arXiv study: WCAG-configured prompts lift AI education accessibility to 97%

Generative AI is increasingly used to produce educational materials — documents, slides, images, audio, and video — yet questions remain about whether those outputs are accessible to people with disabilities. A new paper from Hector R. Amado-Salvatierra, posted on arXiv, tackles this gap head-on with a reproducible protocol for evaluating AI-generated educational content against the Web Content Accessibility Guidelines (WCAG). The methodology compares three prompt conditions applied to the same AI tool: a generic instruction with no accessibility language, a single prompt explicitly configured with WCAG criteria, and a persistent accessibility profile loaded once rather than re-specified each time. The evaluation pairs a WCAG rubric per content type with heuristic validation by human accessibility experts, deliberately sidestepping the limitations of automated scanners.

The paper's exploratory results are striking: rubric-based compliance scores jumped from a pooled mean of 24.2% under the generic condition to 96.7% under the WCAG-configured condition, using one model, one artifact per condition and type, and single-evaluator scoring. The persistent-profile condition was not exercised in this first run. Beyond compliance, the author argues that AI-synthesized content suffers from visual and informational overload — an issue WCAG checklists often miss. The main contribution is methodological: an open evaluation instrument and protocol that the community can apply and extend at full benchmark scale. For educators and tool developers, the takeaway is clear: a small prompt-engineering effort can dramatically improve accessibility, moving AI-generated materials from mostly inaccessible to near-compliant in one step.