AI Isn’t the Accessibility Problem

I was watching a developer livestream using AI to generate a complete UI component in real time. The code came out fast, it looked clean, and the developer running the demo was clearly excited. The developer wasn’t specifically prompting for accessibility, and nobody in the thread seemed to notice that the interactive elements had no keyboard support, the form inputs had no labels, and the color contrast was nowhere near passing.

That moment captures the problem pretty well.

I tested this across multiple AI tools

Accessibility doesn’t come up unless someone deliberately puts it in the conversation, and most of the time, no one does.

In the past few months, I did a couple of livestream experiments testing how well AI tools produce accessible code (see Vibe Coding Accessibility Part 1 and Part 2). I tried ChatGPT (using the chat interface directly), Codex, Claude Code, and Lovable. These four tools represent a range of approaches, from chat-based to agent-based to purpose-built app generation.

The results were mixed. Some tools did better with certain types of components than others. Some responded well to even basic prompting. But the through-line was consistent: not one of these tools reliably produced accessible code by default. Every time, I had to explicitly ask for it. And even then, the output needed careful testing before I’d trust it.

That matters because most developers using these tools are not prompting for accessibility. They’re prompting for functionality, for speed, for something that looks right. Accessibility doesn’t come up unless someone deliberately puts it in the conversation, and most of the time, no one does.

Why the defaults are so bad

The models aren’t broken. They’re reflecting the web back at us.

This isn’t really the models’ fault. AI tools learn from existing code, and the existing web is nearly inaccessible.

The 2026 WebAIM Million report, which tests the homepages of the top one million websites, found that 95.9% of them have detectable WCAG (Web Content Accessibility Guidelines) failures. That number actually went up from 94.8% in 2025, reversing six years of slow but steady improvement from 97.8% in 2019. The models aren’t broken. They’re reflecting the web back at us. They were trained on inaccessible code, so inaccessible code is what they produce.

And it’s not just live websites. The tutorials, Stack Overflow answers, and example components that developers learn from are mostly written without accessibility in mind either. When a model has seen ten thousand buttons built without a label, that’s what it gives you when you ask for a button.

Could AI be making the regression worse?

That reversal in the 2026 data caught a lot of people’s attention. Several accessibility professionals have pointed to vibe coding (using AI to rapidly generate large amounts of code with minimal human review) as a likely contributor. I haven’t seen data that confirms this directly, but the theory is plausible. If developers are shipping AI-generated code at speed without properly testing it, accessibility is probably one of the first things to fall through the cracks.

It’s worth paying attention to.

The tool isn’t the problem. How we use it is

Some people in the accessibility community have concluded that because AI causes harm, we should avoid it entirely. I understand the impulse. I just don’t think it holds up.

AI isn’t the first technology to be misused in ways that hurt specific groups. I’m a big crafter, and I know independent pattern designers are dealing with this right now. AI-generated craft patterns are flooding the market with images that look appealing but don’t actually produce a usable result when someone tries to knit, crochet, or sew from them. Real, human-made patterns have to compete with that. But the cause isn’t AI as a category. It’s people using AI irresponsibly, without regard for who gets hurt.

You could make the same argument about search engines, social media, or the internet itself. These technologies have all caused real harm in specific contexts. We didn’t solve that by opting out. We pushed for better norms, better design, better regulation, and better practice. The same approach applies here.

What AI could actually do for accessibility

AI-assisted testing isn’t just a nice feature; it may be the only way to close the gap at any meaningful speed.

Here’s what I think is genuinely worth being optimistic about.

There are millions, probably billions, of websites, apps, and digital products in the world. There are not enough trained accessibility professionals to manually test even a fraction of them. At that scale, we are going to need help. AI-assisted testing isn’t just a nice feature; it may be the only way to close the gap at any meaningful speed.

Some of that is already happening. Traditional accessibility scanners can’t accurately calculate color contrast when text sits on a background image, a gradient, or anything else that complicates how the background renders. On a typical site, that’s dozens to thousands of items punted to humans for review. We trained a model in AAArdvark that samples the background at three points around each letter and checks contrast against what’s actually rendered. That’s work no human could realistically do at scale, and no traditional scanner can do at all.

The code generation side is moving too. Developers have released prompt libraries and model instructions specifically designed to guide AI toward accessible output. When I tested models with explicit accessibility context built into the prompt, the results improved noticeably. The tools can do better. They just need to be pointed in the right direction.

We’re nowhere near a place where AI tools can completely replace developers or accessibility testers in the process. But we are at a place where they can help humans get more done in better ways. We just have to be thoughtful and responsible.

And beyond development, AI has real potential as assistive technology: tools that can read signs, text, and labels for blind users; browser agents that can help a user complete a task when a website has accessibility barriers; models that can interpret poorly structured content that would otherwise be unusable; agents that can produce plain-language explanations or summaries of complex text. Some of these exist in early forms right now.

The accessibility community needs to be in the room

Opting out of the AI conversation doesn’t slow AI adoption. It just means the people shaping these tools don’t include anyone who cares deeply about accessibility.

Opting out of the AI conversation doesn’t slow AI adoption. It just means the people shaping these tools don’t include anyone who cares deeply about accessibility.

I get the frustration. AI is being used in ways that cause harm, and sometimes that harm falls on people who are already marginalized. That’s serious. But dismissing anyone who engages with AI as unserious isn’t protecting anyone. It’s just ceding the space. The models that developers are using right now were not designed with accessibility in mind. That’s something we can actually work on. Through advocacy. Through better training data. Through prompting standards and tooling. Through community pressure on the platforms building these tools.

But only if we show up.

If you don’t like the direction things are going, get involved and try to change it.