Playbooks · Everyday tasks
CodeA11y: a coding assistant that watches accessibility
Written November 12, 2026
Beginners building web pages with Copilot rarely ask for accessible code, and rarely get it. CodeA11y, a Copilot extension studied at CHI 2025, builds the prompts and reminders into the assistant itself — and its novice users produced measurably more accessible code.
A study worth knowing, from 2025
CodeA11y is not a product launch. It is a research system from a paper presented at CHI 2025 in Yokohama, Japan, by Peya Mowar, Yi-Hao Peng, Jason Wu, Aaron Steinfeld and Jeffrey P. Bigham. The question it answers is one every beginner learning to code with an AI assistant runs into: if the developer never thinks about accessibility, will the assistant think about it for them? On the evidence of this study, the plain assistant does not — and a modified one does measurably better.
The problem, in the paper's numbers
The paper opens with a stark figure: about 96 percent of web pages contain accessibility violations. Specialized accessibility tools exist, but novice developers often remain unaware of them. An AI coding assistant such as GitHub Copilot could in principle generate accessibility-compliant code by default. Whether it does so in practice was, in the authors' words, uncertain — so they studied it before building anything.
What the first study found
The formative study involved 16 developers without accessibility training, working with an AI assistant. The authors found three recurring failures. Developers did not prompt the AI for accessibility. They omitted crucial manual steps, such as replacing placeholder attributes in generated code. And they had no way to verify whether the result complied. None of these is a knowledge problem that a documentation page fixes: the developer does not know what to ask for, does not know which generated parts are unfinished, and cannot tell good output from bad.
What CodeA11y does differently
CodeA11y is a GitHub Copilot Extension, integrated into Visual Studio Code as a chat participant in the Copilot Chat window. It has three features, each aimed at one of the failures the first study found. It produces user interface code that complies better with accessibility standards by default, without waiting to be asked. It prompts the developer to resolve existing accessibility errors in their site. And it reminds the developer to complete parts of the generated code that need manual work, including the manual validation steps an automated check cannot do. Behind those features sit three language-model agents: a Responder Agent, a Correction Agent and a Reminder Agent.
How it was tested, and what the authors report
The evaluation was a controlled study with another 20 novice developers. The authors report that CodeA11y significantly improved the accessibility of the user interface code over GitHub Copilot, and that it guided novice developers by reinforcing accessibility practices throughout the interaction, rather than in a single check at the end. That last point is the design lesson that travels beyond this one system: the reminders worked because they arrived inside the coding conversation, at the moment the relevant code was being written.
Limits
- This is a 2025 study, read as research. The samples are small — 16 developers in the formative study, 20 in the evaluation — and the participants were novices working on study tasks, not experienced teams shipping production sites.
- CodeA11y is a research prototype. The paper links its source code, but it is not a supported product with a maintenance promise.
- The improvement the authors report is on the accessibility of code produced in the study. It is not a claim about any individual site's compliance, which still depends on manual validation — the very step CodeA11y reminds developers to do.
Related playbooks
- T21: a research assistant that reads results first
- ClearPath: browser accessibility extension
- Hatz AI 2.0: chat that plans, browses and saves automations
- Luma: an NVDA add-on that describes the screen
Sources
- CodeA11y: Making AI Coding Assistants Useful for Accessible Web Development — arXiv
- CodeA11y: Making AI Coding Assistants Useful for Accessible Web Development — ACM, CHI 2025 proceedings
Back to: Everyday playbooks hub · Also: GeoVisA11y: maps a screen reader can explore · ClearPath: browser accessibility extension · Luma: an NVDA add-on that describes the screen · Chrome's modern-web-guidance: skills for your agent