Image Accessibility Needs Useful Alt Text, Not Automated Descriptions Alone
Image accessibility begins with a deceptively simple question: what does this image contribute here? Useful alt text answers that question for someone who cannot see the image, while also supporting people whose connection fails to load it. Automated descriptions can help identify visible objects, but they cannot reliably decide why the image was chosen, what matters in this context or whether it should be announced at all.
That distinction matters as organisations publish more content, use larger image libraries and introduce artificial intelligence into everyday authoring. A system may produce fluent descriptions at speed, yet fluency is not the same as usefulness. If the description adds irrelevant detail, misses a functional purpose or repeats nearby text, the page may technically contain alt text while still creating a poorer experience.
Image accessibility begins with purpose
The same photograph can need different treatment in different places. A picture of a crowded train might illustrate peak-hour demand in a transport report, set a mood on a campaign page or act as a link to service updates. The pixels do not change, but the information a user needs does. Image accessibility therefore cannot be separated from the purpose of the surrounding content.
The Australian Government Style Manual distinguishes decorative, informative and functional images because each category calls for a different response. An informative image needs a concise equivalent of the meaning it conveys. A functional image needs text that explains the action or destination. A decorative image should usually be ignored by assistive technology through null alt text. One generic instruction to describe every image cannot handle those differences.
Alt text is a content decision
Alt text is sometimes treated as technical metadata added after design and writing are complete. In practice, it is part of the content. It requires an author to understand the message, the audience and what a person would lose if the image were unavailable. That judgement belongs alongside decisions about headings, labels, captions and plain language, not at the end of a publishing checklist.
A useful description is often shorter than people expect because it focuses on meaning rather than inventory. If a photograph supports a story about an accessible entrance, the relevant detail may be the step-free path and automatic doors, not the colour of the sky or every object near the building. More words do not automatically create more access. They can force a screen reader user through detail that sighted readers would never consider important.
Automated descriptions cannot replace editorial judgement
AI tools are increasingly able to recognise people, objects, text and settings. That makes them valuable as a starting point, particularly when an organisation has a large backlog of unlabelled images. They can prompt an author to notice details and reduce the effort involved in drafting. The risk begins when a generated description is accepted as the finished answer without checking the image's role.
The Australian Human Rights Commission's 2025 guidelines acknowledge the use of AI to generate image descriptions but caution that AI applications are not yet reliable in most situations. A model may confidently misidentify an object, infer a person's identity or emotion, overlook text embedded in an image, or describe visual details that distract from the intended message. It also cannot know the organisation's communication intent unless that context is deliberately supplied.
Automation should support a decision, not make it invisible. Authors still need to decide whether an image is informative, functional, complex or decorative, then review the proposed wording against nearby content. The stronger workflow uses AI to suggest, a person to interpret, and testing to confirm. It does not make a generated sentence the definition of compliance.
Decorative images need silence, not descriptions
A common misconception is that every image should have a written description. Decorative images are the important exception. If an image adds no information and has no function, null alt text allows a screen reader to skip it. Describing a decorative flourish, background texture or redundant stock photograph adds noise and makes the content harder to navigate.
This is another area where automation can cause harm. A model will often describe whatever it can see, even when the best accessible outcome is silence. It may turn a simple visual separator into a detailed announcement or repeat a message already expressed in the heading beside it. The W3C alt decision tree starts with purpose for precisely this reason: the right alternative is not always more text.
Complex images need more than alt text
Charts, diagrams, maps and process models often contain too much information for a short attribute. Compressing an entire chart into one long sentence makes the content difficult to understand and awkward to navigate. Short alt text should identify the image and its purpose, while the significant data, relationships or conclusions appear in nearby body text, a table or an extended description.
This approach also improves content for people who can see the image but may struggle to interpret it. A written summary can explain the main trend, the comparison that matters and the decision the image supports. Accessibility does not require translating every visual mark into prose. It requires ensuring that the essential information and function are available in a form people can perceive and use.
Context changes what useful alt text says
Consider a staff photograph on three pages. On an organisational profile, names and roles may be important. In an article about workspace design, the relevant information may be how people are using the space. On a decorative banner, the photograph may need null alt text because the page already communicates everything it contributes. Reusing one stored description across all three contexts may be efficient, but it is not necessarily accurate.
Consistency should apply to the decision process, not force identical wording into every use. A media library can store a factual base description, especially for distinctive branded assets, but authors should confirm whether the current page needs that text, a contextual variation or silence. This is particularly important when an image carries text, links somewhere, changes meaning through its caption or appears beside information that already describes it.
Build image accessibility into the publishing workflow
The most reliable improvement is to give image accessibility a clear owner before publication. Designers can identify the intended purpose, content authors can write the equivalent message, and developers or platform specialists can ensure it is implemented correctly. When responsibility is left undefined, alt text is often copied from a filename, generated automatically or omitted because everyone assumes someone else will add it.
Record the image's purpose: State whether it is informative, functional, complex or decorative before drafting any description.
Provide the surrounding context: Give the author the heading, caption and nearby text so the alternative does not repeat or contradict the page.
Keep informative descriptions concise: Include the meaning needed in this context and leave incidental visual detail out.
Separate complex information: Use nearby prose, data tables or extended descriptions when a short alternative cannot carry the full message.
Review generated suggestions: Check accuracy, relevance, privacy and tone before accepting any AI-assisted description.
These steps are small enough to add to existing publishing systems. The harder work is cultural: recognising that alternative text is not a field to be filled, but an editorial decision with a real audience. Templates and automated checks can confirm that a field exists. They cannot confirm that its content makes sense.
Test image accessibility with people and real content
Automated accessibility testing can detect missing alt attributes and flag some suspicious values, such as filenames. It cannot reliably judge whether a description communicates the right meaning. Quality review should therefore include reading the page without the images, listening with a screen reader and asking whether the same task or understanding remains possible.
Testing should include people who use assistive technology, particularly for important services and complex visual content. Their experience can reveal repetition, ambiguity and cognitive load that a checklist misses. It can also show when the nearby body copy needs improvement rather than a longer description. The objective is equal access to the information and function, not a perfect verbal reconstruction of every visual detail.
Useful alt text improves more than compliance
When organisations focus only on the presence of an alt attribute, they miss the larger opportunity. The act of writing a useful alternative exposes unclear visual communication, decorative clutter and images that are carrying information they should not carry alone. It encourages teams to ask what the content is trying to do and whether that purpose is available to everyone.
Image accessibility needs useful alt text, but useful does not mean automatically detailed or universally reusable. It means accurate, contextual and proportionate to the image's role. AI can accelerate the first draft, but human judgement must decide what matters. The measure of success is not whether every image has words attached to it. It is whether people receive the same essential meaning and can complete the same task.
