Short answer: Useful AI alt text describes what the image adds to the page, in one short sentence, with no keywords and no "photo of". Give the model the surrounding paragraph, check what it flags as uncertain, and leave the alt empty when the image is only decoration.
Alt text is having a strange moment. Writing it used to be a chore delegated to whoever remembered, which meant it was usually missing. Now AI can write it in a second, which means it is often awful in a new way: keyword-stuffed, caption-flavored, and written for the search engine the author imagines is reading, instead of the person who actually is.
That person is easy to picture. A screen reader user reaches your image and hears whatever you put in the alt attribute. They cannot skim past it, they cannot glance at it, they hear every word. When every word is "premium quality durable ergonomic office chair comfortable seating solution home office furniture", they hear all of that, in the middle of your sentence. Google's own Search Central documentation warns against exactly this: filling alt attributes with keywords results in a negative user experience and may cause your site to be seen as spam.
So this piece has a narrow job: how do you use AI to write alt text a real reader would thank you for, and where does AI reliably misbehave?
What is alt text actually for?
The W3C's Web Accessibility Initiative maintains the reference tutorials on accessible images, and its sorting is the whole game:
- Informative images carry information the page needs: a product photo, a diagram, a chart. They get a short description of the meaning.
- Functional images are links or buttons. They get the function, not the picture: a magnifying-glass icon's alt text is "Search", not "magnifying glass".
- Decorative images add visual flavor and nothing else. They get an explicitly empty alt attribute:
alt="". - Complex images, such as graphs and infographics, cannot fit in an attribute. They get a short alt plus a complete text equivalent of the data elsewhere on the page.
That last bullet is the one most people miss: the alt attribute was never meant to hold a paragraph. For a chart, WAI recommends a full text equivalent of the data nearby, with the alt carrying a summary. The audience is wider than screen reader users: alt text also serves people on connections where images fail to load, and search engines. Google's image SEO documentation calls alt text the most important attribute for image metadata and says Google uses it together with computer vision and page content to understand what an image shows. Useful alt text and search visibility point the same direction. Stuffing breaks both.
How does AI get alt text wrong?
Paste an image into any vision-capable chatbot and say "write alt text", and you will reliably get the same four failure modes.
Caption voice. "A high-quality photo of a modern office chair on a white background." The reader does not need to be told it is a photo. Screen readers already announce "image" before reading your alt text, so "photo of" repeats that fact.
Marketing adjectives. The model has read a million product pages, so it reaches for "stylish", "premium", "sleek". None of those describe anything a person could verify by looking, which means they carry zero information.
Invented details. This is the dangerous one. Vision models describe approximately: they guess between similar shades, miscount objects, and confidently misread small text. Anthropic's own vision documentation says Claude can give approximate counts that "might not always be precisely accurate", and advises developers to always carefully review and verify Claude's image interpretations. Ship an unverified guess and you are describing an image that might not exist, to a person who cannot check. It is the same overconfidence behind AI hallucinations in general, in visual form.
Keywords, the moment you mention SEO. Tell the model the page is about "ergonomic office chairs" and watch it work that phrase into every description. The model optimizes for what you told it, and what you told it was the wrong goal.
There is a fifth, quieter failure: the model never asks whether the image needs alt text at all. Hand it a decorative divider and it produces "abstract gradient wave divider adding elegance and visual flow". For a screen reader user, that is interruption, not information.
Which short prompt makes AI write better alt text?
The fix is a prompt that supplies the two things the model cannot infer: the surrounding page, and the standard to write to. The same principle behind any good prompt library: constraints beat instructions.
Here is an image and the paragraph it sits next to on the page. Write alt text for a reader who cannot see the image. Rules: describe only what the image adds to the page, not what it looks like as a photograph. Under 150 characters. Do not start with "image of" or "photo of". If text inside the image is the point, transcribe the text. Include only details you can clearly see. After the alt text, list any details you were unsure about, so I can check them.
Every clause earns its place. The neighboring paragraph matters because the same photo needs different alt text beside different text: next to a caption reading "our most durable chair", the alt should skip "durable" and describe only what the picture shows. The character cap forces summary judgment. The transcribe rule catches screenshots and charts whose labels are the message. The last line converts the model's biggest weakness into a worklist: it surfaces what to verify instead of hiding it.
Then verify: open the image and check the flagged details. The model is a drafting tool with no accountability; the person publishing the page is describing the image to a real reader.
What should the alt text say for five common images?
A product photo. Navy office chair, studio shot, white background. Bad AI output: "High-quality photo of a stylish, comfortable ergonomic office chair with premium materials, perfect for home office and workspace." That is ad copy, and it steals words your caption should own. What you would ship: "Navy office chair with a high back and armrests, shown from the side." It tells the reader the one thing the product copy does not: what the object looks like.
A chart. Bar chart of monthly signups. Bad: "Bar chart displaying monthly data with an upward trend across categories." The reader learns nothing, not even the direction, and models sometimes misread small axis labels. What you would ship: a short alt carrying the takeaway, "Bar chart: signups grew every month from January through June", plus the actual numbers as text or a table in the page. The attribute holds the headline; the data lives as data.
A screenshot. App settings page, cropped to one toggle. Bad: "Screenshot of the settings interface showing various options and menu elements." What you would ship: "The Auto-save toggle in Settings > Documents, off by default." The model can read the toggle's label; only you know why the screenshot exists, and the point of the image is the alt text.
A meme. Memes need alt text too, and they are the hardest case. A stock photo of a distracted boyfriend is not informative because of the man, the woman in red, or the upset girlfriend; it is a template making an argument. A literal description hands the screen reader user the scene and none of the joke. What you would ship: the template's name plus your specific mapping, "Distracted boyfriend meme: the boyfriend eyes 'the shiny new AI tool' while 'the tool you already pay for' glares." AI can recognize the template. The mapping comes from your paragraph, not the pixels.
A decorative divider. A gradient wave separating two sections. Bad: anything at all. What you would ship: alt="". The W3C's alt decision tree is unambiguous here: if the image is purely decorative, the correct text alternative is an empty one. Note the difference between empty and missing: alt="" tells the screen reader to skip the image entirely, while a missing alt attribute often makes it announce the filename, the worst of the three outcomes.
When should the alt text be empty?
Decoration is one case for silence. The decision tree gives another: the image is redundant with real text nearby. A conference banner reading "ToolbitConf 2026, Lisbon, June 12", sitting under a headline that says the same thing, should be silent. Reading it twice helps no one.
AI will not make this call. The informative-versus-decorative judgment depends on what the rest of your page says, and a model given only the image defaults to describing, because describing is what it does. The decision is yours.
How should you use AI for alt text?
AI is genuinely good at one part of this job: looking at an image and drafting a factual description faster than you can type. Use it for exactly that, with a tight prompt, the surrounding paragraph, and a verification pass. What it cannot do is decide what the image is for on your page, and that decision is the actual skill. Every failure above, from stuffed keywords to poetic dividers, comes from skipping it. Keep the alt to one sentence under about 150 characters, so it can be heard in one breath, and put full chart data in the page. Make the call, then let the machine help you phrase it. If you want a starting point, try an AI alt text generator and run its output through the checks above.
