Nobody Reads Alt Text Until a Lawyer Does: A QA Playbook for AI-Generated Image Descriptions
Photo: Rawpixel, CC0, via Wikimedia Commons
A vision model can describe an image in about 400 milliseconds and cost a fraction of a cent. For a team sitting on 200,000 product photos, campaign assets, and user uploads, that math is hard to argue with. So the alt text gets generated in bulk, pushed into the CMS, and everyone moves on to the next backlog item.
Months later, a screen reader user files a complaint about a product page where every single image is described as "a pair of shoes on a white background." Because it's true. Technically. For 4,000 different products.
The mistake is treating alt text as metadata
Alt text gets filed under the same mental category as tags, captions, and EXIF timestamps, which is why it's the first thing teams automate and the last thing they review. But alt text isn't metadata. It's content. It gets read aloud by assistive technology, indexed by search engines, and quoted in legal filings.
In the United States, the legal scaffolding is well established even where the specifics are murky. Title III of the Americans with Disabilities Act has been applied to retail websites in a long line of federal and state cases. The Department of Justice's 2024 rule sets WCAG 2.1 AA as the technical standard for state and local government web content, and Section 508 does something similar for federal agencies and their contractors. Private-sector cases don't cite a single technical standard, but they routinely point to WCAG 2.2 Success Criterion 1.1.1, which asks for a text alternative that serves the equivalent purpose.
"Equivalent purpose" is doing a lot of work in that sentence. A generic label is not equivalent to anything.
Where machine-generated descriptions fall apart
The failure modes are consistent enough that you can plan around them:
- Sameness at scale. One template applied to a thousand variant products. Technically accurate, functionally useless for someone trying to choose between two colors.
- Decorative vs. informative. A model can't tell that the swoosh graphic behind your hero copy is ornament, not information. Decorative images should carry an empty alt attribute, and auto-tagging rarely gets that right.
- Text baked into images. Price callouts, promo badges, sizing charts. The model sees an image; the user needs the words.
- Hallucinated specifics. Colors, materials, brand names, and quantities that sound plausible and are wrong. This is the failure mode that creates the most risk, because it's confidently incorrect.
- Context collapse. The same photo means different things in a support article, a checkout flow, and a lifestyle banner. Alt text depends on purpose, and purpose isn't in the file.
- Function over description. If an image is the only content inside a link, the alt text has to describe the destination, not the picture.
None of these are arguments against using AI. They're arguments against shipping it unreviewed.
A QA workflow that holds up under scrutiny
Tier your library by risk and traffic
Not every image needs the same treatment. Build three tiers: high-risk images (checkout flows, product detail pages, forms, interactive controls, anything with legal or medical claims), medium (category pages, marketing landing pages, help center screenshots), and low (blog inline art, dated campaign assets). Spend human review minutes where a failure is both likely and expensive.
Use AI as a first draft, not a final answer
Generate descriptions in bulk, then route them. High-confidence, high-similarity clusters (a repeated product grid pattern you've already reviewed) can clear with sampling. Anything flagged as low confidence, containing text, or belonging to the high-risk tier goes to a human queue. A reviewer can approve 30 descriptions in the time it takes to write five from scratch, and the review is what makes the process defensible.
Lint before you publish
Automated checks catch a surprising share of problems cheaply. Flag alt attributes that are empty on informative images, longer than roughly 125 characters without good reason, identical across more than N assets, starting with "image of" or "picture of," or missing entirely. Run a scanner like axe on a sample of key templates to catch the structural issues the content checks miss.
Keep an audit trail
Record who or what produced each description, when it was reviewed, and by whom. If you're ever asked to demonstrate a process, "we have a policy" is weaker than "here are 40,000 review records with timestamps." This is also where vendor contracts matter: ask whether your accessibility tooling or content vendor carries any indemnification, and read the answer carefully.
Close the loop at upload
Block new uploads that lack alt text in your CMS or digital asset manager. A hard requirement at intake beats a cleanup project every single time. For user-generated content, provide an optional description field and default to an empty alt attribute when it's blank, rather than fabricating a description.
Test with the thing humans actually use
Automated rules catch maybe a third of real accessibility problems. Once a quarter, sit with a screen reader for an hour and navigate your own checkout as a blind customer would. NVDA on Windows or VoiceOver on a Mac will find phrasing that no linter will: descriptions that repeat the product name three times, descriptions that read like a parts list, descriptions that skip the one detail that matters.
That hour is uncomfortable. It's also the cheapest accessibility audit you will ever run.
What a defensible process looks like
It isn't perfect alt text on every asset. It's a documented standard, risk-based review, a remediation path for complaints, and evidence that you act on what you find. Companies that get sued successfully tend to share a different trait: they were told about the problem and did nothing.
AI can get you 80 percent of the way there in an afternoon. The remaining 20 percent is where the descriptions actually become useful, and it's the part that keeps you out of a courtroom.