A systematic study evaluating GPT-4's ability to generate accessible HTML code reveals consistent failures: missing form labels, redundant or misapplied ARIA attributes, incorrect use of `required` on checkbox groups, missing fieldset/legend structures, and color-only error states. The author built a Python testing framework that sent prompts to Azure OpenAI's GPT-4, generated 10 iterations per prompt at high temperature, and manually reviewed each output for accessibility compliance. A key finding is that including diff-based accessibility hints in prompts dramatically improved output quality, suggesting the root cause is insufficient accessible code in training data rather than fundamental model incapability. Proposed solutions include curating open-source accessible code datasets, accessibility-aware fine-tuning, integrating accessibility linting into IDE AI tools, and embedding accessibility guidance into default system prompts.
