A model can write accessible code

AIMAC, the AI Model Accessibility Checker, comes from the GAAD Foundation in partnership with ServiceNow. It prompts 70 models to build web pages across 28 categories with no accessibility guidance at all, then runs axe-core against WCAG 2.2 Level AA. The score is called AIMAC Debt, and lower is better.

GPT 5.4 Mini finished first with a median debt of 0.00, and the whole 28-category run cost $0.89. At least half its pages had zero critical or serious violations, with zero accessibility instructions. Kimi K3.1 Pro (3.18) and Inkling Small (3.43, for $0.34) came next. Thirty-eight of the 70 models cost under $1.

Plenty of models did much worse. Claude's four models landed between 42nd and 54th, with Haiku 5.5 at 4.92, Sonnet 5.5 at 5.03, Opus 5.5 at 5.55, and Fable 5.1 at 6.11. That's the floor: what you get when you ask for a page and say nothing about accessibility. It tells me models can do this, and that the result depends on which one you pick and what you tell it.

Where I'd hand work to a model

These are jobs where the fix is mechanical or a first draft saves time. Each one still needs a person to look at the result.

  • Triage the scan. Paste axe or Lighthouse output and ask for the findings grouped by cause. AIMAC notes that one bad CSS color can produce dozens of violations, so a long report often comes down to a handful of fixes.
  • Fix contrast. Low contrast is the most common failure by a wide margin: 87.7% of AIMAC's generated pages, and 83.9% of real home pages in WebAIM's 2026 Million report. Finding a color that clears 4.5:1 is arithmetic, and a model handles it well. Then check hover states, focus states, disabled buttons, and text over images, which are easy to miss.
  • Name the unnamed. Missing form labels (17.0%) and empty links (16.2%) were the next most common errors on AIMAC's pages. A model can add labels and accessible names across a whole codebase in one pass.
  • Replace text baked into images. I've flagged this as a severe problem in audits, because screen readers and search engines can't read it. A model can transcribe a promo graphic into real HTML text in seconds.
  • Refactor the markup. Heading order, landmarks, lists built out of divs, links that should be buttons. This is tedious by hand and quick for a model.
  • Draft alt text. A first draft beats a blank field.

Where a person has to step in

Alt text and link text carry meaning. The right alt text depends on what an image is doing on that page. The same photo can need a full description in one spot and an empty alt in another, where it's decoration. A model sees the image and a slice of code. You know why it's there.

The same goes for a “Learn more” link that only makes sense to someone who can see the paragraph above it.

ARIA goes wrong fast. Models add it eagerly, and bad ARIA can leave a page worse off than none. AIMAC found ARIA role errors on 4.2% of generated pages. The first rule of ARIA is to use a native HTML element when one exists, and a model needs to be told that.

Automated checks see part of the picture. Axe-core catches contrast, labels, empty buttons, and ARIA errors. It can't catch keyboard traps, confusing screen reader flow, or whether someone can finish the task. AIMAC says so on its own methodology page. A model working from code never hears the page read aloud or tabs through it either.

Even the top model's 0.00 is a median, so some of its pages still had violations.

Versions drift. GPT 6.1 Sol Pro scores 3.97 against 3.32 for the GPT 6 Sol Pro it replaced, and its violations climbed from 165 to 266. Gemini 4 Argon sits at 13.12 against 4.40 for the Gemini 3.1 Pro Preview before it.

Anthropic has its own examples: Sonnet went from 4.80 to 5.03, and Fable from 4.94 to 6.11. It runs the other way too, since Gemma 4 31B dropped from 10.30 to 4.60.

AIMAC is upfront that these are separate runs that can't isolate what changed in the model. The pattern still holds. Swapping or upgrading a model changes your build.

Someone owns the sign-off. Conformance is a judgment call, and a person has to make it, ideally with input from people who use assistive technology every day.

A workflow I'd trust

  • Tell the model the target. Name WCAG 2.2 AA, 4.5:1 contrast for body text, a label on every input, and visible focus states. AIMAC tests with no guidance, so I haven't seen a number for how much this helps. I treat it as a good bet and keep checking.
  • Ask for changes you can review. Small diffs you can read beat a rewrite of the whole site.
  • Run axe DevTools or Lighthouse on the rendered page.
  • Tab through the page without a mouse. Can you see where you are? Can you reach everything? Can you get out of the menu?
  • Listen to it once with a screen reader. NVDA is free.
  • Read every alt text and link label yourself.
  • Do all of it again when the model changes.

The models are getting better at this, unevenly, and they're cheap enough that skipping them is hard to justify. The part they can't do is use the page.

Let the model do the typing. Keep the testing, the judgment calls, and the sign-off for a person.

Related reading: Somebody made your website and WCAG Levels A, AA, and AAA explained. What an audit from me covers is on Web Accessibility.