How to Run an AI Accessibility Audit Before You Publish

10 min read

Make this article actionable

Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.

Open in Agent

Most teams now ship AI-assisted work faster than they can check it. A typical week: an agent drafts a 12-slide launch deck, a 1,800-word announcement, four product images and a 60-second clip. Then a procurement questionnaire arrives asking whether the material meets WCAG 2.2 AA — and nobody on the team can answer.

That gap is fixable. This guide is a repeatable, five-pass AI accessibility audit for AI-generated documents, decks, images, spreadsheets and video, with the prompts to run it and an honest list of what still needs a human.

Why AI output fails accessibility by default

The failures are rarely dramatic. They are structural: the model returns something that looks like a heading, sounds like alt text, and reads fine to a sighted reviewer on a laptop.

The scale is documented. In a W3C Web Accessibility Initiative post published 10 November 2026, WAI Director Shawn Lawton Henry writes that "74% of new web pages in April 2025 included AI-generated content" — a figure the post attributes to Ahrefs — and argues that the tools, not individual authors, are the leverage point: "If the tools we use to create content better supported accessibility, it would have a tremendous impact on making the web accessible to people with disabilities" (W3C WAI, 10 Nov 2026).

The failure patterns are equally well documented. WebAIM's 2026 Million report analyzed one million home pages in February 2026 and found 95.9% had detectable WCAG 2 failures, up from 94.8% in 2025, averaging 56.1 errors per page — 10.1% more than the year before. Six failure types accounted for 96% of all detected errors: low-contrast text (83.9% of pages), missing image alt text (53.1%), missing form labels (51%), empty links (46.3%), empty buttons (30.6%) and missing document language (13.5%) (WebAIM Million 2026).

Two caveats, stated plainly. That report measures the web, not AI output specifically; it does not show that AI content is worse than human content. And automated tools only find detectable failures — WebAIM notes that full conformance is "certainly lower" than the 4.1% of pages that passed its automated scan. What both sources support is narrower and still useful: the errors that dominate are mechanical, and mechanical errors are exactly what an AI pass can find at scale.

Mid-read shortcut

Turn the useful parts into next steps

Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.

Create a brief

The five-pass audit

Run the passes in order. Each has one job, produces one artifact — a findings table — and ends with a decision: fix, accept with a written reason, or block.

Pass 1 — Structure: real headings, not bold text

AI drafts often use bold paragraphs where a heading belongs, skip levels (H2 straight to H4), or produce two H1s in one document. To a sighted reader it looks fine; to a screen-reader user navigating by headings, the outline is noise.

What to do:

  1. Export or copy the document as Markdown or HTML.
  2. List every heading in document order with its level.
  3. Confirm exactly one H1, no skipped levels, and that each heading describes the section beneath it rather than repeating a label like "Introduction" three times.
  4. Convert decorative bold lines into body text; convert real section breaks into headings.

Tables get the same treatment: a header row must be marked as a header, not just bolded, or a screen reader reads cells with no column context.

Pass 2 — Images: alt text is a draft, not a verdict

W3C's guidance is specific about what a good tool does: it requires alt text or a confirmation that none is needed, suggests alt text based on context, and — when an image is reused — offers the saved description as a draft and asks the author to confirm it still fits the new context (W3C WAI).

That last point matters for AI workflows, because generated images get reused across pages, decks and social posts. A description that was right for a hero banner is often wrong inside a slide where the point is the chart on the right.

For every image, decide one of three things:

  • Informative: describe what it conveys, not what it is. "Revenue grew 3x between Q1 and Q4" beats "a chart".
  • Decorative: empty alt (alt=""), so screen readers skip it.
  • Text-bearing: transcribe the visible text exactly, because reading text out of a generated image is unreliable.

Never accept "image of a person working" as alt text. It adds words without adding information.

Pass 3 — Contrast: measure it, don't eyeball it

Low contrast is the single most common failure on the web — 83.9% of home pages in the 2026 WebAIM analysis — and AI-generated decks are a reliable way to add more of it: gradients behind white type, brand grey on off-white, a tint that looked fine on one monitor.

WCAG 2.2 Level AA sets the thresholds: at least 4.5:1 for normal text, and at least 3:1 for large text, defined as 18 point or 14 point bold (W3C, Understanding SC 1.4.3). The same page is blunt about rounding: a computed 4.499:1 does not meet 4.5:1.

