
If you can't answer those questions, then the claim isn't much of a claim... so don't use it to decide how YOUR team builds with AI.

AI x Accessibility Builder & Strategist. Author. Speaker. Teacher. #Woodworking feather.ca/x

If you can't answer those questions, then the claim isn't much of a claim... so don't use it to decide how YOUR team builds with AI.

That doesn't mean it's perfect across the board. It means some models do better than others by default, and that you CAN shape the output to get the right results.

with a bare prompt, 12% fully held keyboard focus; with accessibility guidance, 64%.

Since then I've generated more than 8,000 modal dialogs with AI models and driven every one in a real browser. In one study of about 3,000, I measured six of the models, from Anthropic, OpenAI, Google and DeepSeek:

I read a claim from someone with zero detail: no data, no measurement, no evidence... just a blanket statement that AI generated UI isn't accessible. That's part of the reason I ran my first experiments in the lab.

1. Conditions: was this a one-off attempt, or done with experimental rigour? 2. Constraints: what prompt, guidance and tools did they give it? 3. Model: which AI model, and which version? 4. Measures: how did they check?

Every time you read someone claim "AI generated UI isn't accessible" -- OR that it IS -- I want you to ask yourself a really important set of questions:

I look at older photos of my family, and I do the math on how old my parents were in that photo, how old was I in that photo, and how does it relate to how old I am now, how old my children are… and so on. For some reason it has started to happen with professional contacts and friends that I’ve made through the web as well. I learn their age reading an interview, or through conversation. And then I find myself trying to remember what I was doing at that age. Where my head was at, what I was experiencing, doing… how work was going, what computer I was using, who I was meeting and working with… all of it. Do you do it too? Just me?

For anyone following along, almost NONE of the 2 dozen floppies were readable or had anything useful on them. I'm sure I've got another batch, and hopefully there's something of interest there. I did find in this batch readable copies of: - Netscape 0.9a - Windows 3.1 - Borland Turbo Pascal 7.0 Related: I found an old 250 GB LinkStation Network HDD that has backup files from 2001-2008 ish. I used ChatGPT to help me get it connected to my network and sync the files to dropbox directly using rclone. Hopefully much fun comes from that :)

You'd think system colour pairs like HighlightText on Highlight would just work together by default. In Chrome's dark mode on Android, @AaronGustafson found they fail contrast requirements. Same kind of work we did at WaSP: show the evidence, push the browser makers. Tell the browser teams about these pairs too.
I expected paired system colors to work together. I was wrong. To help others avoid my mistake, I built a system color study you can explore and contribute to. aaron-gustafson.com/notebook…

I expected paired system colors to work together. I was wrong. To help others avoid my mistake, I built a system color study you can explore and contribute to. aaron-gustafson.com/notebook…

Building and launching a wireless application serving .wml? Hardly anyone did that. Serving different data from the same backend that was dependent on the context? Outrageous. Found some files showing some design decisions I made back in 2000-2003. They're screens from a mobile site. At that time, mobile meant WAP (Wireless Application Protocol) serving up .wml files (Wireless Markup Language). We were bringing online the Student Information System for a college. And, while doing that, we decided we'd serve up their timetables and grades on a WAP version of the application. And we designed the AppServer to take in a context object and return the right data depending on where your request came from → if you were wireless, you got TODAY's timetable first rather than showing the entire week. On desktop, you got the week all at once. Everything was expensive, so we didn't just assume they wanted their timetable first, and we certainly didn't want to return their entire week's schedule -- they had to choose it willingly from the menu. In those days, mobile data was charged by the byte. It was the pinnacle of innovation. Award-winning. And today? it's completely forgotten... 👉 Today's "we use AI" is tomorrow's serving up an application with WML. In 2002 great teams asked what the student on the phone actually needed. They built in requirements like accessibility and ease of use. They used their judgement to design the right solution. That's fundamentally the same questions great teams answer today... they need to use their judgement to design and build the right solution, based on what people actually need. The tools have changed. The materials have changed. The technology has changed. But user needs like accessibility and ease of use are still fundamentals of a great design.

Me: That site is still live Claude: Checking to see if the site is still live Also Claude: Yes, the site is still live Meanwhile, Claude: You still haven't sent that email I drafted you since yesterday. Me: When did you last check my email? Claude: *checks email* You did send it at 4:53pm yesterday. I SWEAR my AI fact checks me at 10x the rate it fact checks itself and it should be the other way around.

How does that change now? How are you thinking about it? I share my thinking about these types of topics all the time, and you can enjoy them on my list, algorithm free... feather.ca/x/tofo

For me, the point of one step at a time was moving one improvement forward in your practice so you get good at it, and then moving to another, as opposed to trying to do everything all at once and not making progress on any of them.

Or does one step at a time still make sense, but the steps have to go much, much quicker? And does it apply to entire teams and organizations?