Can AI Write Good Alt Text? An Honest Answer

We build an AI alt text tool, so you'd expect this post to say "yes, perfectly, use ours". The honest answer is more useful: AI writes good alt text for most images, excellent alt text for some, and confidently wrong alt text for a predictable few — and knowing which is which is what makes the workflow safe.

Here's where the line sits, from watching thousands of images go through our Alt Text Writer.

Where AI Is Genuinely Good

• Straightforward photos. Products, people, places, food, screenshots of recognisable apps. Modern vision models describe these accurately and — importantly — already follow the rules humans forget: front-loaded subject, no "image of", sensible length.
• Volume. This is the killer argument. The alternative to AI alt text on a 400-image media library is not lovingly handcrafted alt text; it's no alt text. AI at 90% quality beats empty attributes at 0% every single time.
• Consistency. Humans write great alt text at 9am and "photo" by 4pm. The model's hundredth description is as careful as its first.
• Visible text. Good models reliably transcribe text inside images — the "50% off until Friday" banner problem — which humans skip surprisingly often.

Where AI Fails, Predictably

• Context only you have. The model sees "woman presenting to a room". It cannot see "our CEO announcing the Colombo expansion". The most valuable alt text on important images is made of facts that aren't in the pixels — names, events, significance. AI can't supply what only you know.
• Purpose. Whether that decorative swoosh deserves alt="", or whether the linked logo should say "Cocoon — home" rather than describe the logo, depends on what the image is for on the page. Function is invisible in the image file. (The rules are in our alt text guide.)
• Charts and data. Models describe what a chart looks like ("bar chart with six green bars") when the reader needs what it says ("AI job postings doubled in six months"). The takeaway needs a human who knows which finding matters.
• The confident miss. Rarely, the model misreads an image — an unusual angle, a niche object, a culturally specific scene — and describes the wrong thing fluently. Rare, but it's why "generate and publish unreviewed" is the one workflow we tell people not to use.

The Workflow That Gets It Right

AI drafts, human seasons. In practice, per image:

1. Generate a draft with the Alt Text Writer — 3 seconds
2. Check it against the image — is anything simply wrong? — 3 seconds
3. Add what only you know — a name, the event, the chart's takeaway — 5 seconds where needed
4. Zero out decorative images with alt="" — instant

Ten seconds an image instead of a minute of staring at a blank field. For a media library, that's the difference between "accessibility project we'll do someday" and "done by Friday". And because honest, accurate descriptions are exactly what search rewards, the same pass quietly improves your image SEO — see Alt Text and SEO: What Google Actually Rewards.

The Verdict

Can AI write good alt text? For the bulk of ordinary images — yes, better than most humans bother to. For the images that matter most, it writes a strong first draft that needs your ten seconds of context. The failure mode isn't using AI; it's using AI unreviewed, or using its limitations as an excuse to leave 400 images silent.

Try it on your trickiest image and judge for yourself: the Alt Text Writer is free, browser-based, and doesn't ask for an account. It's part of Cocoon's accessibility toolkit, alongside SignBridge and our A11y Widget.