How to describe an image with ChatGPT without turning guesses into facts

Say what the description is for

The same image can need several different descriptions. A picture on a product page may need to explain the product's visible features. A picture inside an assembly guide may need to show how parts fit together. A photo in a personal message may need a friendly overview. A chart in a report may need its values and relationships. Ask ChatGPT for the job you need, not just a general description. Otherwise it may produce a pleasant paragraph that misses the important point.

Start with a sentence such as: I need a short text alternative for this image in a guide about organizing a desk. Add the surrounding heading or caption if it changes what the image means. Also say whether the reader needs every visible object or only the detail that supports the page. A description should help someone understand the information the image provides. It does not need to create a story for every decoration in the frame.

Give the image as input, not a vague reference

The official image-input documentation describes adding an image to a ChatGPT prompt and stating what to inspect. In the web composer, an image may be attached, pasted or dragged in. Use the method available in your current interface and confirm that the image is present before sending. This guide does not claim a particular plan, button label or account test. If image input is unavailable, you can supply a manual description, but the assistant is then working from your words rather than inspecting the image.

Use a readable version. A tiny thumbnail can hide labels; a screenshot of a screenshot can make small text harder to read. When a detail matters, provide a close crop and the original context image, labeled clearly. For example: Image A is the full chart; Image B is a close-up of its legend. Do not crop away the unit, axis or surrounding instruction needed to interpret the detail. If you provide several images, state whether you want separate descriptions or a comparison.

Separate observations from interpretations

An observation names what is visible: a yellow cup sits beside a blue notebook. An interpretation goes beyond that: a student is preparing for an exam. The first may be supported by the image; the second may be a plausible story. Ask for these in different sections. Useful uncertainty sounds like the small label is unreadable or the dark object may be a cable. It should identify the exact detail that needs checking, not wrap every sentence in vague doubt.

Do not infer identity, occupation, health, emotion or ownership from appearance alone. A uniform-like garment does not tell you a person's exact job. A smile does not establish how someone feels. A vehicle in front of a building does not prove who owns it. Even without people, objects invite guesses: a mug may hold tea, coffee or nothing. If your supplied context establishes a fact, say that it comes from the context rather than pretending the image itself proves it.

A conceptual example you can compare carefully

The example here is a described, imaginary illustration, not an uploaded photograph or a live ChatGPT image test. Imagine a rectangular wooden desk viewed from above. On its left is a yellow mug with a dark interior. In the center is a closed blue notebook, with a pencil lying just to its right. A small plant in a white pot is at the upper-right corner. Light from outside the frame casts shadows. There are no people, words, logos or visible clock.

This gives us a controlled reference for a description exercise. A good answer should preserve the objects and their relationships. It should not add a laptop, open pages or a city visible through a window. Nor should it claim the mug contains coffee or that the scene takes place in the morning. You can use this reference to understand how checking works, then apply the same checks to your real image. The reference answer below is an editorial example, not a measured output from an account.

Reference description for the imaginary illustration:
A top-down view of a wooden desk with a yellow mug on the left, a closed blue notebook in the center and a pencil to the notebook's right. A small plant in a white pot sits in the upper-right corner. Shadows fall across the desk. No text is visible.

Unsupported additions to reject:
A student's morning study desk with fresh coffee and exam notes.

Choose detail based on the image's role

If the illustration appears in a desk-organization article, a short description could say: Desk arranged with a mug, closed notebook, pencil and small potted plant. If it demonstrates the placement of a plant away from writing space, keep that relationship: Potted plant at the desk's upper-right corner, away from the notebook and pencil in the center. The second answer includes fewer objects but better serves that task. Length alone does not decide whether a description is useful.

An inventory is different. A reader counting objects needs one row per visible item, perhaps with position and uncertainty. A design review may need colors, spacing and composition. A neutral photo description may start with the overall scene and then the relevant details. Tell ChatGPT which of these you want. If you ask for all of them at once, make their boundaries explicit so a marketing caption does not quietly replace an objective account of what is shown.

Short alt text and a long description do different jobs

Alt text is a text alternative associated with an image. Its wording depends on the image's purpose. W3C's image tutorial distinguishes informative, decorative and functional images. A decorative image may need an empty text alternative; an image used as a link or button needs to communicate that function. Do not ask ChatGPT to describe every image identically. First establish what the image contributes to the page. The same picture can be decorative in one place and important in another.

A short alternative can identify a complex image, but a chart or diagram may require a longer explanation containing its essential information. W3C's complex-image guidance describes this combination of a short identification and a more complete text account. You can ask ChatGPT to draft both, then place the long description where readers can find it. Avoid forcing a table's complete contents into one huge alt-text string. A readable nearby table or explanation is usually easier to use when substantial detail matters.

Describe charts with their labels and units

For a chart, ask for the title, chart type, axes, units, categories and visible pattern before asking for a conclusion. A line that goes up might represent a larger cost rather than a better result. A bar's height is meaningful only against its scale. If the chart gives exact labels, compare those labels with the output. If exact values cannot be read, do not accept precise figures guessed from the picture. Use the source data table when available.

Consider another invented example: a bar chart labeled Orders, with January 40, February 55 and March 50. A suitable description says orders rose from 40 to 55, then fell to 50, leaving March above January. It should not say revenue grew or that a marketing campaign worked. The chart shows counts of orders, not money or causes. If a reader needs the exact figures, supply them in a small text table. Keep the description attached to the chart's own units and period.

Read screenshots without inventing the interface

A screenshot can help explain a visible error, but it captures only one state. Ask ChatGPT to identify the message exactly, relevant controls and the part of the screen you are asking about. An error dialog does not reveal every step that happened before it. Describe your action separately: I clicked Save after changing the title. That gives the assistant a fuller problem without asking it to infer unseen history from the screenshot.

Check private details before sharing a screenshot: names, email addresses, browser tabs, document titles, address-bar parameters and notifications. After cropping or hiding a detail, inspect the actual exported image. When the answer mentions a control, compare its label with the screenshot. If the text is blurry, provide a clearer crop or type the exact message yourself. Do not follow a suggested click path simply because it sounds familiar; the visible image may show another version of the interface.

Repair a wrong object or relationship

Correction works best when it points to the exact failure. For the conceptual desk example, you might say: The notebook is closed, and the pencil is to its right. Remove the invented open pages and laptop. Keep the other verified objects. Ask for the corrected description only, then compare it again with your reference. This is more useful than asking the assistant to be more accurate because it tells it which facts changed and which ones should remain.

If two images are being compared, label them and ask for changes tied to each one. Image A has the plant at the top right; Image B has it at the top left. Avoid saying the plant moved if the images might show different scenes. A comparison can report visible differences without explaining their cause. When orientation matters, specify that left and right mean the viewer's perspective. If the subject's left is required, say so explicitly.

Check the final text in its real setting

Inspect the description for wrong nouns, extra objects, omitted key details, incorrect positions and unsupported numbers. Check quoted text character by character when it matters. Then look at the surrounding page. Does the description repeat the caption unnecessarily? Does it help the reader understand the image's role? If a nearby paragraph already gives a chart's values, you may only need a short identification that directs attention to the fuller explanation.

A good description also has clear limits. It cannot verify that a photo is authentic, prove a time or location that is not evidenced, or establish what happened outside the frame. For safety, medical interpretation or other consequential decisions, an image description is only a description, and the relevant expert or original data may be needed. Keep the final version grounded in what is visible and what you explicitly supplied. That makes the text more useful than a richer-sounding story that the image cannot support.