Make an image with AI: describe it well, check it, and know what you can use it for

Short answer

Write the prompt (your request to the AI) like a brief: what is in the picture, where, in what style, how it is framed, what shape it is, the exact words it shows, and what to leave out. Then change one thing at a time. Before you use it, check every word, the hands and small details, any logo, every fact in a diagram or map, and whether anyone real appears. Keep the original file, and say it was made with AI. The vendors' terms let you use what you make, but in the US a prompt on its own does not make you the author.

To check an image someone sent you, see Was this image made or changed by AI? A five-minute check. This page is for images you make.

Step 1: Describe it well (5 minutes)

Google suggests starting with "<Create/generate an image of> <subject> <action> <scene> and then build from there", and thinking about "how you want the elements in your image to be arranged (composition)" (Gemini). Its help page adds: "Describe the image style you want" (Gemini Apps Help). OpenAI's prompting guide says to "Put literal text in quotes or ALL CAPS", to spell tricky words "letter-by-letter", and to "State exclusions and invariants explicitly" (OpenAI Cookbook, April 2026).

In ChatGPT, ask for an image in a chat, or select More and then Images. You can set the shape with the aspect ratio (width-to-height shape, such as 16:9) picker or in the prompt (OpenAI Help). Pick it from where the image will go.

Create an image of <SUBJECT> <ACTION>, in <SETTING>.
Use: <USE>. Aspect ratio: <ASPECT_RATIO>.
Style: <STYLE>.
Composition: <COMPOSITION>.
Text in the image, spelled exactly as written, and no other text:
"<EXACT_TEXT>"
Spelling of unusual words, letter by letter: <SPELLED_OUT>.
Leave out: <LEAVE_OUT>.

Replace the following:

• <SUBJECT>, <ACTION>, <SETTING>: such as a bicycle courier, handing over a parcel, a rainy street at dusk (illustrative).
• <USE> and <ASPECT_RATIO>: such as a slide picture and 16:9, or a printed A4 poster and portrait.
• <STYLE>: such as flat vector illustration, two colours or soft studio product photo.
• <COMPOSITION>: where things sit, such as subject on the left third, empty space on the right for a title.
• <EXACT_TEXT>: the words, short. Write none if you will add text yourself later.
• <SPELLED_OUT>: names or uncommon words, such as Q-U-I-N-C-E (illustrative), or none.
• <LEAVE_OUT>: such as real people, real logos and brand names, extra text, borders.

Step 2: Change one thing at a time (2 minutes per edit)

OpenAI suggests starting "with a clean base prompt and refine with small, single-change follow-ups", and to "repeat the preserve list on each iteration to reduce drift" (OpenAI Cookbook). Gemini's page puts it simply: "If you don't like what you see, just ask Gemini to change it" (Gemini).

Change only this: <ONE_CHANGE>.
Keep everything else exactly the same, including: <KEEP_LIST>.

Replace the following:

• <ONE_CHANGE>: such as make the sky clear blue (illustrative).
• <KEEP_LIST>: such as the courier's pose, the parcel, the colours, the framing.

To fix one area in ChatGPT, you can "Select part of the image with the selection tool, then describe the changes in chat". OpenAI warns that "edits may extend beyond the area you selected" (OpenAI Help), so look at the whole image again after each edit. If three edits in a row make it worse, start a fresh chat with a better first prompt (our suggestion).

Step 3: Check the result (3 minutes)

Open the image at full size and go down this table in order (our suggestion). OpenAI's March 2025 launch listed "Precise graphing", "Multilingual text rendering" and "Dense information with small text" among its known limits, and said it can "occasionally crop longer images, like posters, too tightly, especially near the bottom" (OpenAI). The image tools have changed since, so check every image.

A worked example (illustrative)

A teacher needs a title-slide picture for a lesson on recycling.

1. First prompt. Subject a child, action sorting bottles into a green bin, setting a school playground. Use a slide picture, aspect ratio 16:9. Style flat vector illustration, two colours. Composition child on the left third, empty space on the right for a title. Text none. Leave out real people, real logos and brand names, extra text, borders.
2. What came out. A clean picture, but one bottle carries a label that looks like a real drinks brand, and the bin has a made-up word on its side.
3. One-change edit. Change only this: remove all labels and writing from the bottles and the bin. Keep list: the child's pose, the colours, the framing, the empty space on the right.
4. Check. No text left anywhere, the child's hands have five fingers each, nothing is cropped, no logo. She adds the title "Where does it go?" in her slide tool and puts "Illustration made with AI" in the slide notes.

