A Cross-Model Evaluation of Accessibility in AI-Generated Educational Games

Read full text Summarise with VIRAT
Authors: Thea Eaton
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This research investigates how large language models handle accessibility when generating educational content for children. Nine commercial AI models (Claude 3.5, ChatGPT 4o, ChatGPT 5.2, Gemini 3 Pro, Grok 4.1, LLaMA 3, Mistral Large 3, DeepSeek V3, and Microsoft Copilot) were tested using an identical prompt requesting a simple number-matching memory game for young students. None mentioned accessibility, grade level, or student needs in their initial clarifying questions. The result: 100% accessibility failure rate across all models. Games were evaluated against WCAG 2.2 Level A standards and universally failed keyboard navigation, semantic HTML, ARIA labeling, and screen reader compatibility requirements. Follow-up questioning revealed that every model possesses comprehensive accessibility knowledge and can generate accessible code when explicitly instructed. The failure occurs at the interpretation and prioritization stage: models learned from web-based training data that accessibility is optional, not foundational. This research introduces the concept of "adversarial training data" - corpus content that works against a system's stated purpose of helpfulness and inclusion. When AI is trained on repositories where inaccessible code is the norm, models reproduce those exclusionary patterns by default, regardless of capability. Key findings: Accessibility knowledge exists but isn't applied by default Training data quality, not model capability, determines inclusive design outcomes Accessibility is incorrectly treated as user preference rather than baseline requirement The solution requires ethical corpus curation, not smarter models This research has implications for AI safety, educational technology, accessibility policy, and the broader question of how training data shapes AI values and priorities.

Fields of Study & Subjects

Digital Accessibility for Disabilities Ethics and Social Impacts of AI Artificial Intelligence in Education

Publication details

Journal / source
OSF Preprints (OSF Preprints)
Publication date
19 March 2026
Language
EN
Document type
Preprint