Every major tech company is racing to get AI into production. Almost none of them have checked whether their AI interfaces work with a screen reader.
That's not a niche concern. It's a design failure hiding in plain sight.
What happened
A detailed practical guide has just been published on making LLM-powered chat experiences accessible — covering the real, hands-on work of ensuring these interfaces function for people using screen readers, keyboard navigation, and other assistive technologies. The fact that this guide needed writing at all tells you something. Chat-based AI interfaces are everywhere now, but the foundational interaction design work hasn't been done.
Meanwhile, the adoption race keeps accelerating. Disney has struck a deal with Adobe to weave generative AI into its creative workflows. Analysis of how organisations are approaching AI adoption argues many are repeating the mistakes of early digital transformation — treating it as a technical bolt-on rather than something that requires genuine design thinking.
The conversation we skipped
Here's what frustrates me. The discourse around AI in design has been dominated by two questions: ethics (whose work trained the model?) and economics (whose job does it replace?). Both matter enormously. But we've somehow skipped a third question that should have come first — does this interface actually work?
Not "work" as in "does it generate a plausible response." Work as in: can a blind user navigate it? Can someone using voice control interact with it? Does it behave predictably for people with cognitive disabilities?
Chat might feel like a simple pattern. A text input, a response, a scrolling thread. But the moment you add streaming text, dynamically updating content, multi-turn context, inline code blocks, and embedded media, you've built something genuinely complex. And complexity without structure is where accessibility collapses.
Most AI chat interfaces treat the conversation as a visual artefact rather than a semantic one. Screen readers can't reliably distinguish between your message and the AI's response. That streaming effect where words appear as the model generates them? Often not announced to assistive technology at all. The user sits in silence while sighted users watch text pour onto the screen.
This isn't an edge case. It's a fundamental interaction design problem we've collectively decided to ignore.
This isn't just an accessibility problem
I know some of you are thinking: "this only affects a small percentage of users." Two things.
First, that percentage isn't small. Roughly 15% of the global population lives with some form of disability. And accessible design benefits far more people than that — anyone in bright sunlight, on a slow connection, or multitasking with their hands full.
Second — and this is the part I think designers should really sit with — the accessibility failures in AI chat reveal that we haven't actually designed these interfaces. We've defaulted to them. Chat became the AI interface not because someone designed it to be the best possible interaction pattern, but because it was the easiest one to ship.
That's the real problem. The rush to adopt AI is outpacing the design work required to make AI usable. Disney's deal with Adobe, organisations layering AI onto existing workflows, product teams bolting chat windows into every SaaS app — it's all moving at the speed of technology, not the speed of design.
A study on AI adoption in education found something worth noting: when organisations actually prepared for AI integration thoughtfully — with clear strategies, support, and training — people felt less threatened and more capable. The ones that just dropped the technology in and hoped for the best? Anxiety went up and talent started heading for the door.
The same principle applies to interfaces. When you design AI interactions with proper semantics, clear structure, and genuine accessibility, you build something that works for everyone. When you ship fast and assume chat is a solved pattern, you build something that works for the default user on a MacBook with perfect vision — and quietly excludes everyone else.
What to do this week
Before your team ships, updates, or even discusses another AI-powered feature, do one thing: test it with a screen reader. VoiceOver on Mac, NVDA on Windows — both free.
Navigate your AI chat interface without looking at the screen. Can you tell whose message is whose? Can you tell when new content arrives? Can you reach every action?
Ten minutes of this will teach you more about the actual quality of your AI interface than any amount of prompt tuning. And if it fails — which, honestly, it probably will — you've just found the most important design work your team isn't doing.