Total time: about 12 minutes, two prompts.

Real people

OpenAI's usage policies say you may never use its services for "use of someone's likeness, including their photorealistic image or voice, without their consent in ways that could confuse authenticity" (OpenAI usage policies, effective 29 October 2025). Google's policy prohibits "Impersonating an individual (living or dead) without explicit disclosure, in order to deceive" (Google Generative AI Prohibited Use Policy, last modified 17 December 2024). Both ban non-consensual intimate imagery.

Say it was made with AI

Images from ChatGPT and Gemini carry hidden provenance signals (marks that record where the image came from), and editing or converting the file can strip them. What they are and how to read them is on Was this image made or changed by AI?. Google's terms forbid "misleading others into thinking that generative AI content was created by a human" (Google Terms of Service), and its use policy forbids "claiming it was created solely by a human, in order to deceive".

What you can use it for

What the vendors say. OpenAI's terms (effective 1 January 2026): "As between you and OpenAI, and to the extent permitted by applicable law, you (a) retain your ownership rights in Input and (b) own the Output." They add that "output may not be unique and other users may receive similar output" (OpenAI Terms of Use). Google's terms (effective 30 July 2026): "Some of our services allow you to generate original content. Google won't claim ownership over that content" (Google Terms of Service). Both vendors' usage policies also forbid infringing other people's intellectual property.

Microsoft Designer is stricter: its FAQ (updated 11 September 2026) says "Designer is licensed for personal, non-commercial use with your Microsoft account", so don't use it for a business flyer or client work (Microsoft Designer FAQ, vendor help).

Models you download carry their own licence. Qwen-Image-2.1 is released under the Qwen Research License, which allows use "for research or evaluation purposes only" and says "You shall not use the Materials for any commercial purpose without obtaining a separate commercial license" (Qwen licence, licence text). Read the licence of the model, not only the app.

These terms settle things between you and the vendor. They do not promise the image is protected by copyright, or that it looks nothing like someone else's work.

What the US Copyright Office says. In its January 2025 report on copyrightability, the Office concluded: "Based on the functioning of current generally available technology, prompts do not alone provide sufficient control." And: "Copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements." Human authors can still own "creative modifications of the outputs" (US Copyright Office, Part 2).

What that means for you (our reading, not legal advice).

• A slide, a lesson, a blog post or an internal poster: fine to use under the vendor terms, with the checks above.
• Something others might also get from a similar prompt: expect that they can, because output "may not be unique".
• A logo, a product design or anything you need to stop others copying: the AI's part alone cannot be protected in the US, so ask a lawyer first.

How to tell it worked

• Every word in the image matches your text, letter by letter, or there is no text in it.
• Hands, details and edges pass a close look at full size.
• No real person, logo or brand is in it without permission.
• Every fact in a diagram or map is checked, or the image makes no factual claim.
• You kept the original file, and the image is labelled as AI-made wherever it could pass as a photo.
• It took a handful of edits. If you are past three that made it worse, restart with a fuller first prompt.

Sources

1. OpenAI. GPT Image Generation Models Prompting Guide. OpenAI Cookbook, April 2026. Vendor guide. Link
2. Google. Gemini image generation overview. Vendor page. Link
3. Google. Generate and edit images with Gemini Apps. Vendor help. Link
4. OpenAI. Images in ChatGPT. Vendor help. Link
5. OpenAI. Introducing 4o image generation, March 2025. Vendor announcement. Link
6. OpenAI. Usage policies, effective 29 October 2025. Vendor policy. Link
7. Google. Generative AI Prohibited Use Policy, last modified 17 December 2024. Vendor policy. Link
8. OpenAI. Terms of Use, effective 1 January 2026. Vendor terms. Link
9. Google. Terms of Service, effective 30 July 2026. Vendor terms. Link
10. US Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability, January 2025. Government report. Link
11. Microsoft. Frequently asked questions about Microsoft Designer, updated 11 September 2026. Vendor help, read 2026-10-19. Link
12. Qwen. Qwen Research License Agreement (Qwen-Image-2.1-Turbo), 20 September 2026. Licence text, read 2026-10-19. Link

Changes

1. v1.010 Oct 2026first version