Missing or inadequate alt text is the single most common accessibility violation on the web. According to the WebAIM Million report, which audits the top one million websites annually, over 55% of all images on the web are missing meaningful alt text. For screen reader users, an image without alt text is a black hole — they know an image exists, but have no idea what it shows. AI vision models have changed this equation dramatically, making it possible to generate accurate, descriptive alt text at scale.
Why Alt Text Matters
Alt text (the alt attribute on <img> elements) serves several critical functions:
- Screen reader users — Alt text is read aloud by screen readers, providing blind and low-vision users with the information the image conveys.
- Slow connections and broken images — When images fail to load, alt text displays as a fallback, ensuring information is not lost.
- SEO — Search engines use alt text to understand image content. Proper alt text improves image search rankings and overall page relevance.
- Legal compliance — WCAG 1.1.1 (Non-text Content) is a Level A requirement. Every informative image must have alt text. No exceptions.
How AI Vision Models Generate Alt Text
Modern AI vision models (such as Anthropic's Claude, OpenAI's GPT-4V, and Google's Gemini) can analyze an image and produce a natural language description of its contents. The process works as follows:
Step 1: Image Detection
The accessibility tool scans the DOM for images missing alt text. This can happen on initial page load and continuously via MutationObserver to catch dynamically loaded images (infinite scroll, lazy loading, SPA navigation). Images are filtered to exclude decorative elements, tracking pixels, and spacer GIFs based on dimensions and context.
Step 2: Batching and Optimization
To minimize latency and API costs, images are batched. Instead of sending one API request per image, a typical implementation groups up to 20 images per batch request. Each image URL (or base64-encoded image data for cross-origin images) is sent alongside contextual information: the page title, surrounding text, and the element's role in the page layout.
Step 3: AI Analysis
The vision model receives the image and context. It analyzes visual content — objects, people, text within the image, actions, settings, and composition — and generates a concise description following alt text best practices. A well-designed prompt instructs the model to be descriptive but concise (typically 10-30 words), avoid phrases like "image of" or "picture of" (the screen reader already announces it as an image), focus on the information the image conveys in context, and describe text visible in the image verbatim.
Step 4: Caching
Generated alt text is cached (typically in Redis or a CDN edge cache) keyed by image URL hash. Subsequent visitors to the same page receive cached alt text instantly without another AI API call. Cache TTLs are typically set to 30-90 days, with invalidation when the image URL changes.
Step 5: DOM Injection
The generated alt text is injected directly into the alt attribute of the image element in the DOM. The image immediately becomes accessible to screen readers without any page reload. The remediation is logged for review by the site owner in the dashboard.
Quality Considerations
AI-generated alt text is not perfect. Current limitations include:
- Context ambiguity— An AI may accurately describe what is in an image but miss why the image matters on that specific page. A photo of a building might be described as "modern glass office building" when the relevant context is "Accessify headquarters in Paris."
- Text in images — OCR accuracy has improved dramatically but is not 100%, especially for stylized fonts, handwriting, or low-resolution images.
- Complex diagrams — Charts, technical diagrams, and infographics require extended descriptions that go beyond what most AI models generate in a single pass.
This is why Accessify's approach includes a human review layer. AI-generated alt text is stored as "pending" remediations in the dashboard, where site owners can approve, edit, or reject each suggestion. Approved alt text is served with high confidence; pending alt text is served as a better-than-nothing fallback until reviewed.
AI Alt Text vs Manual Alt Text
The ideal approach is manual alt text written by a human who understands both the image content and the page context. But the reality is that most websites have hundreds or thousands of images, content is updated frequently, and user-generated content (forums, marketplaces, social platforms) makes manual coverage impossible at scale.
AI alt text fills this gap. It provides immediate, reasonably accurate descriptions for every image that would otherwise have none. Combined with human review for high-priority images, it delivers both coverage and quality. A website with AI-generated alt text on every image is dramatically more accessible than a website where 55% of images have no alt text at all.
Getting Started
Accessify's widget includes built-in AI alt text generation powered by Anthropic's Claude vision model. When you install the widget, it automatically detects images missing alt text, generates descriptions, and injects them into the DOM in real time. You can review and approve all AI-generated descriptions in the Accessify dashboard, giving you both immediate coverage and editorial control.