AI Keeps Getting Better at Accessibility. Human Eyes Are Still the Point.

Our own accessibility scanner ran on our homepage and gave it 100 of 100, with zero decided failures. If we were the kind of company that stops at the score, we would have taken a screenshot and moved on.

The same scan reported 105 items it could not decide. They sit in a category we call needs review, our manual check bucket, and each one is a question the machine would not answer. Every one of the 105 is a color contrast check: text sitting on background colors that blend from one shade into another. The checking engine cannot pick a single background color behind that text, so it does not guess. It abstains, and it hands the question to us.

That is the honest part of the result. A 100 that hides 105 open questions would be marketing. A 100 that shows them is a worklist.

The figures in this post come from dated records: axe-core 4.14, run across all 82 pages of this site on November 2, 2026, with an independent second checker run the same day on our core pages, and the AAA contrast count from a deeper letter level measurement recorded on October 30, 2026.

It is tempting to treat an undecided item as a tool failure, something a smarter model will simply fix next year. Some of it will get fixed. Pixel sampling (taking a few pixel readings to estimate color) and better rendering models (ways to simulate how the page is drawn) will decide more of these cases, and we are building in that direction ourselves.

But look at the scale of what is left. Across all 82 pages of our site, 8,615 contrast items are undecided for the same family of reasons: over gradients, background images, and places where elements overlap, the engine cannot confirm the exact color behind the words. And it is not just our tool being careful. A second widely used checker, run the same day, also found no errors on our homepage and set seven items aside for human review rather than deciding them. Where a question calls for judgment, the tools stop.

That is not a flaw in the tools. It is an honest report of a measurement limit. The standard asks for contrast to be measured at the point where the blend is hardest on the reader, and today's engines cannot reliably find and measure that exact point across a gradient, so they say so instead of guessing. Better measurement will decide more of these cases. Until it does, the undecided count is the truth about where the tooling stands.

Beyond the undecided items, there is a harder category: checks that no AI or scanner can settle on its own. For these, it is a property of the questions that a person has to answer them. We keep a written register of them for our own site. It currently holds 26 checks, including:

• Listening to the page with a real screen reader, the software many blind and low vision people use every day, such as JAWS, NVDA, and VoiceOver, and judging whether what it says makes sense in order.
• Seeing the keyboard focus indicator (the visible outline or highlight on the active item) with your own eyes and confirming a sighted keyboard user can always tell where they are on the page.
• Watching moving and flashing content and judging whether it is comfortable and safe to look at, not merely whether it crosses a numeric threshold (a cutoff value).
• Reading the words as a person and deciding whether they are plain enough for the audience they serve.
• Judging whether a heading actually describes what follows it, and whether a link label means anything out of context.

We build accessibility tooling for a living, so this is not an argument against automation. It is an argument for automation that knows where it ends. In our own product, that principle shows up three ways:

• The scanner shows needs review as its own category, next to the score, never folded into it and never quietly dropped. If the tool could not decide, the report says so.
• Our screen reader section is a simulation (a mock view that imitates behavior), and we label it as a simulation, in those words, every time. A simulation is useful for finding structural problems early. It is not a screen reader pass, and we do not let it borrow that credibility.
• Our AAA program, the level above the AA standard most sites aim for, is down to 48 remaining items. Our AA standing does not rest on any scan by itself: it comes from automated scans reporting zero decided failures, checked alongside the manual work in our 26 check register, because a scan alone never establishes conformance. All 48 are enhanced contrast items: color pairings that pass the AA contrast bar but fall short of the stricter AAA bar of 7 to 1. They are a separate count from the undecided contrast items above, not a slice of them. The standard scan could not decide that blended text at all; a deeper letter level measurement could, and at the AAA bar these 48 fail. Each one is now a decision about a specific component: a pairing a person has to choose, weigh against the design, and sign off on. Software can line those decisions up. It cannot make them.

Here is where we have landed, and it shapes everything we ship. AI is excellent at what it can decide: scanning every page, applying every rule the same way every time, and finding mechanical failures at a scale no human team can match. Let it do that work. It is genuinely transformative at it.

Accessibility conformance (meeting the standard), though, is a judgment discipline. The standard describes outcomes for people, and people are the only instruments that can confirm many of those outcomes. The moment a tool pretends otherwise, polishing its score by guessing at questions it cannot answer, it has stopped measuring accessibility and started manufacturing comfort.

So yes, the AI frontiers will keep advancing, and our tools will decide more every year. The undecided list will shrink. But the last mile of this work is a person looking at a page, listening to it, tabbing through it, and saying: this works for someone like me, or it does not. People are not the backup plan for web accessibility. They are the point of it.