I'll be attending #ASSETS2025 next week to present a poster on a novel way to automatically tag PDFs with high accuracy! Catch me in Poster Session B on Tuesday.
Read the paper here: dl.acm.org/doi/10.1145/...
I'm also exploring internship opportunities for summer 2026 and would love to chat!
I'll be attending #ASSETS2025 next week to present a poster on a novel way to automatically tag PDFs with high accuracy! Catch me in Poster Session B on Tuesday.
Read the paper here: dl.acm.org/doi/10.1145/...
I'm also exploring internship opportunities for summer 2026 and would love to chat!
come present your accessibility research in our Accessibility seminar at CMU!
Do you conduct accessibility research? We want to hear about your work! Sign up here to be a guest speaker at CMU's Accessibility Lunch Seminar in Fall 2025: forms.gle/KvVRvsLRdgyc...
Do you conduct accessibility research? We want to hear about your work! Sign up here to be a guest speaker at CMU's Accessibility Lunch Seminar in Fall 2025: forms.gle/KvVRvsLRdgyc...
An image of a data visualization for assessing the quality of output for accessible web code generation from GitHub Copilot and CodeA11y. These are ranked on a scale of 0-2 where 0 is unacceptable, 1 is average, 2 is good. Color Contrast: copilot 0.75, codea11y 0.88 Form labeling: copilot 0.88, codea11y 1.5 Link labeling: copilot 1.75, codea11y 2 Adding alt-text: copilot 0.25, code11y 0.7
Happy Monday! Check out this new @acm-sigchi.bsky.social paper from @peyajm29.bsky.social and colleagues on making web code more accessible using AI
dl.acm.org/doi/10.1145/...
jeff in front of a slide with two turtles on it -- it says, "we've done pretty good at the first turtle", but there is another turtle", grand challenges in accessibility have depth
Had a wonderful post-CHI visit to the Miraikan Museum, where Chieko Asakawa is the Director. She has brought a bunch of cool accessibility exhibits.
I gave a talk revisiting a few of the grand challenges in accessibility, reflecting on what has changed and the new problems that have emerged
I find studies like this so interesting -- even basic functional "performance" assessment can be so unique depending on who is doing the evaluation and what their needs & priorities are.
the tool does three main things -- 1) accessible by default code suggestions 2) built-in accessibility error detection, and 3) reminders to complete template code
the tool does three main things --
1) accessible by default code suggestions
2) built-in accessibility error detection, and
3) reminders to complete template code
doubled accessibilitiy evaluation on color contrast and form labeling, slight improvement on link labeling, and tripled on adding alt text
using CodeA11y, people created more accessible code --
code generation presents a new opportunity to make UI code accessible, but @peyajm29.bsky.social's work shows current codex models mostly fail on #a11y;
her system CodeA11y improves accessibility of UI code w/ three strategies other tools could adopt
4:20pm in AnnexF206
dl.acm.org/doi/10.1145/...
Thanks Mina!! :)
Thank you! :)
Thanks Frank! :)
Thank you so much :)
I can't stop smiling today as I've accepted an offer to start my PhD and continue working with my wonderful advisors, @jeffreybigham.com and Aaron Steinfeld, at @cmurobotics.bsky.social!
Big thanks to my collaborators: @jeffreybigham.com @yihaopeng.bsky.social, Jason Wu and Aaron Steinfeld
A bar graph with mean scores for web accessibility tasks, comparing outcomes with GitHub Copilot and CodeA11y AI assistants. CodeA11y shows a statistically significant improvement over Copilot in maintaining appropriate colour contrast, form labelling and adding alt texts.
A technical evaluation of CodeA11y with 20 developers demonstrated a statistically significant improvement in producing accessible UI code without compromising their satisfaction and ease of use.
(5/n)
Comparison between user interactions with a baseline AI assistant (GitHub Copilot) and CodeA11y. CodeA11y reinforces accessibility practices in the text response, generates a reminder to check for button contrast and adds appropriate colours in both default and hover states. Copilot gives code for a button with the same colours in the default state, but significantly poorer color contrast in the hover state.
We built CodeA11y, a GitHub Copilot Extension that addresses the observed gaps by consistently reinforcing accessible development practices throughout the conversational interaction.
(4/n)
Our formative study with 16 developers revealed 3 barriers to this opportunity: i) developers failed to prompt for accessibility considerations explicitly, ii) uncritically accepted incomplete suggestions, and iii) struggled to detect accessibility issues in their code.
(3/n)
Prior work shows getting front-end developers to write accessible UI code has proven exceptionally difficult. We argue that AI coding assistants (e.g., Github Copilot), already widely adopted, could offer an opportunity to make UI code more accessible.
(2/n)
The interface of the CodeA11y extension is displayed along with its features: (1) Accessibility-by-default Code Suggestions, (2) Relevant Error Identification and (3) Code Completion Reminders.
Excited to share our #CHI2025 preprint:
CodeA11y: Making AI Coding Assistants Useful for Accessible Web Development
arxiv.org/abs/2502.10884
🧵(1/n)
Thrilled to share that our work on making AI Coding Assistants (e.g, Github Copilot) better at helping developers produce accessible UI code has been accepted to #chi2025 🎉🌸
Thank you! :)
Thank you so much for this! Could I be added too?
