Write better alt text with AI: start with the image's job

The same photograph can need different words

A close-up of a canvas bag can do several jobs. On a product page, it might show a zippered inside pocket. In a sewing tutorial, it might show where a seam meets the pocket opening. In a decorative banner, it might add atmosphere without contributing information that the page needs. A description based only on the picture misses those differences.

That is the limitation I would address before asking AI to generate alt text. The assistant needs to know why this image is here. Otherwise it may return a fluent inventory of colors and objects while leaving out the one detail that helps someone understand the page.

Alt text is a text alternative associated with an image. It is not automatically the same as the visible caption, the product specification, or the page's search description. Those pieces of writing can support one another, but they have different jobs. Combining them into one promotional sentence usually makes the alternative less useful.

W3C's image decision tree organizes the choice around information, function, redundancy, and decoration. I would use that distinction to frame the AI request rather than assume every uploaded image needs a descriptive sentence of the same length.

This article uses fictional image scenarios and proposed wording, not results from a named AI tool or a live accessibility audit. There are no unseen product photographs being analyzed here. The examples show the editorial decisions to make when the actual image and its context are available.

The objective is not to describe more. It is to preserve what a reader needs when the visual information is unavailable or difficult to perceive.

Sources: W3C: the alt decision tree

Give the assistant the page, not just the picture

For a useful first draft, I would supply four things: the actual image, the section's purpose, the nearby text, and whether the image is interactive. The same file can appear in several places, so include the particular placement being edited. A filename alone rarely provides enough context to make the decision responsibly.

Imagine our fictional canvas bag appears under a heading called Inside storage. The paragraph already says that the bag has one internal pocket, while the image shows the pocket's zipper open. That context suggests a description about the opening and what the view reveals. It does not justify calling the bag waterproof, lightweight, or suitable for every laptop.

Separate what is visible from what comes from an approved specification. If dimensions or materials matter to the surrounding content, provide their verified source. Do not ask the assistant to estimate capacity from perspective or infer a fabric certification from appearance. A plausible product attribute is still an invented attribute if the evidence does not establish it.

Also supply the visible caption, if there is one. A caption might explain the photographer's point while the alternative supplies a visual detail. Or the caption might already carry all the information, leaving the image redundant in that context. The assistant should explain its reasoning before suggesting an empty alternative, especially where the image is clickable.

When the image cannot be opened or its text is unreadable, the right result is a review question, not a polished guess. Ask for a better source or provide a verified transcription. A system that fills every field without acknowledging missing evidence is completing a batch, not necessarily improving the page.

Help draft a text alternative for this image in this specific placement. Image: [ATTACH IMAGE]. Page purpose: [PURPOSE]. Nearby heading, text, and caption: [TEXT]. Link destination or button action, if any: [ACTION OR NONE]. Verified facts that are relevant but not visually inferable: [FACTS AND SOURCE, OR NONE]. First identify the information or function the image contributes. Then suggest an alt value, or explain why an empty alternative is appropriate. Keep uncertain observations and review questions outside the proposed alt text. Do not invent image details, product specifications, identities, or unreadable text.

A product photo needs a distinction, not a sales pitch

For the fictional bag, suppose the image is an unlinked detail photograph showing an open zippered pocket inside an otherwise plain lining. The surrounding page already identifies the product. A candidate alternative is Interior pocket with its zipper open, showing the plain lining. That wording earns its place by explaining the detail view.

Compare it with Premium durable travel bag for professionals and students. That sentence may sound commercially useful, but it contributes no reliable description of the photograph. It also introduces judgments about durability and audience suitability that the image does not prove. I would remove those claims rather than try to make them sound less promotional.

Another weak draft would describe the whole scene: a bag on a table beside a plant, with a light wall and a soft shadow. Those details might be visible, but they distract from the pocket unless the page is specifically discussing the setting or composition. Relevance is not the same as exhaustiveness.

For a gallery, ask what each view adds. A front view, a pocket detail, and a shoulder-strap attachment may need different descriptions. Repeating the product name three times without distinguishing the views gives a reader little help. Conversely, inventing differences to make every description unique is worse than admitting that two images communicate the same thing.

Keep the alternative aligned with the accepted image version. If the picture is later cropped so the zipper is no longer visible, the sentence needs review. Treat the image and its description as a pair. Approval of the wording should not automatically survive a replacement image with the same filename.

Context: an unlinked pocket detail beneath Inside storage.
Visible detail supplied for the exercise: an open zipper and plain lining.

Reject: Premium waterproof travel bag for every adventure.
Reason: promotional and unsupported by the supplied view.

Candidate: Interior pocket with its zipper open, showing the plain lining.
Review: confirm both details in the actual accepted photograph before use.

Sometimes the right alternative is deliberately empty

Now move a cropped texture photograph into a banner where its only purpose is decoration. The heading and body already convey the complete message, and the image is not a link or a control. Writing a detailed texture description would add information that the reader does not need to complete the page's task.

For a genuinely decorative HTML image, W3C's decision tree recommends an empty alt attribute. MDN distinguishes an empty alternative from omitting the attribute: absence does not communicate the same intent and can result in an unhelpful filename being exposed. This is an implementation choice to verify, not merely a blank cell to leave for somebody else to interpret.

