Evaluating AI Support for Implementing WCAG Requirements in Web Development

Sammanfattning

This thesis investigates how a Large Language Model (LLM) can support theimplementation of Web Content Accessibility Guidelines (WCAG) in webdevelopment. Web accessibility is an important aspect of software engineering, asdigital services must be usable by people with diverse abilities. Despite the importanceof accessibility, many development teams struggle to implement WCAG requirementsdue to the complexity of the guidelines and limited accessibility expertise. Recentadvances in Artificial Intelligence (AI)-based development tools suggest that LLMsmay assist developers by identifying and correcting accessibility issues but it remainsunclear how effectively these tools can improve compliance in practice.The purpose of this study was to evaluate whether prompt specificity influencesthe ability of an LLM to resolve WCAG accessibility issues. A controlled comparativeexperiment was conducted using 38 real-world webpages provided by Clira AB. Threeexperimental conditions were compared: a baseline condition with noLLM-modification, a general prompting condition where the LLM received only abroad instruction to improve accessibility and a targeted prompting condition where theLLM received structured feedback detected from the baseline condition. An automatedevaluation pipeline was developed using tools such as Puppeteer and Pa11y togetherwith a large language model.The results show a clear difference in accessibility improvements betweenprompting strategies. The baseline evaluation identified 553 accessibility issues acrossthe tested webpages. Under the general prompting condition, the number of issues wasreduced to 318, corresponding to a reduction of 42.5%. Under the targeted promptingcondition, the number of issues was reduced to 51, corresponding to a reduction of90.8%. The findings show that structured accessibility feedback improves the ability ofan LLM to implement WCAG requirements compared to general prompting alone. Thestudy suggests that LLMs can provide strong support in accessibility implementation,especially when prompts contain structured information about detected issues.