A photo shows two people working beside a raised planting bed. Your AI draft calls them “local volunteers celebrating a successful harvest.” That sounds plausible, but the image does not establish where they live, whether they are volunteers, what they feel, or whether anything has been harvested. A fluent caption has quietly invented the story.
A useful photo caption joins what readers can see with context you have verified. This worked example complements our guide to writing alt text for different contexts. Here the task is the visible caption beneath a photograph: a small piece of editorial copy that must carry its own factual weight.
Start with three separate sets of information
Before asking for a caption, separate visible observations, verified background, and unknowns. This takes only a few lines. It also gives you a clear way to reject attractive wording that has no support. A model's confidence is not an additional source.
For this fictional editing exercise, the photograph shows two adults placing seedlings in a raised bed. A watering can stands nearby. The event organiser's approved notes identify the people, from left to right, as Mira Shah and Alex Chen. The notes say they are participants in a planting workshop at Eastbank Community Garden on 12 September 2026. The photographer is Rowan Lee, and the publication has permission to use the image with that credit.
The notes do not say that either participant lives nearby, is a volunteer, has gardened before, or grew the seedlings. They give no harvest results or attendance total. Those gaps should remain gaps. Do not ask AI to infer identities, personal circumstances, or an event's success from appearances.
Identify the invented story in the first draft
Suppose the draft reads: “Local volunteers Mira Shah and Alex Chen proudly celebrate their first successful harvest at Eastbank Community Garden, where a popular programme is transforming the neighbourhood.”
The names and place come from the record, but almost everything connecting them is unsupported. “Local volunteers” assigns residence and status. “Proudly celebrate” supplies emotion and purpose. “First successful harvest” changes planting into harvesting and adds both a history and an outcome. “Popular” and “transforming” claim impact without evidence.
Do not merely replace “proudly” with “happily” or shorten the sentence. Remove each unsupported claim. Then check that the remaining action matches the photograph. A caption can be concise and still wrong; word count is not the main test.
Build a caption from the approved facts
A defensible version is: “Mira Shah, left, and Alex Chen plant seedlings during a workshop at Eastbank Community Garden on 12 September 2026. Photo: Rowan Lee.”
Each part has a job. The names identify the subjects, “left” resolves which person is which, the verb describes the visible action, and the location and date supply verified context. The credit follows the fictional permission record. It does not claim that mentioning a photographer is sufficient permission to reuse an image.
If the photo will appear directly beneath a heading that already names the workshop, you may be able to shorten the caption. Check the actual page before doing so. A photo reused in a newsletter or social preview may lose that surrounding context, so it may need a fuller caption there.
Use a prompt that makes missing information visible
Try: “Draft one photo caption of no more than 45 words using only the observations and approved notes below. Keep names, left-to-right order, location, date, and photo credit exact. Do not infer emotions, occupations, residence, motives, popularity, or outcomes. Put any questions for the editor in a separate list, outside the caption.”
Then provide the three sets of information. If the name order has not been confirmed, say that explicitly. An editor question such as “Can you confirm which person is on the left?” is useful. A guessed order hidden in polished copy is not. Resolve the question with the source owner or omit identification until it is verified.
Use only material you are authorised to share with the tool. If an unpublished photo or its notes contain sensitive information, work with a fictional example or approved placeholders. The point of the prompt is to constrain the rewrite, not to upload every detail available.
Give captions and alt text their own review
The W3C guidance on informative images explains that text alternatives should convey the image's meaning in context and can account for information already supplied by nearby text. That is why copying a visible caption into an alt attribute is not an automatic accessibility solution.
Review the actual image, caption, and surrounding paragraph together. Decide what visual information remains necessary for someone who cannot see the image. Our linked alt-text example walks through that separate decision. Keep this caption review focused on whether the published identification and context are accurate.
Check the final crop before publishing
A correct caption can become misleading after an image change. If an editor crops out Mira, “Mira Shah, left” no longer works. If a different workshop photo is substituted, the action or date may no longer match. Read the caption against the exact image that will go live, including its mobile crop if it changes what is visible.
Finally, compare every proper name, date, action, and credit with the approved record. You can use AIUndetectable to explore a more natural phrasing of an approved caption, then repeat that factual comparison after editing. Natural wording does not establish truth, permission, or provenance. The finished caption should help a reader understand this photograph without making up the part you cannot verify.