In a content system, ask what its decorative setting actually outputs. Does it create an empty alternative? Does another component add an accessible name anyway? The editor's form and the rendered page are not always identical. A reviewer should inspect the result instead of assuming a cleared input has produced the intended markup.

I would be cautious about declaring an image decorative solely because it looks attractive. A photograph showing a venue's step-free entrance may be essential information. A diagram with no visible caption may explain a process the paragraphs do not cover. The relevant question is what meaning would disappear if the image were removed.

The same caution applies to interactive images. An image-only link cannot simply lose its accessible name because the picture is decorative. Review the link as a complete control, including any real text or other naming mechanism. Empty alt is useful when justified by context; it is not a shortcut for avoiding a difficult description.

Sources: W3C: informative, redundant, and decorative decisions; MDN: the image element and alt attribute

For a linked image, ask where it takes the reader

Consider a fictional icon showing a downward arrow over a document. If it is the only content of a link to a workshop worksheet, an alternative such as Black arrow and white rectangle is not useful navigation. The reader needs to know what the link offers, not how the icon was drawn.

W3C's guidance for functional images emphasizes the action or destination rather than visual appearance. For this example, a label such as Download the workshop worksheet may fit, provided the link really does download that worksheet. If it opens a preview instead, the name should reflect that behavior.

Now place the same icon inside a link that already contains the visible words Download the workshop worksheet. The icon may add nothing to the accessible name. Repeating the whole phrase through the image alternative could make the control unnecessarily repetitive. This is why the assistant must see the complete link, not just the icon asset.

The reviewer should inspect how the final control is named. Text, image alternatives, and attributes can interact, so do not approve each piece in isolation. The result should be understandable once, as a coherent action. The goal is not to maximize the number of places where the same phrase appears.

Ask about mobile behavior too. If visible words disappear at a narrow width, make sure the remaining control still has an appropriate accessible name. A decision that was correct for the desktop layout may become incomplete when the layout changes.

For image galleries, also distinguish a product link from a button that opens a larger image. View pocket detail and Open bag product page describe different tasks. AI can help write both, but the implementation determines which task actually happens.

Sources: W3C: functional images in links and buttons

A screenshot should support an instruction, not replace it

Suppose a tutorial includes a fictional settings screenshot. The relevant detail is that Weekly digest is selected while Daily alerts is not. An AI description that inventories every menu, avatar, and toolbar item makes the instructional point harder to find.

I would first put the actual instruction in normal page text: Select Weekly digest in Notification frequency, then save the change. The screenshot can confirm what the selected state looks like. A candidate alternative might be Notification frequency set to Weekly digest. Whether that adds useful information depends on the surrounding text and what the screenshot contributes beyond it.

Do not rely on the image as the only location of an important instruction or error message. If the screenshot contains a warning that the tutorial discusses, transcribe the relevant wording accurately and make it available as text. Do not ask the assistant to reconstruct blurred words from what such a warning usually says.

Keep a distinction between what the interface displays and what the action accomplishes. A screenshot of a Saved notification shows that a message was displayed; it does not independently prove that the underlying setting persisted. The image description should not turn a visible state into an unverified technical conclusion.

Before sharing screenshots with an assistant, remove information that is unnecessary for the writing task, such as private messages or account details, using an approved workflow. Then check that the redaction did not obscure the exact control the article needs to explain. Privacy and accuracy both depend on providing an appropriate source.

Finally, record which interface version the screenshot belongs to. When the tutorial is updated, verify the picture, its alternative, and the written instruction together. Correct words attached to an outdated screenshot can still mislead the reader.

Sources: W3C: deciding how to handle meaningful image text

A chart needs its evidence, not a squeezed-down essay

A chart can carry more information than a short alternative should attempt to hold. W3C's complex-image guidance describes a two-part approach: a short description that identifies the image, and a longer text equivalent for its essential information. That is a better starting point than asking AI to compress every label and value into one enormous sentence.

Our fictional chart compares workshop registrations across three months: April has 40, May has 55, and June has 45. A short candidate description is Workshop registrations peaked in May; monthly values and explanation follow. That last phrase is appropriate only if the promised accessible content really follows the chart.

The longer explanation can state that registrations increased from 40 to 55, then declined to 45. June remained five above April. A nearby properly marked-up table can provide all three values with their month labels. If the chart includes a meaningful target line, uncertainty interval, or additional series, the text equivalent must account for that information too.

Do not let the assistant turn the chart into a causal story. These three numbers do not establish that a marketing campaign caused May's peak. Nor do registrations equal attendance, revenue, or unique customers without further definitions. A useful description preserves the metric as well as the numbers.

Provide the source data when asking AI for a chart description. A small image can make labels difficult to read and close values difficult to distinguish. If the chart and the supplied table disagree, flag the discrepancy instead of silently choosing whichever version produces the cleaner narrative.

The fictional values below are for teaching only. Before publication of a real chart, verify the complete visual against its data source and inspect how the short alternative connects to the longer explanation. Merely adding a table somewhere else on the site does not make the relationship clear to a reader.

