AI-Coded Web Pages Plagued by Accessibility Failures, Experts Warn

Breaking: Nearly 96% of the top one million homepages now fail Web Content Accessibility Guidelines (WCAG), reversing six consecutive years of improvement, according to the 2026 WebAIM Million report. The average page contains 297 accessibility issues—even among companies actively investing in accessibility, per AudioEye’s Digital Accessibility Index.

Experts point to a root cause: large language models (LLMs) have been trained on an inaccessible web, and when developers use AI-based tools to generate code, those systemic flaws get baked into new pages. “With AI-based tools, the gap that exists today is structural, not incidental. The root of the challenge is that LLMs have been trained on an inaccessible Web,” said Mike Paciello, Chief Accessibility Officer at AudioEye.

How AI-Generated Code Fails Real Users

An LLM might build a navigation menu that looks clean in code review but adds conflicting ARIA labels, structures headings by visual size rather than semantic hierarchy, and traps keyboard users inside the component. These problems only surface when a real person using a screen reader or keyboard-only navigation tries to use the page.

When page headers are semantically incorrect, readers with disabilities pick up sections out of order. Focus management—a key accessibility construct—becomes a challenge. If a click opens a new window without proper ARIA labels, low-vision or blind users cannot reliably navigate back and forth; they get stuck in a keyboard trap, often requiring a full system shutdown.

Record Lawsuits Mirror AI-Generated Errors

Accessibility oversights are costly. Retailer Target paid $6 million in damages and $3.7 million in attorney fees after a 2006 lawsuit by blind and low-vision plaintiffs. Since 2020, total accessibility lawsuit filings have more than doubled, with 78% targeting e-commerce businesses. The barriers driving those suits—keyboard navigation failures, missing labels, broken screen reader support—are exactly what AI-generated code most commonly gets wrong.

“Accessibility incidents are far from isolated,” Paciello added. “The barriers driving a record number of lawsuits are structural defects that LLMs inherit from their training data.”

Background: Training on an Inaccessible Web

LLMs are trained on enormous scrapes of public web content. Because the vast majority of that content fails WCAG standards, the models internalize flawed patterns—incorrect heading hierarchies, missing alt text, broken keyboard navigation. When developers prompt an LLM to generate a button or a form, the output often replicates those same errors.

Mike Paciello compares the problem to making lasagna with “noodles, peanut butter, and pears.” The structure is there, but the ingredients don’t add up. The accessibility gap is “baked in.”

What This Means

The current trajectory suggests web accessibility will continue to decline unless AI development practices change. Without deliberate effort to train and validate models on accessible code, AI will amplify exclusion rather than reduce it.

Developers must treat accessibility as a first-class requirement in AI-generated code—testing with real assistive technology, fixing semantic errors, and not relying solely on visual checks. Otherwise, the web risks becoming even more inaccessible for the 15% of the global population living with disabilities.

“AI can be a powerful tool for inclusion,” Paciello said, “but only if we acknowledge that the training data is broken and actively work to fix it.”