There is a real thrill in it. You upload a spreadsheet of 400 product images, run them through an AI captioning tool, and thirty seconds later every single one has alt text. No copywriter, no intern, no billable hours. Just done.
Except it isn't done. It's described. And description is not the same job as alt text.
Here's the gap nobody tells you about when they're selling you the one-button fix: AI is very good at telling you what's in a picture. It is far less reliable at telling you what a picture is for. Those sound like the same task. They are not, and the difference is exactly where automated alt text quietly breaks.
The Stat: A CHI 2026 study found that unreviewed AI-written image descriptions can measurably reduce screen reader task completion compared to human-reviewed alt text. (Source: CHI 2026)
Why "accurate" and "useful" are not the same word
Ask an AI model to describe a photo of a woman laughing on a beach, and it will tell you, correctly, that there is a woman, a beach, laughter, sunlight. That description is accurate. Whether it's useful depends entirely on why the image is on the page.
If that photo is decorative filler next to a paragraph about vacation refund policy, a screen reader user does not need three clauses about the woman's expression. If that same photo is the entire testimonial, the emotional tone is the content, and a generic caption throws away the one thing the image was doing.
AI models don't know which situation they're in, because they're never shown the page. They see a cropped image file, in isolation, stripped of the surrounding context that tells a human writer what actually matters. That's the core failure mode running through every category below.
Where AI alt text genuinely nails it
Give credit where it's due. There are categories of images where automated description is fast, cheap, and reliably good enough to ship with a light human pass.
Product photography on a plain background. A single item, centered, well lit, no scene to interpret. "Red canvas sneaker, side profile" is exactly the kind of literal, closed-set description a model is built for. This is the single best use case for AI alt text at scale, and it's also the case the vendors always demo, which is a little misleading, because it's the easiest possible version of the problem.
Simple, single-meaning icons. A trash can icon, a magnifying glass, a printer symbol. One object, one accepted meaning, minimal ambiguity. Models trained on huge image sets have seen these exact shapes thousands of times.
Generic stock photography. "Group of people in a meeting room" or "hands typing on a laptop" is often literally all a stock image is for. There's no deeper function to miss because the image itself has no deeper function.
Notice the pattern. AI succeeds when the image's job is its literal content. The moment an image is doing something other than showing what it shows, the model has no way to know that, and that's where things fall apart.
Where it lies (or, more fairly, where it confidently guesses wrong)
Decorative versus meaningful. This is the single biggest cause of bad automated alt text on real websites. A model can't tell that a swirl graphic between two sections is purely visual and should get an empty alt="", versus a screenshot embedded in the same layout that carries real information. It describes both, adding noise to screen reader output for the decorative one and sometimes under-describing the meaningful one because it looks visually similar to something ornamental. The W3.org decision tree for image alt text exists precisely because this decision requires knowing the image's role on the page, not just its pixels, and that's a judgment call no model sees enough context to make on its own.
Brand logos. Models frequently either over-describe a logo ("a red and white circular emblem with stylized lettering") when the correct alt text is simply the company name, or they misread stylized wordmarks entirely, especially for smaller or newer brands that weren't well represented in training data. Either way, the output needs a human who actually knows the brand.
Charts and graphs. This is the category where the CHI 2026 finding above really bites. A model can usually tell you "a bar chart with several colored bars," which is true and almost worthless. What a sighted user gets from a chart is the trend, the comparison, the outlier. Reproducing that requires reading the actual data and reasoning about what it means, and that's a fundamentally different task than image captioning. WebAIM's alt text guidance is explicit that complex images like charts often need a longer description or a data table alongside brief alt text, not a one-line caption pretending to be equivalent.
Emotional and cultural nuance. Facial expression, tone, cultural context, and implied narrative are exactly the things a caption model tends to flatten into generic descriptors. "A man looking concerned" might be technically true of a photo that's actually conveying grief, urgency, or fear, and the wrong emotional register changes how the surrounding content lands for someone who can't see it.
A quick honest note from our own testing
When our team ran a batch of client images through a popular AI captioning tool as an internal check, the pattern above showed up almost immediately: clean, confident captions on product shots and icons, and vague or misleading captions on anything decorative, branded, or data-driven. That's not a rigorous published study, just our own hands-on experience reviewing the output before it ever reached a client site, and it lines up with why review matters so much here.
A practical checklist before you trust the output
- Product photos, plain-background items, simple icons: AI draft is usually fine with a quick skim.
- Anything decorative: check it should probably be
alt="", not described at all. - Anything with a brand name or logo: replace the guess with the actual name.
- Any chart, graph, or infographic: rewrite by hand, describing the trend or data point that matters, not the shape of the visual.
- Any image carrying emotional weight or cultural context: rewrite by hand.
That's the dividing line. It's not "AI alt text good" or "AI alt text bad." It's knowing which pile a given image falls into before you hit publish.
If you want the full workflow for building that review step into your process without losing the speed advantage of automation, our guide to using AI for alt text the right way walks through exactly how to structure it.
The takeaway
AI alt text is a genuinely useful first draft tool, not a finished product. Treat its output the way you'd treat a fast but literal-minded intern: great at describing what's there, unreliable at knowing what it's for. A five-minute human pass on the categories above catches almost every real failure before it reaches a screen reader user.
If you'd rather have that pass done for you, you can run a free scan of your current alt text through our free tools and see exactly which images need a second look.
Want a person to walk through the results with you instead of a report? Email experts@wcag.world and our team will help you sort it out. You can also start anytime with our free tools.