Upload an image to almost any modern AI tool and it will hand you a caption before you've finished reaching for your coffee. "A golden retriever running on a beach at sunset." Done. Two seconds, zero effort, and it even sounds like something a person would write. It's genuinely one of the best content life-hacks available right now — until you realize the AI has no idea why that dog is on your homepage, or that the point of the photo was actually the leash brand stitched into the collar.
That gap between "plausible-sounding caption" and "actually useful alt text" is exactly where most teams get burned. AI is a fantastic first draft machine. It is not, yet, a reliable final answer. And the web's largest accessibility dataset just handed us the receipts.
The Stat: 53.1% of home pages have missing alternative text for images — one of the six most persistent failure categories, essentially unchanged for seven consecutive years. (Source: WebAIM Million 2026)
Why AI-written alt text sounds right and still goes wrong
An image model is trained to describe what pixels look like. It is not trained to know what a picture is for on your particular page. Those are two different jobs, and conflating them is where AI-generated alt text quietly fails.
A few ways this shows up in practice:
- It describes the scene, not the function. A photo of a person clicking a button in your app screenshot gets captioned "a person using a laptop" — technically accurate, useless to someone trying to follow a tutorial via screen reader.
- It hallucinates specifics. AI captions sometimes invent brand names, locations, or text that isn't actually in the image, because the model is predicting plausible words, not verifying facts.
- It misses text embedded in the image. Infographics, screenshots, and memes often carry critical text baked into the pixels. A generic scene description ("a chart with bars and text") tells a screen reader user nothing about the actual data.
- It doesn't know when alt text should be empty. Purely decorative images — a spacer graphic, a repeated background flourish — should have
alt=""so assistive technology skips over them. An AI captioner will happily describe them anyway, adding noise instead of removing it. - It repeats itself across a page. If every product thumbnail gets a similarly generic AI caption, you end up in the same "repetitive alt text" bucket the WebAIM Million calls out — technically not blank, but not actually informative either.
That last point matters because the data isn't just about missing text — it's about text that exists but doesn't do its job.
More than 1 in 4 images on popular home pages have missing, questionable, or repetitive alternative text. (Source: WebAIM Million 2026)
And when alt text is missing outright, it's missing at real scale: 16.2% of all images on home pages — an average of 10.8 images per page — had no alt text at all, according to the same report. That's not a handful of edge cases slipping through; that's a systemic gap, and it's one AI drafting tools can help close if a human checks the output before it ships.
The two-second draft, the two-minute check
Here's the actual life hack, and it's not "avoid AI." It's "use AI for speed, use a human for judgment." A workable workflow looks like this:
- Let AI generate a first-pass caption for every image that needs one. This alone is faster than most teams' current process of skipping alt text because writing it feels tedious.
- Ask: is this image decorative or informational? If decorative, override the AI draft with
alt="". Don't let the model talk you into describing a background texture. - Check the draft against purpose, not accuracy. The question isn't "is this a correct description of the pixels" — it's "does this tell a non-sighted user what they need to know to use the page." W3C WAI's alt text decision tree is the fastest way to make that call consistently, image by image.
- Rewrite anything that names a function. Buttons, links, and icons need alt text describing the action ("Submit order"), not the appearance ("blue arrow icon"). AI rarely gets this right unprompted.
- Scan for embedded text. If the image contains words — a quote graphic, a chart, a screenshot — the alt text (or a nearby text equivalent) needs to include that content, not just describe the layout.
- Spot-check for hallucination. If the AI caption names a specific brand, number, or location, verify it's actually in the image before you publish.
| Step | AI does this well | Human still needs to check |
|---|---|---|
| Describing visible objects/scenes | Yes | Verify no hallucinated details |
| Deciding decorative vs. informational | No | Always — set alt="" where appropriate |
| Describing button/link function | Rarely | Rewrite for action, not appearance |
| Capturing embedded text/data | Rarely | Confirm all on-image text is represented |
| Avoiding repetition across a page | No | Vary and tighten wording |
Small syntax details that trip people up regardless of who wrote the caption
Even a perfectly worded caption fails if the markup is wrong. A few basics worth double-checking, whether the text came from a human or a model:
- Every meaningful
<img>needs analtattribute — an image with noaltattribute at all is worse than one with an empty string, because some assistive tech falls back to reading the filename. MDN's documentation for the img alt attribute covers the exact behavior across browsers and screen readers. - Alt text shouldn't start with "image of" or "picture of" — screen readers already announce it as an image.
- Keep it concise. A good target is roughly one sentence; if you need a full paragraph, put the detail in surrounding body text or a caption instead.
- Don't stuff keywords into alt text for SEO purposes. It reads as spam to assistive tech users and search engines increasingly penalize it too.
Where to spend the two minutes you saved
The whole point of letting AI draft the caption is that it frees up human time — so spend that saved time where it counts most. Prioritize review on:
- Homepage and landing page images, since these get the most traffic and the most scrutiny, and the WebAIM Million measured its 53.1% missing-alt-text figure on home pages specifically.
- Product photos and icons that double as controls, where the wrong caption can block someone from completing a purchase or a signup.
- Charts, infographics, and screenshots, where an AI's scene-level description almost never captures the actual data or instructions being conveyed.
- Anything auto-published from a CMS or e-commerce feed, since these are exactly the high-volume, low-attention pipelines where "questionable or repetitive" alt text tends to pile up unnoticed.
None of this requires a specialist. It requires a checklist, five minutes per batch of images, and a habit of treating the AI draft as a starting point rather than a finished product.
The bottom line
AI has made writing a first draft of alt text nearly free, which is genuinely great news — it removes the excuse of "it takes too long." But the WebAIM Million shows that the industry's alt text problem was never really about typing speed. It's about attention: knowing which images matter, what they're for, and whether the description actually helps someone using a screen reader. AI can hand you the raw material. Only a human review step turns that into alt text that actually works.
If you're not sure how your own site stacks up, it's worth finding out before a visitor — or an auditor — finds out for you. You can get a free scan that flags your images with missing or weak alt text and see exactly where the gaps are, starting with a free accessibility audit.