Comprehension without load, engagement without persistence: Measuring whether generative AI is accessible to disabled learners

Generative artificial intelligence (GenAI) tools are increasingly offered to disabled learners to make learning accessible. Accessibility is usually evaluated by conformance: checking an interface against the success criteria of a technical standard such as WCAG 2.2. Such checks show that content can be reached, not how much effort using it requires. This systematic analysis of 92 studies (2023 to March 2026) examines whether research measures how much mental effort GenAI-supported learning demands of disabled learners, whether individual learners keep using these tools across repeated sessions or stop, and whether their confidence as learners grows. Seventeen outcome constructs were coded against Cognitive Load Theory, the Technology Acceptance Model and Self-Determination Theory. Studies mostly measured immediate responses: comprehension (57.6%), engagement (53.3%) and attitudes (48.9%). Few measured the processes that decide whether these results last: cognitive load (10.9%), disengagement (4.3%) and self-efficacy (17.4%), giving asymmetries of 5.3:1, 12.3:1 and 2.8:1. Only 27.2% of studies followed learners beyond a single session. A second coder agreed poorly with the original coding and judged that 16 of 23 sampled studies did not meet all inclusion criteria, so these figures are upper estimates. Cognitive load is defined as the demand a task places on working memory; only the load that arises when the output format does not match a learner’s way of accessing information is treated as an accessibility issue. A three-axis framework adds evidence from use to conformance-based evaluation.