Building accessible AI: Embedding inclusion across the SDLC

Hello everyone, my name is Carie Fisher. I use she/her pronouns, and I'm a senior accessibility program manager at GitHub. Today I'm talking about building accessible AI, and what it takes to embed inclusion across the software development lifecycle.

But first, I want to tell you about a product my team made accessible — and the person it still excluded.

When I worked at Deque, my team spent months fixing accessibility bugs in a major retailer's app before it shipped. It passed its audit. We thought the app was in a good place. Then we tested it with people who are blind and low vision, and within ten minutes we knew it wasn't.

One tester was asked to buy a pair of shoes from a card grid. She was on an Android phone using TalkBack, which read the grid as a stream: a product name, a number, another number, on and on. The prices were there. The availability was there. But nothing connected them to a product, so none of it meant anything. She opened one product page, came back, and lost her place. She asked if the site was broken. It wasn't.

This wasn't a company that ignored accessibility. They'd invested heavily in it, and the audit was clean when we ran it. But the audit came before release, and the product kept changing after that. Everything that shipped after that check went out unexamined. One moment, one version — a snapshot. Accessibility can't live in a single checkpoint. It has to run through how you plan, design, build, and ship.