AI-Generated Alt Text on Product Photos: Helpful or an Accessibility Liability

A retailer ships a 5,000-product catalog and the accessibility audit comes back with the same finding every quarter: thousands of images with empty alt attributes, product pages with no meta descriptions, and summaries nobody had time to write. The AI alt text tool promises to close the gap in an afternoon. It does, and it creates a new one: alt text that describes the pixels and misses the point.

What AI alt text gets wrong

The failure traces to one limitation: the model sees pixels, not purpose. It does not know the page around the image, the audience, the purpose of the section, or where a link goes. A bar chart becomes bar chart showing statistics, which says nothing about the figures. A product photo gets a confident description of the wrong product, because the model guessed at a brand name or misread text inside the image. A decorative divider gets a full sentence, which forces screen reader users to listen to information that adds nothing. Research cited by practitioners puts the miss rate at 15 to 30 percent for context-critical details, which means roughly one in five AI-written alt texts is wrong in a way that matters.

That matters legally as well as practically. Missing alternative text ranks among the most common WCAG failures, and WCAG 1.1.1 requires text alternatives that serve the equivalent purpose of the image. Purpose depends on context the model does not have. WebAIM reports a 300 percent increase in accessibility-related lawsuits between 2022 and 2026, and small businesses accounted for nearly 42 percent of cases in the Web Accessibility Lawsuit Report for 2024. AI alt text on its own is not a reliable path to ADA compliance, because unreviewed output misdescribes images, labels decorative images that should be hidden, and misses what a linked image does.

Where it helps, and the workflow that works

The honest use is drafting at scale with human review on the images that matter. Vision-language models like GPT-4o, Claude 3.5, and Gemini 2.5 can generate five alt text candidates per image in parallel, and a workflow that selects the most descriptive one and human-reviews the top 20 percent, heroes, product shots, and instructional diagrams, closes most of the gap for a fraction of the manual cost. The remaining 80 percent still needs a spot check, because the errors are not evenly distributed: they cluster on the images where the description matters most.

Two rules make the workflow safe. First, decorative images get an empty alt attribute, alt="", so screen readers skip them entirely, and the AI never sees them. Decide what is decorative before generation, not after, because the model describes everything it is given. Second, never leave the AI default in place on a product page. The product photo is the purchase decision, and its alt text has to name the product, the variant, and the detail that distinguishes it from the next photo. A model that writes a woman holding a bag has not done the job. A human who writes tan leather tote with brass clasp, front view has. The size-chart version of the same problem is AI-Generated Size Charts Measure Wrong. Here Is Why..

The catalog-scale answer

For a catalog of thousands, the answer is not to write every alt text by hand and not to publish AI output unreviewed. It is to generate on publish, land the output in a typed field the schema validates, and keep an editor in the loop to review before anything ships. That is the workflow the AI CMS vendors describe, and it is the only one that scales without shipping errors at scale. The field has to be structured: a 155-character meta description, a 40-word summary, and a plain-language alt string, each returned as exactly those values, each reviewed before publish.

The cost of skipping the review is not theoretical. An AI catalog pipeline can output more than 1,000 images a day, each needing unique alt text, and at a 15 to 30 percent miss rate that is 150 to 300 wrong descriptions a day, live, on product pages, read aloud to the customers who rely on them most. The tool is not the problem. Publishing its first draft is. For the speed half of the same storefront, see Shopify LCP: Where Your Store Is Actually Slow.

Frequently asked questions

Can AI-generated alt text satisfy WCAG 1.1.1?

Not on its own. WCAG requires text alternatives that serve the equivalent purpose of the image, and purpose depends on page context the model does not have. AI drafts plus human review can, unreviewed AI output cannot reliably.

How often does AI alt text get important details wrong?

Research cited by practitioners puts the miss rate at 15 to 30 percent for context-critical details, which means roughly one in five AI-written alt texts is wrong in a way that matters.

What should alt text be for decorative images?

An empty alt attribute, alt="", so screen readers skip them. AI describes everything it is given, so decide what is decorative before generation, not after.

What is a safe workflow for AI alt text at catalog scale?

Generate on publish into a typed field, select the best of several candidates, and human-review the top 20 percent: heroes, product shots, and instructional diagrams. Spot-check the rest, and never publish the first draft unreviewed.