AI Wrote Your Alt Text. Now It Says the Wrong Thing on 400 Products.

AI-generated alt text scales fast and drifts fast. A model that calls a navy sweater black, or describes a lifestyle shot instead of the product, ships errors no human ever reads.

A merchandising team turns on AI alt text generation for a catalog of 400 products. The first week looks great: every image has alt text, the accessibility scan goes green, and nobody reads any of it. Then a customer complains that the screen reader called a navy sweater black, and a spot check finds lifestyle shots described as if the model in the photo were the product for sale.

Alt text is not a checkbox. WCAG 2.2 criterion 1.1.1, Non-text Content, requires a text alternative that serves the equivalent purpose, and a wrong alternative is worse than a missing one, because a missing one at least signals a gap. Monitoring AI alt text means sampling what the model wrote, not counting that something was written.

What the models get wrong

Color is the first failure. Models guess at color from compressed thumbnails and get it wrong often enough to matter, especially navy versus black and cream versus white, which are exactly the distinctions a shopper needs to buy the right variant. Pattern is second: stripes read as checks, a subtle herringbone reads as solid.

Context is the deeper failure. A lifestyle photo of a model wearing the product on a beach gets described as a beach scene, while the shopper needs to know the product, the color, and the fit. And the model will happily invent details that are not in the image at all, like a pocket that does not exist, because its training data says jackets like this usually have one.

A monitoring routine that catches it

Sample, do not scan. Pull twenty product images a week across categories, read the alt text against the image, and log the error rate by category. A flat error rate hides the truth: the model may be fine on studio shots and wrong on half the lifestyle images, and only a category split shows that.

Flag three patterns automatically. Alt text longer than about 150 characters is usually a caption, not an alternative. Alt text that repeats the product title word for word adds nothing. Alt text containing words like image of or picture of wastes the first words on what the screen reader already announces.

The rule for keeping it

AI alt text is a draft, not a deliverable. Keep it for the long tail of old products, but require a human pass on new arrivals, bestsellers, and anything where color or pattern decides the purchase. And keep the model's output out of decorative images entirely: a spacer or a background texture needs an empty alt attribute, not a generated sentence about a grey rectangle.

Related reading: Monitoring Skip Links After a Navigation Redesign, Monitoring Table Headers on Order History and Account Pages, Monitoring Focus Visibility When Sticky Headers and Banners Keep Changing.

Frequently asked questions

Is AI-generated alt text good enough for WCAG 1.1.1?

Only when it is accurate. The criterion requires a text alternative that serves the equivalent purpose, and a wrong description fails it just as a missing one does. AI output has to be sampled and corrected, not trusted on volume.

What do AI alt text models get wrong most often?

Color and pattern, especially navy versus black and stripes versus checks, and context: describing a lifestyle scene instead of the product. Models also invent details that are not in the image, like a pocket that does not exist.

How do I monitor AI alt text at scale?

Sample about twenty images a week across categories, read the alt text against the image, and log the error rate by category. Also flag alt text over roughly 150 characters, alt text that repeats the product title, and alt text that starts with image of.

Should decorative images get AI alt text?

No. Decorative images need an empty alt attribute so screen readers skip them. Generated sentences about a grey rectangle or a background texture add noise for no benefit.