If your image’s alternative text is turning into a novella, Adrian Roselli wants you to know what that experience can be like. In this post he stuffs the alt attribute, then tests what browsers and screen readers do with it.
Spoiler: the experience is awful on all counts, but you should really read it though to fully grasp the horror.
Worth keeping handy the next time a live-region announcement seems to arrive on its own timeline. Russ Weakley walks through two different parts of the journey: the browser exposing an update to assistive technology and the screen reader deciding when to speak it.
Back at Ability Summit, I moderated a conversation about shifting accessibility left in the age of AI. During that discussion, Yumeng Ma described a problem she’d encountered in her research: there wasn’t a magical prompt that made generated code accessible. The component mattered. The model mattered. And what we chose to measure mattered too.
That research is the subject of a paper I co-authored with Yumeng Ma, Jacob O. Wobbrock, Joe Devon, Michael Fairchild, Jennifer Mankoff, and Venkatesh Potluri: “Prompt and Circumstance: Evaluating Accessibility in LLM-Generated HTML Form Components”, for ASSETS ’26, the ACM SIGACCESS Conference on Computers and Accessibility.
In an internal analysis of accessibility bugs, a colleague found that roughly half could be traced back to design decisions. These were issues we should have caught earlier, before those decisions became codified in code and long before anyone encountered them in a shipped product.
That’s one of the reasons I started working on Accessibility Assistant for Figma. I wanted to help designers catch issues while they worked and help them build their accessibility knowledge along the way. Version 53 brings more of that work together, organizing annotations, assessments, issues, and handoff around projects.
A while back, Eric Meyer emailed to ask about my old post arguing that CSS variables were a bad idea. Did I still feel that way?
No. I’ve changed my mind.
Many years ago, I made a little jQuery plugin called Easy Sortable Tables. It works exactly as promised: point it at a table and it upgrades the UI so a reader can sort the rows by clicking a column heading. The functionality is still useful, but I can’t even remember the last time I added jQuery to a site.
The web platform’s changed quite a bit since I wrote that plugin, so I decided to modernize it as a web component: <table-sortable>.
Apple’s WebKit team recently opposed the WebMCP proposal, arguing that it creates a separate semantic layer for agents when we should instead improve the shared layers people, assistive technology, and agents already use. Jason Grigsby’s post led me to Apple’s comment, and I share the concerns raised in both. In fact, I’d made a similar case over email back in April, when I first learned about WebMCP.
I keep coming back to tabbed interfaces.
I’ve spent the better part of the last 2 years both using coding models and trying to help these models deliver higher quality code and I honestly could not agree more with Zach here:
I will continue to make the argument that craft defined as caring about the quality of output (independent of the tools used to generate the output) is more important than ever. Caring about what you ship will always have real-world value to visitors/consumers/users real people.
EstiShay filed a bug report that’s concerning. They found that Claude Code treats accessibility fixes as optional, even when the project’s requirements file explicitly specifies “WCAG 2.2 AA minimum.” When asked, the model explains:
Claude treats accessibility fixes as optional trade-offs rather than requirements, even when the project’s own rules say otherwise. That’s distinct from just not knowing the right ARIA pattern — it’s a values problem in how the model weighs competing priorities.
So… not a lack of knowledge (the model performs well when reviewing code); it’s a conscious choice, a prioritization failure. Accessibility becomes a “nice to have” in pursuit of coding speed, just like we’ve deprioritized it in human-made software for decades.
The models learned from us and are repeating all of our mistakes. At scale.