The model was prompted using threedifferent instruction strategies, with five websites generated for each strategy. Theresults indicate notable differences in accessibility compliance depending on the promptused. The prompt that followed OpenAI’s documentation produced the most accessibleresults, highlighting the importance of prompt engineering in generating accessible webcontent.