AI writes alt text: how to check whether it conveys the information people need
How to review alt text generated at scale against the image role, factual attributes and actual place of publication.
Article language: English

In this article
A model saw a photograph of a wooden organizer and wrote: “An eco-friendly organizer handcrafted in a Polish workshop.” The sentence is so persuasive that you can almost smell freshly planed wood. Yet the file establishes neither its origin nor how it was made, and the material might be veneer. A fluent description is not necessarily a verified one.
When alt text is generated at scale, two things need separate checks: whether the image was recognized accurately and whether the information is useful where it appears. A model may name an object correctly while saying too much, too little, or something that does not help someone use the page. A detailed photo caption, meanwhile, is not always suitable for an alt field, the text alternative for an image.
Review the whole workflow: the image, page context, instruction to the model, output and publication decision. Asking only whether a sentence “fits the photo” overlooks the image's function. It is like accepting a signpost because its arrow is drawn straight without asking where it points.
The same file can serve different purposes
Imagine a fictional photograph of an organizer on a desk. In one place it decorates a general introduction to a workshop. In another it is the main image for a particular product. In a third it shows the position of a divider in an instruction. The pixels are the same, but the information a reader needs changes with the task.
For decoration, a detailed description may repeat an atmosphere without adding necessary meaning. On a product page, identifying the pictured variant may matter. In instructions, the reader may need to understand the arrangement of parts or the result of an action. One automatically generated description permanently attached to the file cannot account for these differences.
The W3C alt decision tree relates the choice of a text alternative to the image's function and meaning in context, including an empty alternative for decoration. This gives the review a concrete basis. An AI workflow should therefore receive information about how the image is used instead of working from the file alone.
Record the context in the input: the type of placement, nearby text, what an associated link does, and what the reader needs to understand from the image. You do not need to send the entire website, only the material needed for the established role. The more general the instruction “describe the picture,” the easier it is to get an accurate caption that still does not serve the page.
For an image linking to a product, check the name of the entire link in the finished interface. Alt text may be read together with a visible name or become the only description of the destination. A text file alone will not show that relationship. Acceptance must cover the published placement and the actual element structure.
Do not confuse alt text with a caption visible to everyone. A caption can explain the variant, attribution or context of a photo shoot. The alternative should convey the meaning needed to use the image. If the workflow generates both fields, establish their roles and check how they are read together. Two identical long sentences may produce needless repetition.
The image's role is determined when selecting brand photography for a website, a decision discussed in the article on choosing brand photography (in Polish). In an automated workflow, that decision should travel on to the description process. Otherwise the model has to guess a decision the team already made but stored elsewhere.
Azure Vision documentation on image descriptions describes generating proposals with confidence scores and ranking them by those scores. That is information about an image-description mechanism. It does not establish that a proposal suits the function of a particular place on a page.
Do not use the confidence score alone as the condition for publication. Assess object recognition, unsupported additions and fitness for the role separately. These are different questions. A high score for recognizing an organizer does not establish that the description explains how to fit its divider.
Build a small set of cases that exposes different errors
Start with images representing real uses. The collection does not have to be enormous, but it should contain more than a few easy photographs of a single object against a plain background. Include similar variants, a detail without the whole object, an element containing text, and an image whose meaning emerges only from the adjacent paragraph.
Prepare an assessment card for each: confirmed information, role, surrounding text and additions that are not allowed. Do not make one exemplar sentence the only acceptable answer. The model may convey the meaning accurately in other words. Criteria should assess the information, not the reviewer's literary taste.
For the fictional organizer, distinguish three kinds of claim. The first concerns what can be seen, such as the number of visible compartments or the object's position. The second comes from the approved product record, such as material and dimensions. The third is unsupported inference, such as environmental credentials, origin or durability. The result should let you identify the basis for every material claim.
A detail tests overconfidence. A close-up shows part of a joint and an edge. “A sturdy fastening for years of durability” goes beyond recognizing the image. Hold such a result for review rather than treating the praise as harmless style. Product performance claims need their own evidence.
Similar variants test assignment. If a file shows two compartments but the product page concerns a version with three, the text will not fix the contradiction. Establish the right image or the right catalog item. Correcting one word in the alt may merely hide that the sources describe different products.