Compute the ratio instead of judging it. Any correct implementation of the standard formula works; the point is that a machine does the arithmetic and you record the result:

python
def ratio(fg, bg): def lum(c): c = [x / 255 for x in c] c = [x / 12.92 if x <= 0.03928 else ((x + 0.055) / 1.055) ** 2.4 for x in c] return 0.2126 * c[0] + 0.7152 * c[1] + 0.0722 * c[2] a, b = lum(fg), lum(bg) hi, lo = max(a, b), min(a, b) return round((hi + 0.05) / (lo + 0.05), 2) print(ratio((255, 255, 255), (120, 120, 120))) # 3.95 -> fails for body text

Do this for each text/background pair in the deck template, not for each slide. Fixing the token fixes every slide.

Pass 4 — Links, buttons and form labels

Empty links (46.3%) and empty buttons (30.6%) are mostly icon-only controls with no accessible name, and AI-generated pages produce them readily: an arrow icon, a search glyph, a menu button with no label.

Replace "click here" and "read more" with text that makes sense out of context, because screen-reader users often pull up a list of links. Give every form field a visible label programmatically tied to its input. Give every icon-only button an accessible name.

Pass 5 — Media and exported documents

For video: captions for anything with speech, plus a transcript that includes speaker labels. Check the captions after generation — auto-captioning mishears product names and numbers most often, which are exactly the words that matter.

For documents: PDF export is where structure dies. A PDF printed from a slide deck frequently has no tags, no reading order and no document language. If the PDF is the deliverable, test it with a screen reader, or ship an accessible HTML version alongside it.

Copy-ready prompts

These four prompts turn the audit into an agent task. Run them against the actual file, not a summary of it.

Structure map

text
You are auditing a document for accessibility. Read the attached file. Return a table with one row per block, in document order: | # | block text (first 8 words, exact) | current style | should be | problem | Flag: multiple H1s, skipped heading levels, bold text used as a heading, headings that do not describe their section, tables without header rows. Do not rewrite the document. Do not claim it passes any standard.

Alt text with context

text
For each image in the attached deck, write: 1. role: informative | decorative | text-bearing 2. alt: <= 125 characters, describing what the image conveys in THIS slide 3. confidence: high | low, and why If the image contains text, transcribe it exactly and mark the transcription as unverified. Never reuse a description from another document without saying so.

Link and label rewrite

text
List every link and button in the attached page. For each, give the current accessible name, whether it is empty, and a replacement of 2-5 words that makes sense when read alone. Keep the destination the same.

Findings table

text
Merge your findings into one table: | location | quoted text | WCAG area | severity (blocker/major/minor) | exact fix |. Quote the original text verbatim; if you cannot quote it, write "not located" instead of guessing.

The instruction "do not claim it passes any standard" is deliberate. A model cannot certify conformance, and a report that says "this deck is WCAG compliant" is worse than no report at all.

What automation cannot catch

Automated checks cover the mechanical layer. These need a person, ideally one who uses assistive technology:

  • Meaning: whether an alt text actually conveys the point of the image in this context.
  • Reading order: whether the sequence a screen reader follows matches the visual logic — especially in slides and multi-column layouts.
  • Keyboard reach: whether every interactive element can be reached and operated without a mouse, with a visible focus indicator.
  • Caption accuracy: names, numbers, jargon.
  • Cognitive load: whether a plain-language version exists at all.

Budget an hour for a 15-slide deck. That hour is the difference between a checklist and an accessible artifact.

A publish gate you can keep

Before anything AI-generated goes live:

  • One H1; heading levels never skip; tables have header rows
  • Every image has a role and a description (or empty alt if decorative)
  • Every text/background pair measured; 4.5:1 body, 3:1 large text
  • No empty links or buttons; no "click here"
  • Every form field has a programmatic label
  • Video has captions and a transcript; captions spot-checked for names and numbers
  • Exported PDF tested with a screen reader, or an HTML version supplied
  • Findings table attached to the release ticket, with accepted exceptions written down

Common mistakes

Treating alt text as SEO copy. Keyword-stuffed descriptions are noise for the people who need them.

Fixing slides instead of templates. If the contrast token or the heading style is wrong, every future deck inherits it.

Trusting an AI self-report. Ask for findings with quoted text and locations; "looks accessible" is not a finding.

Auditing after publication. Retrofitting a live document means re-checking every downstream export.

Running it inside an AI workspace

The audit fits naturally where the content is already being made. In Vife, the draft, the deck, the images and the transcript live in the same task: the agent can read the actual files, produce the findings table, apply the fixes and export the deck, then hand you the exception list to review. Keep the audit prompt as a saved instruction so every new document gets the same five passes instead of a fresh improvisation.

The goal is not a compliance badge. It is that the thing you published in twenty minutes with a model still works for the person reading it with a screen reader.