Accessibility got worse the year AI was supposed to fix it
Two things happened to web accessibility this year, and they don't agree with each other. WCAG failures went up for the first time in six years. The federal government, watching the same tools everyone else is watching, decided it needs more time before making anyone comply.
The number that broke a six-year trend
WebAIM's 2026 report, scanning the top one million home pages, found 95.9% carrying at least one detected WCAG 2 failure, up from 94.8% the year before. That single point of movement ends six straight years of the number inching down.
| WebAIM Million | 2025 | 2026 | Change |
|---|---|---|---|
| Home pages with a detected WCAG failure | 94.8% | 95.9% | +1.1 points |
| Average distinct errors per page | 51.0 | 56.1 | +10.1% |
| Home pages with low-contrast text | 79.1% | 83.9% | +4.8 points |
The average page now carries 56.1 distinct errors, up 10.1% year over year. The single most common one, low-contrast text, now shows up on 83.9% of home pages. More sites are shipping more errors than at any point since this report started improving.
This is the year AI-assisted coding went from a minority habit to the default way most interfaces get built. Those two facts sitting in the same twelve months is not proof of a mechanism, but it is not nothing either.
The government said the same thing, from the other direction
That same month, the Department of Justice extended the deadline for state and local government websites to meet WCAG 2.1 AA. The original date was April 24, 2026. It moved to April 26, 2027 for entities serving 50,000 people or more, and April 26, 2028 for smaller ones. Legal coverage of the interim rule lists resource constraints and remediation technology that isn't yet reliable enough to lean on, generative AI included, among the stated reasons.
A regulator does not usually cite a specific technology's shortcomings in a compliance extension unless the shortfall showed up in its own review process. The people writing the rule looked at what AI-assisted remediation could currently do and decided it wasn't enough to hold agencies to the original date.
Why speed and coverage pull in opposite directions here
An AI coding agent optimizes for finishing the visible task: does the button work, does the layout render. Contrast ratios, focus order after something loads asynchronously, and whether a screen reader announces a modal correctly are not things a prompt asking for a working feature will surface on its own, because none of them block the page from looking done.
Automated scanners, AI-assisted or not, catch static, structural failures well: missing alt text, contrast math, missing labels. They are far weaker at anything that depends on runtime behavior, which is exactly the category a fast-moving, prompt-driven workflow is most likely to skip.
Speed and coverage were never the same axis, and treating them as one is the actual mistake here. A team can ship faster and more accessibly at once, but only if something in the workflow is explicitly checking for the second property, because nothing about shipping fast checks for it on its own.
What to actually do this week
Pick the page on your site with the highest traffic. Run it through a free automated scanner (WebAIM's own WAVE tool takes about two minutes), and fix the single most common failure type it finds, most likely contrast, since that's the one hitting 83.9% of the web already.
That is not the whole job. It is the ten-minute fraction of the job an AI-generated interface is least likely to have done for you by default, which makes it the fraction worth checking yourself before you ship.