Why AI-Built Sites Are Failing WCAG More

For six straight years, the WebAIM Million report told a slow, encouraging story: the share of top home pages with detectable accessibility failures kept inching down, from 98.1 percent in 2020 to 94.8 percent in 2025. The 2026 report broke that streak. Ninety-five point nine percent of the top one million home pages now carry at least one detectable WCAG 2 failure, and the average page carries 56 of them, a 10 percent jump from the year before, across roughly 56 million distinct errors sampled. That's the first year in six the number went the wrong way.

Two numbers explain a lot of the reversal. Page complexity is up 22.5 percent year over year, and ARIA usage, the markup that tells a screen reader what a custom component actually is, is up 27 percent, with most of that growth implemented incorrectly. Sites are shipping more interactive layers than ever, and the accessibility scaffolding underneath hasn't kept pace. That's not a coincidence in a year when AI-assisted page generation went from novelty to default. The same tool that made shipping a five-tab dashboard trivial also made shipping five tabs' worth of mislabeled ARIA trivial.

The Department of Justice noticed. When it extended the ADA Title II digital accessibility compliance deadline by a year, its stated reasoning included the growth of AI-generated content as an emerging accessibility risk, alongside more familiar factors like resource constraints. That's a federal regulator naming AI output as a specific risk category, not a hypothetical one.

The legal exposure is already showing up elsewhere. Federal ADA website lawsuits hit 3,117 in 2025, a 27 percent jump over 2024, with trackers estimating 2026 closing in on roughly 6,000. Close to 40 percent of last year's federal ADA Title III filings came from people representing themselves rather than a law firm, because generative AI can now draft a legally coherent complaint in minutes and a free scanner can point straight at the violation. The barrier to filing dropped at the same time the barrier to building did. Both came from the same technology.

A 2026 study presented at the Mindtrek conference tested this directly. Researchers ran Wix's and Framer's AI website builders through a battery of prompts, then checked the output with the WAVE accessibility tool and manual review against WCAG. The same failures showed up on every builder, every time: missing or vague alt text, broken heading hierarchies, keyboard navigation that couldn't reach every interactive element, inconsistently labeled components. The study's most important finding wasn't the list of failures. It was that they appeared regardless of prompt complexity or website type. A careful, detailed prompt didn't fix it. A lazy one didn't make it worse. The rate held steady either way.

That's the tell. If better prompting fixed the problem, this would be a skill issue, something a more careful writer or designer could work around. It isn't. The failure sits upstream of the prompt, in what the model was actually trained and scored to get right.

Call it the Default Reviewer: the implicit profile, sighted, mouse-driven, with no assistive technology running, that shaped both an AI model's training data and the human raters who judged its output as good or bad. Nobody wrote that profile down anywhere. It's just who was doing the looking, and it explains the failure pattern with uncomfortable precision. Contrast ratios are invisible to a reviewer with good vision on a well-lit monitor. Missing alt text is invisible to a reviewer who can already see the image. A keyboard trap is invisible to a reviewer who never stopped using a mouse. Wrong ARIA is invisible unless you're one of the people actually running a screen reader that reads the label out loud.

None of this is malicious, and none of it is unique to AI. Hand-coded sites fail these same categories constantly, which is the entire reason WebAIM has been running this report for years. What's different now is speed. A hand-coded site that fails accessibility took someone hours or days to build. An AI-generated one takes minutes, which means the same blind spot is shipping at a volume nobody could hand-review even if they wanted to.

This is a specific version of a pattern we keep running into. We've written about what AI-generated code debt actually costs an MVP: the code runs fine in a demo while quietly accumulating security gaps and duplicated logic that only surface later, at due diligence or when a second developer opens the file. An AI-built site's accessibility failures follow the identical arc. The homepage looks finished. It looks fast. Nothing about a WCAG violation shows up in a screenshot, a client walkthrough, or a "ship it" message. It shows up when a screen reader hits it, or when a demand letter does.

Our breakdown of why AI agent pilots stall before production makes a related argument from the other direction: the pilots that reach production aren't the most capable ones, they're the ones with a visible layer where a human can catch the one output in twenty that's wrong before it compounds. Accessibility is that same idea applied to a static page instead of a running agent. The fix was never a smarter model. It was a specific, deliberate check that someone has to own, because the model was never going to flag its own blind spot.

That's also the real argument to bring to the newer wave of AI wireframing and page-building tools, and it's narrower than "don't use them." Figma, Claude Design, Google Stitch, and the rest are genuinely good at compressing the distance between an idea and a screen. None of them, on their own, replace the accessibility review a senior designer used to do as a matter of habit. The tools got faster. The review step didn't get automatically included, and a lot of teams didn't notice it had quietly disappeared until legal did.

One finding from accessibility auditors makes this concrete: the reliable predictor of whether a site passes isn't which tool generated it. It's whether a human accessibility review happened at all. That's the same finding, three different ways, across code, agents, and now page generation. The AI changes what's fast. It doesn't change what needs a person checking it.

Two things won't fix this, and they're worth naming before the useful list. First, a better prompt, the Mindtrek study already ruled that out. Second, an overlay widget. Overlay tools inject a JavaScript layer that adjusts fonts or contrast on top of a page without touching the underlying HTML, and the data on how that holds up in court isn't encouraging: researchers tracking lawsuit records found more than a thousand businesses, over a quarter of all digital accessibility defendants, were sued despite having one installed. The FTC fined one overlay vendor $1 million in 2025 for overstating what its product could do. An overlay can be a genuine convenience for some users. It isn't a substitute for fixing the page.

What actually works is a short review before launch, the same one a senior developer used to do out of habit before "ship a site by lunch" became achievable:

1. 1.Run an automated scanner, then don't trust it completely. Tools like WAVE catch missing alt text and contrast failures reliably. They can't judge whether an image's description is accurate, whether a heading order makes sense read aloud, or whether a keyboard user can escape a menu they've opened.
2. 2.Tab through the entire page without touching a mouse. If you can't reach every link, button, and form field in a logical order, a screen reader user can't either.
3. 3.Check contrast on the actual brand colors, not the design tool's rendering. A gradient that reads fine on a calibrated monitor regularly fails the 4.5:1 ratio WCAG AA requires against a busy background.
4. 4.Have someone who didn't build it do the review. Whoever prompted the page already knows what it's supposed to say and do. That's exactly the knowledge a first-time visitor, sighted or not, doesn't have.

None of this requires slowing the build back down to a pre-AI pace. It requires putting one person back in the loop who's specifically checking for what the Default Reviewer never could.

A page that passes the Default Reviewer's test and a page that works for the roughly 1.3 billion people relying on assistive technology worldwide are two different claims. 2026 is the year that gap started showing up in federal deadlines, not just audit reports.

If you're about to ship a site an AI helped build, get someone who wasn't in the room for the prompt to tab through it before your customers do.