A photograph of a bicycle can do several jobs. In a shop, it might show the shape of the frame. In a repair guide, it might reveal a damaged cable. In a travel essay, it might establish how someone crossed a city. Ask an AI to describe only the picture and all three uses can receive the same competent sentence.
That is where a small editorial task becomes an exercise in judgment. Useful alternative text connects an image to the page around it. The work is less about naming every visible object than deciding which information a reader would otherwise miss.
The W3C Web Accessibility Initiative's image decision tree makes context central. It distinguishes images containing text, images used as controls or links, informative pictures, and purely decorative material. That is a more useful starting point for AI assistance than a command to describe everything.
Tell the assistant what the image is doing
Begin with the image, its caption, the surrounding paragraph, and its role. Is it evidence, an illustration, a navigation control, or atmosphere? Supplying that information helps the assistant work toward the reader's task.
Consider an invented repair-page example. The page explains where a bicycle's rear brake cable has frayed. A broad description of a silver bicycle against a brick wall misses the reason the image is there. A candidate description might instead identify the frayed cable beside the rear brake. A person must check that the damage is actually visible and the named part is correct.
For a travel essay using the same photograph, those mechanical details may be irrelevant. The useful context might be the bicycle locked outside the station where the journey begins. Context changes selection; it does not license invention. If the source does not establish the station's identity, the assistant should leave it unnamed.
Ask for two concise alternatives and the information each preserves. Comparing the options can help the editor decide what belongs in the text without treating the first generated sentence as finished.
Sometimes the best description is empty
An ornamental image can become an interruption when its description repeats material already present or contributes no information. W3C's decorative-image guidance recommends an empty alternative, expressed in HTML as alt="", for these cases. It also explains why omitting the attribute is different: some screen readers may announce a file name.
The practical question is what the reader loses if the image disappears. A texture behind a title may contribute visual character without requiring an audible description. A photograph that supplies evidence needs different treatment. A person should make that contextual call rather than asking the model to fill every empty field.
Linked images need another kind of attention. If the picture is the only label for a destination, the text must help someone understand where the link goes. Describing its colors may be much less useful than naming the action. Check the whole interactive element, including visible link text, before deciding whether an additional description adds information or repeats it.
Charts need their underlying information
Charts expose the limits of a one-sentence assignment. A useful short description may identify the chart, but the reader can still need its values, relationships, and main pattern. W3C's complex-image guidance describes pairing a short identification with a longer text alternative and offers several ways to make that detail available.
Give the AI the source table as well as the picture. Ask it to separate what the data says from what the editor thinks the pattern means. A model reading a low-resolution screenshot should not become the authority for small numbers or axis labels.
For an illustrative chart comparing three stores, the longer explanation could describe the period, measurement, largest change, and any exception. The underlying table lets readers inspect exact values. Keep claims about causes out unless other evidence supports them. A rising line can show a change without explaining why it occurred.
The longer description can help readers who see the image too. Someone unfamiliar with the topic may need the same plain-language explanation. Accessibility work often reveals where an article has assumed that a picture will do too much teaching by itself.
Read the result as part of the story
Finish by reviewing the page in reading order. Listen with assistive technology where available, and check keyboard access when an image is interactive. Compare the description with the caption and surrounding text. Repetition, a missing destination, or an unexplained chart can be easier to notice in sequence.
This method does not establish accessibility conformance, and this publication has not tested a particular image model. It gives editors a concrete way to use AI for candidate wording while keeping purpose, factual accuracy, and the reader's experience in human hands.
The final question is generous and demanding: after this image is translated into words, can the reader still follow the story? That is a better target than a description that merely proves the machine can see a bicycle.
Sources and limitations
- W3C WAI, “An alt Decision Tree”: checked October 15, 2026. Supports selecting an image alternative according to purpose and context; not an evaluation of AI-generated descriptions.
- W3C WAI, “Decorative Images”: checked October 15, 2026. Supports empty alternative text for decorative images and the distinction from omitting the attribute.
- W3C WAI, “Complex Images”: checked October 15, 2026. Supports short and long alternatives for complex visual information. Implementations require context-specific review.