Building the world’s first accessible AI chat interface

" "

On August 21, 2023, the University of Michigan launched what we believe to be the first accessible web chat interface to send prompts to GPT-4 and other large language models — U-M GPT. This application is designed for accessibility according to international standards, optimized for usability with common assistive technology, and built without the barriers found in some other examples out on the web.

This achievement was made possible by the combined efforts of accessibility experts and web developers, who put their skills together to design an accessible, usable user interface for all (not just some) members of the University of Michigan (U-M) community. U-M GPT is designed to meet WCAG 2.1 AA standards, which define digital accessibility for users with disabilities, and the app has been optimized for usability with assistive technology.

U-M Information and Technology Services (ITS) constructed the entire platform in a mere two months. Operating with the agility and drive of a start-up, ITS ensured the rapid and robust delivery of cutting-edge tools for the start of the academic year. Most importantly, the platform is inclusive, equitable, secure, and accessible.  

In order to develop an accessible web application in the context of a short timeline and changing scope, the accessibility and generative AI teams within ITS partnered to include accessibility at every stage of development in an agile process. This meant testing for accessibility not at the end of a long run up to launch, but regularly at every point along the way. 

The teams worked iteratively and more intensively than in typical collaborations. A release would prompt an accessibility review, the review would result in changes, the changes would be verified in the next iteration. In some cases, accessibility experts and developers collaborated directly in co-working sessions to fix and validate accessibility issues in real time. And given the importance and visibility of the end product, members of the accessibility team performed final reviews separately then conferred to discuss results, learn from each other, and merge reports for the developers. 

This collaborative and iterative approach demonstrated the value of working as One ITS organization and learning from each other. Accessibility experts taught developers how to implement an accessible pattern, then the developers applied that knowledge in future designs. The developers taught accessibility experts about development constraints, then the accessibility team included that context in future feedback. By leveraging partnership and communication, this approach turned early issues into learning opportunities and professional growth, rather than technical debt.

U-M GPT is a web application where users can send prompts to AI large language models using a chat interface, adjust the AI models used, and save or browse previous chats. The user interface was built using Tailwind UI components, Vue.js javascript framework, and Typescript programming language; was inspired by ChatGPT; and has been customized throughout for enhanced accessibility. 

Accessibility features in U-M GPT include good contrast between text and background, keyboard accessibility for all interactive components, visible keyboard focus, thoughtful label text, good usability for screen readers, use of ARIA for handling dynamic content, and reflowable content and controls. This application and other U-M AI offerings will continue to be updated and optimized for accessibility. 

The accessible and usable interface of U-M GPT is a key example of U-M ITS’s commitment to serving its entire community, and of positioning powerful technology tools to be used and leveraged by all members of the university.