Metric: workshop registrations, not attendance.
April: 40
May: 55
June: 45

Short alternative candidate: Workshop registrations peaked in May; monthly values and explanation follow.

Longer explanation: Registrations rose from 40 in April to 55 in May, then fell to 45 in June. June was five above April. These figures do not explain why registrations changed.

Implementation requirement: provide the values as accessible page text or a properly marked-up table, not another screenshot.

Sources: W3C: short and long alternatives for complex images

Shorten the draft by removing the wrong details

When an AI draft is too long, I would not begin by giving it an arbitrary character target. I would identify which information is necessary and which details merely describe the scene. That keeps the revision focused on meaning rather than chopping the last words off a useful sentence.

W3C's tips recommend concise alternatives that reflect purpose, with longer descriptions for information that needs more space. They also recommend placing important information first and avoiding generic image or picture wording when it adds nothing. Those principles are more helpful than a universal length rule applied to every image.

For the bag detail, Interior pocket with its zipper open leads with the relevant object and state. There is a photograph of an item on a surface delays that information and tells the reader little. But the medium can matter in some contexts: an illustration proposing a room layout should not be described as if it were a photograph of a completed room.

Remove unverified emotional and commercial language. Beautiful, luxurious, effortless, and perfect for everyone usually add evaluation rather than information. If a person's expression is relevant, describe only what the image reasonably supports and avoid inventing a private emotional state or identity.

Also check what the description repeats. Repetition may be necessary when the alternative needs to stand in for important image content, but it should be deliberate. Do not mechanically paste the caption into every alt field or assume that nearby text always makes the image redundant.

Read the surrounding paragraph and the proposed alternative in sequence. Does the image's contribution become clearer? Does the sentence introduce a fact the page cannot support? This editorial pass is useful before any technical accessibility check, although it does not replace that check.

Sources: W3C: concise, purpose-led text alternatives

Batch by page placement, not by filename alone

Bulk generation is tempting when a site has hundreds of images. The risk is approving text at the asset level when the correct alternative depends on each use. A photograph that explains a feature in one article may be redundant in a card elsewhere. The filename can remain the same while the editorial decision changes.

I would start with a small, varied batch: one informative photo, one decorative use, one functional image, one screenshot, and one chart. These cases expose different decisions. Five nearly identical product photos can make a weak process look reliable because it has not faced a genuinely different task.

Give every proposed edit a page URL, a placement identifier, the current value, and a proposed value. Keep the reason and any unresolved question in separate review fields. A blank proposed value should come with an explicit decorative or redundant decision, not be indistinguishable from a failed generation.

Do not overwrite existing human-reviewed descriptions by default. Some may contain context absent from the image or the model's input. Compare them with the current page, and investigate before replacing a specific useful description with a more fluent generic one. The purpose of the batch is improvement, not making every field look newly written.

For uncertain cases, leave the edit pending. A poor-quality chart, unreadable screenshot, or unexplained link destination needs clarification. Report the number of unresolved placements rather than making the batch appear complete by filling them with guesses.

Keep changes reviewable and reversible in the publishing system. The reviewer should be able to see which placement changed and why without reconstructing a long AI conversation. This also helps when another editor is updating the same article or replacing an image at the same time.

Prepare a review-only batch of image text alternatives. For each supplied page placement, return: page URL, placement identifier, image purpose, existing alt value, proposed alt value, reasoning, evidence checked, and unresolved questions. An intentionally empty alternative must be labeled as a deliberate decision. Do not overwrite existing values or publish changes. Flag unreadable content, unclear link actions, missing page context, and descriptions that need a longer text equivalent. Review each placement even when several placements use the same image file.

Check the actual page before calling the work finished

A good sentence in a review document is not yet a good implementation. After an approved edit, inspect the page that readers will use. Confirm that the intended image has the intended alternative, and that the content system has not inserted a filename, duplicated text, or applied the value to every occurrence of the asset.

For a decorative image, inspect that the intended empty alternative survives rendering. For a linked image, check the accessible name of the complete link. For a chart, verify that the longer description or data table exists, is understandable without the image, and is connected clearly to the short description.

Also check responsive layouts and image replacements. A crop can remove the detail named in the alternative. A mobile control can lose the visible words that previously supplied its name. An updated chart can keep an old description even after its data changes. These are maintenance risks, not reasons to abandon text alternatives.

A screen-reader check can help assess the experience, particularly for image-only controls and complex content. Use qualified accessibility review where appropriate. An AI assistant's assertion that all images are accessible is not evidence of conformance, and this article does not provide a certification checklist.

My final editorial question would be simple: does a reader receive the information or action that this image contributes, without being asked to accept invented details? For the bag, that means the relevant visible feature. For the worksheet icon, it means the destination. For the chart, it means the values and their meaning, not a guessed explanation for the trend.

AI is useful for proposing language and surfacing questions across a large collection. The human contribution is context, evidence, and judgment about the reader's task. Keeping those responsibilities visible is how an alt-text project becomes a genuine improvement rather than another batch of fields marked complete.