When a graphic contains text, pay attention to small lettering and symbols. A model's reading may need separate confirmation. “K22” and “K2Z” differ little visually, but can point to different sections of the catalog. If the code matters to the task, compare it with the approved source. Do not assume a fluent sentence confirms every character.
In an instruction, check the relationship between the action and the image. The description should help explain the current step rather than return to the product's general name. If understanding the joint requires additional explanation in ordinary text, settle that during page editing. Do not try to fit the entire instruction into a single text field merely because the CMS provides one.
When an image contains a label with data or text, also check what may be transcribed. A model can carry a detail into the description that was not intended for publication. The automated removal of data needs a separate review, as explained in the guide to checking automated data removal (in Polish). Omitting a surname from alt text does not remove it from the visible image.
Add an unreadable or ambiguous example. The team should decide in advance when a result goes to a person and what uncertainty remains in the record. This is not a call to publish a lengthy “perhaps this depicts...” It is a way to stop automation when the necessary conclusion lacks a basis.
A similar problem in a voice conversation concerns a misheard code. The distinction between recognizing input and accepting its meaning appears in the article on reviewing AI voice conversations (in Polish). With images too, distinguish the technical analysis result from confirmation of the right product and role of the information.
Accept the publishing workflow, not just a table of sentences
For work at scale, establish a link between the result and the specific use of an image. A file address may not be enough if the same photograph is used as decoration and in instructions. The record should allow the page, placement and material version to be reconstructed. Otherwise correct text can easily end up on the wrong element.
Separate an intentionally empty alternative from an unprepared result. Both may look like a blank field in working data, but they require different handling. The first should record a decision that the image is decorative; the second should record a reason for stopping or a processing error. Otherwise automation may fill in a deliberately omitted description or treat an unprocessed image as accepted. Review the decision status, not just text length.
Check missing results and partial processing. If the model produces no description for one item, the system should not replace it with a random filename or an old text from another image. Handling depends on the role and publication workflow. Flag the case for an appropriate editorial decision.
Assess a sample on the finished page. Check how it is read, repetition with neighboring text and the action's name in a link. One correct field does not establish that an entire site is accessible: the whole usage scenario matters, as explained in the business website accessibility guide (in Polish). An operational review of description generation should state its own scope honestly. If the tool produces a broader report, check the coverage of its automated AI accessibility assessment: which page states it visited and which findings it left for manual verification. The article on the scope of automated accessibility checks (in Polish) explains how to examine what an automated assessment actually covers.
The reviewer needs the image and its context. A table of text alone can help find awkward wording, but it cannot confirm the number of elements or the correct variant. Prepare a review view that makes it easy to move from a result to its material. Otherwise the check will be slow and details will start losing to fatigue.
Group results by error type. An invented material, the wrong link destination and unnecessary repetition call for different corrections. One average “alt quality” score can hide a problem that affects only instructions but matters there. Record the number of uses checked, not just the number of files.
The Aeneia Audit offering describes reviews of AI use in companies, with the scope agreed before work begins.
Before another generation, also check protection of approved editorial work. If a person corrected an instructional description, automation should not replace it with a new proposal without a trace. Record the version, decision maker and reason for the update. A new result can be compared before it is published. This preserves knowledge from the previous review instead of reproducing the same error after every catalog update.
After improving the model instruction, return to the same cases and add a few adjacent ones. A better product-photo result does not necessarily mean decoration has improved. Keep image roles and confirmed information boundaries in AI quality regression tests. For later comparisons, the guide to regression tests for AI output (in Polish) develops this approach. Later changes can then be assessed on more than sentence fluency.
Define reasons for another check: a model change, a different page template, a new use of a photo or a product update. The same text can lose its meaning without the file itself changing. If an image starts linking to a different destination, revisit its alternative. The automation needs to recognize that point of change.
Give editors a short set of examples and a way to report exceptions. Not every editor needs to know the image-recognition technique. They should be able to point out that a description says three compartments when the product has two, or that it duplicates the explanation beside it. Such a note should feed into process improvement rather than disappear into a private correction.
A well-accepted automation is not the one that wrote the longest description for every image. It helps prepare the necessary information and preserves the boundary between what is confirmed and what the model inferred. Sometimes the right decision is a short description, sometimes a different way of conveying the content, and sometimes a pause in publication until a question is resolved. The smell of wood can remain in the author's imagination until the catalog provides a basis for it.







