Accessibility checks on AI-built websites

We gave AI coding tools the same everyday website briefs without accessibility instructions. Then we scanned the pages they built with axe-core.

Results

A run pairs one coding tool with one model. Open a run to see its sites and their scan results.

AIMAC Debt counts only critical and serious axe-core violations. Repeated violations of the same rule on a page add progressively less to its score. Runs are ranked by median page score across the 28 categories that every run captured. Lower scores rank first, with ties broken by the lower average page score. The Violations column counts all impact levels. Smaller pages give axe-core fewer elements to check.

Run rankings by AIMAC Debt
RankRunAIMAC DebtSites scoredViolationsHTML elements
1Codex CLI GPT-6.2-Astra0.02803130
2Codex CLI GPT-6.2-Sol0.02842957
3Antigravity CLI GPT-OSS 120B0.028651166
4Claude Code Fable 5.22.02810316578
5Codex CLI GPT-6.1 Terra5.0286675651
6Codex CLI GPT-6-Luna5.7283174256
7Claude Code Opus 5.57.32857212247
8Antigravity CLI Gemini 3.8 Flash8.0284407705

Method

Each run received the same briefs, one per category. Each tool started with its default model. We chose later models by their rank on aimac.ai. We sent the opening instructions through the /goal command in Codex and as a regular message in Claude Code and Antigravity. A structural check confirmed that every run completed every category before scanning. We opened each saved page in Chromium with external requests blocked and ran axe-core once with its default rules.

Capture tools and settings
ToolVersion or setting
axe-core4.14.1
@axe-core/playwright4.14.0
playwright1.65.1
chromium159.0.8154.6
nodev26.7.0
Injected axe-core script (SHA-256)29348999450cdf2849a8e70b7912fc2da85a849da39c0d0d5394f51d4a9ed95c
Viewport1280 by 1024 pixels
Locale, timezone, color schemeen-US, UTC, light

Limits

  • Each tool used its own Linux container on a Mac mini. Its models ran one after another, each in a fresh conversation and a separate folder. Later models could read earlier folders. Chromium was installed, and some models used it to check their pages. Antigravity ran on Google's free tier. Room checks confirmed access to the public internet and blocked access to the host's home folders, tested local network services and other rooms. Full clean-room qualification is pending.
  • Each model ran once. After the opening instructions, we intervened only with GPT-OSS 120B. We told it to continue five times and shared two errors from the structural checker after it reported finishing with two sites missing.
  • A failed capture leaves the page's accessibility result unknown.
  • Axe-core scanned each page as it first loaded at one desktop size, without clicks, scrolling or other interaction. A full accessibility assessment or a judgment of conformance with the Web Content Accessibility Guidelines (WCAG) requires further testing. The briefs required pages to load without external resources, which can change what a model builds.
  • The rankings reflect critical and serious axe-core violations from one run per model. How much scores vary across repeat runs is unknown.