← lab / researchR-02 · PART ONE — DESK STUDY

Does AI make interfaces kinder?

The web is being rebuilt by AI faster than at any point in its history — and it just got measurably less accessible for the first time in seven years. Part one: what the public data says. Part two: our own scan of 100 AI-native products, this autumn.

blurple lab researchJuly 20268 min readFIELDWORK: PART TWO IN PROGRESS

fig. 1 — one thousand dots, one thousand home pages. The 41 teal ones pass automated WCAG 2 checks. Data: WebAIM Million, Feb 2026.

Every AI product demo looks effortless: type a sentence, get an interface. The question nobody asks in the demo is who gets left outside. We went looking for the answer in the largest datasets that exist — and the early picture is uncomfortable enough that we're now pointing our own scanner at it.

95.9%

of the top one million home pages have detectable WCAG 2 failures — up from 94.8% a year earlier. (WebAIM Million, 2026)

+10.1%

rise in errors per page in a single year — 51 to 56.1. The first reversal after six years of slow improvement.

1,437

elements on the average home page — 22.5% more in one year, nearly double 2019. Error density holds at roughly one per 26 elements: more code, same defect rate.

01 — THE BASELINE

The web got worse the year AI started writing it.

WebAIM has scanned the top million home pages every year since 2019. The 2026 edition is the first to reverse direction: more pages failing, more errors per page, sharply more markup per page. WebAIM itself points at the likely culprits — the shifts, in its words, "likely reflect broader shifts in web development including increased reliance on 3rd party frameworks and libraries and automated or AI-assisted coding practices."

The failures aren't exotic. The same six categories have topped the list for seven straight years — and they're the basics:

Low-contrast text83.9%

Missing alt text>50%

Unlabelled form inputs51%

share of top-1M home pages affected · WebAIM Million, Feb 2026

02 — THE AI EVIDENCE

Generated interfaces inherit generated flaws.

The direct evidence on AI-built interfaces is young but consistent. A 2025 academic study of generative-AI website builders found persistent WCAG violations across every prompt and platform tested — missing or vague alt text, broken heading structures, keyboard traps — and, tellingly, the same prompt could produce different accessibility outcomes on different runs. The tools aren't reliably bad; they're unreliably everything, which is worse for anyone depending on assistive tech. An earlier ACM study of six fully AI-built websites counted 308 distinct accessibility errors, more than half of them cognitive — the category automated checkers see least.

There's a second pattern hiding in the WebAIM data that should worry AI builders specifically: pages using ARIA — the attribute layer that's supposed to help screen readers — averaged 59.1 errors, against 42 on pages without it. Generated code loves ARIA. Bad ARIA is worse than none.

THE ARIA PARADOX

59.1

avg. errors — pages with ARIA

42.0

avg. errors — pages without ARIA

The layer meant to fix accessibility correlates with more failures — because it's pasted, not designed. W3C's own rule: no ARIA is better than bad ARIA.

03 — WHAT WE'RE DOING ABOUT IT

Part two: we scan the AI wave itself.

The stakes aren't academic. The European Accessibility Act has been enforceable since June 2025, and US regulators have pointed to the growth of AI-generated content while setting digital-accessibility deadlines. Products shipping generated interfaces are accumulating legal surface area at exactly the rate they're accumulating users.

And yet nobody has systematically measured the products selling the AI future. So that's the fieldwork: this autumn we're putting 100 AI-native product homepages — assistants, GenAI builders, AI SaaS — through the a11y-ray engine, plus a manual keyboard and screen-reader pass on a 20-product subsample. Our hypothesis, from what you've read above: the more "magical" the product, the worse it treats assistive tech. We'd love to be wrong.

SAMPLE100 AI-native product homepages, drawn from public directories and traffic rankings; categories balanced across assistants, builders and vertical AI SaaS.

INSTRUMENTa11y-ray engine (axe-core), two viewports, rendered DOM — same method class as the WebAIM baseline, so the numbers can be honestly compared.

HUMAN PASSKeyboard-only and screen-reader walkthrough of the core signup flow on 20 products — the third automated scans can't see.

OUTPUTFull ranking, raw CSV, prompts and method — published open, here. Argue with us: hello@blurple.digital

An honest note on method.

Automated scans catch roughly a third of real accessibility barriers; everything above is a floor, not a ceiling. Part one is a desk study of public data — our own numbers arrive with part two, raw data attached.

sources
[1] WebAIM Million 2026 — the accessibility of the top 1,000,000 home pages · webaim.org/projects/million
[2] "Accessibility in the Age of Generative AI Web Based Builders: Evaluating Web Design Tools for Inclusive Practices" — Proc. 28th Academic Mindtrek Conference, ACM 2025
[3] ACM 2024 — accessibility evaluation of six fully AI-generated websites (308 distinct errors)
[4] W3C, Using ARIA — "No ARIA is better than Bad ARIA"
[5] European Accessibility Act (Directive 2019/882), in application since 28 June 2025

blurple lab · research

all studies →