Building accessible AI: Lessons from Microsoft’s Chief Accessibility Officer

Jenny Lay-Flurrie offers insights on building inclusive tech at scale

As AI rapidly transforms how we work, one of the questions facing IT leaders is: How do we ensure this technological revolution is an inclusive one?

Speaking during an interview at the recent Business Disability Forum Global Conference, Jenny Lay-Flurrie, Microsoft's Chief Accessibility Officer, offered Diane Lightfoot, CEO, Business Disability Forum some crucial insights into building inclusive technology at scale.

The first piece of advice is to be patient. Return on investment won't materialise immediately and Lay-Flurrie acknowledged that even businesses as well resourced as Microsoft don’t always get it right first time.

“When you have the scale and volume of products that Microsoft does there's no way we can go after perfection,” she explained. “We really are after progress day after day and continually listening to feedback.

“You also must make sure that you've got strong listening systems, because there's no possible way you can catch it all. Once you've got the foundational items in place such as the right skilling, I think you can then start leaning into the innovation potential.

“I don't think that's unique to tech. I think that is something that any company, regardless of industry, can do. It’s about making sure you got a solid foundation and investing in that and then working through strategically how to connect it all the way from your mission down into the core of the products, no matter what your products are.”

Shift left for accessibility

As AI capabilities expand, responsible deployment becomes paramount. Lay-Flurrie stresses the need for clear principles and strong controls to ensure disability inclusion in both training data and output results. This means working closely with ethics teams to establish the guardrails that prevent AI systems from perpetuating bias or exclusion.

One of the most striking statistics Lay-Flurrie shared is that 50% of accessibility bugs in production originate during the design phase.

“Organisations are creating problems before development even begins,” she said. The solution lies in "shifting left" which involves building accessibility into products by design rather than fixing issues after launch.

AI tools like GitHub Copilot offer tangible opportunities here, helping developers write more accessible code while simultaneously teaching better practices. This represents a fundamental shift from the traditional fix-and-remediate cycle that has characterized most accessibility work.

“We need to shift left so that we are building by design,” Lay-Flurrie said. “We all know that theory. We need to make sure that the actual technology delivers on it, and AI gives us the opportunity to do it.”

The rapid pace of AI advancement creates an urgent need for widespread training, particularly for disabled individuals who stand to benefit most from AI-enabled features. Lay-Flurrie described her own experience as a deaf person, explaining that features like real-time captioning, transcription, sign detection, and automated meeting notes are now standard. Twenty years ago, each would have required a human assistant with considerable notice and might not have been fulfilled even then.

This technological shift could fundamentally change career possibilities for disabled workers, potentially removing longstanding barriers and opening new roles. But realising this potential requires making training free, affordable, available, and itself accessible.

Improving AI-generated imagery

Whilst AI could offer improved career possibilities for many disabled people, it also poses challenges, not least the unrealistic and often negative representations of disabled people created by AI-generated imagery.

Most LLMs are trained with images on publicly available websites, but the problem is that disabilities can be depicted so negatively. This influences the output of images from AI tools, which in turn can misrepresent disabled identities, reinforcing stigma and stereotypes.

Lay-Flurrie explained how Microsoft has been working to improve how disability reflects in Microsoft AI systems.

“One of the biggest issues was the injection of disabled imagery that were being prioritised at the right level. We contacted Little People of America, The Kilimanjaro Blind Trust and several other nonprofits, to say that we really want to fix this.

“We've been able to incrementally launch and embed that data with the right rules and guardrails around it to make sure that if you want to create your own prompt, you're going to get more realistic, representative and authentic results. That's live today for little persons, amputee, limb difference, blindness, autism, Down Syndrome, but we've got plenty more to go.”

Future of disability inclusion

Her work improving the output of Microsoft’s LLM imagery has led to some sobering realisations for Lay-Flurrie.

“The statistic that keeps me up at night is that only 4% of the world’s websites are accessible today. I think we missed this with the first technology wave.”

Nonetheless, she remains optimistic about the future of disability inclusion in businesses around the world.

“We have this enormous responsibility to make sure that we don't repeat the mistakes of the past and we learn from them, and if we do, I think we can see this wave of features and inclusivity and possibility that could finally impact some of the statistics like the unemployment, the underemployment, the availability of societal constructs.

“There is possibility here to impact these if we get it right, but I do think we need to lean in. Success requires active stewardship and sustained investment.”