AI-Assisted Accessible Web Interface Generation and WCAG Validation: A Review of LLM-Based Code Generation, Automated Auditing, Repair, and Human Oversight

Across the thirty-five studies reviewed here, AI helps at each stage of building an accessible page, but with a clear limit. Models are good at simple issues and weak at ARIA, keyboard use and anything that depends on meaning. Automated tools miss a large share of real barriers, so a page that passes them is not necessarily usable. Feedback loops that return checker reports work better than general instructions, and context makes image descriptions better. Evidence from disabled users is still scarce. A layered process that generates, checks with more than one tool, repairs with targeted feedback and ends with human review is the safest use of these tools today. For the reference pipeline, the next steps are to build it, test it on real projects against several checkers, add manual tests with screen-reader and keyboard users, and report violations found, violations fixed, new violations created and time spent. No tool guarantees an accessible page. A checked, logged process at least keeps generated code from copying the barriers already on the web.

[1]     World Health Organization, "Global report on health equity for persons with disabilities," WHO, Geneva, Switzerland, 2022, ISBN 978-92-4-006360-0. [2]     World Wide Web Consortium (W3C), "Web Content Accessibility Guidelines (WCAG) 2.2," W3C Recommendation, Oct. 2023. [Online]. Available: https://www.w3.org/TR/WCAG22/ [3]     European Parliament and Council of the European Union, "Directive (EU) 2019/882 on the accessibility requirements for products and services," Off. J. Eur. Union, vol. L 151, pp. 70–115, Jun. 2019. [4]     Government of India, "The Rights of Persons with Disabilities Act, 2016," Act No. 49 of 2016, Dec. 2016. [5]     National Informatics Centre, Government of India, "Guidelines for Indian Government Websites and Apps (GIGW) 3.0," 2023. [6]     WebAIM, "The WebAIM Million: An accessibility analysis of the top 1,000,000 home pages," 2025. [Online]. Available: https://webaim.org/projects/million/ [7]     A. Vaswani et al., "Attention is all you need," in Proc. Adv. Neural Inf. Process. Syst. (NIPS), 2017, pp. 5998–6008. [8]     T. B. Brown et al., "Language models are few-shot learners," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, pp. 1877–1901. [9]     M. Chen et al., "Evaluating large language models trained on code," arXiv:2107.03374, 2021. [10]   OpenAI, "GPT-4 technical report," arXiv:2303.08774, 2023.