Evaluating the Feasibility of LLM-Based Automation of Manual WCAG Compliance Testing

Digital inclusion requires that websites and online services are accessible to people with disabilities, yet verifying compliance with the Web Content Accessibility Guidelines (WCAG) remains resource-intensive. Many success criteria require manual auditing, as conventional automated tools detect only a limited subset of violations. This study evaluates the feasibility of using a fine-tuned enterprise large language model (LLM) to (semi-)automate such manual WCAG checks. We developed a Chrome extension that operationalizes selected criteria through structured prompting and tested it against prior expert audits of two benchmark pages. Results show moderate ability to localise problematic code, but weak performance in correctly assigning WCAG criteria and generating reliable improvenent suggestions. While the proof-of-concept demonstrates potential as a audit support tool, its probabilistic outputs currently lack the stability and accuracy required to properly support human accessibility auditing.