Aiodic

Back to the table

90 Creative tool Oi OpenAI Images Creative Tools

OpenAI Images

Image generation inside ChatGPT.

What it is

OpenAI Images is the name we use on this table for the picture-making side of OpenAI, the company behind ChatGPT. In everyday use it means one thing: you open ChatGPT, describe the picture you want in ordinary words, and a few seconds later an image appears in the chat. You can then say "make the sky darker" or "put the title in bigger letters" and it changes the picture, the same way you would ask a designer for a revision. OpenAI currently calls this feature ChatGPT Images. Developers can use the same models in their own apps through OpenAI's API, a way for one piece of software to use another, where the models go by names like GPT Image.

The tool has changed names and engines several times, so it helps to know the short history:

What can it do today? It creates pictures from a description, in square, landscape or portrait shapes. It edits photos you upload, for example changing a background, removing an object or restyling the whole image. It keeps track of the conversation, so a character or product can stay consistent across several rounds of changes. And it is strong at text inside images, such as a sign, a menu, a poster headline or a label, which older image generators were notoriously bad at.

Why it matters

It made AI images useful, not just impressive. Early image generators were good at dreamy, surreal scenes but useless for the everyday pictures people need, such as a flyer with a date on it, a diagram with labels, or a social post with a headline. OpenAI's own launch post in March 2025 made exactly this point, arguing that image tools had struggled with the "workhorse" imagery people use to share information. Once the words in a picture came out spelled correctly most of the time, a small business owner could make a menu board or an event poster without design software.

It replaced prompt-writing with conversation. With many image tools you learn a special way of writing prompts, full of style keywords and settings. In ChatGPT you describe what you want the way you would to a person, look at the result, and ask for changes. That lowers the barrier for beginners a great deal. It is a practical example of iterative refinement: the first result is a draft, and the real work happens in the follow-up requests.

It operates at a vast scale. In April 2025, OpenAI said more than 130 million people had made over 700 million images in the first week after the March launch. By September 2026 the company said people were making more than 3 billion images every week across ChatGPT and its developer models. Whatever you think of AI images, tools used on that scale shape what the internet looks like.

It sharpens the hard questions. A tool this capable and this easy to use raises the same questions as every AI image generator, only louder. Is it fair to the artists whose work the models learned from? Who owns the result? How do you tell a real photo from a made one? And what should it refuse to draw? OpenAI's answers to those questions have changed over time, and the examples below show how.

Real-world examples

The Studio Ghibli week (March 2025)

Within hours of the GPT-4o image launch on 25 March 2025, social media filled with photos turned into the soft, hand-painted look of Studio Ghibli, the Japanese animation studio behind Spirited Away and My Neighbor Totoro. People converted family snapshots, wedding photos, pets and famous internet memes. OpenAI's chief executive, Sam Altman, changed his own profile picture to a Ghibli-style portrait.

The demand was enormous. Within days of the launch Altman posted that it was "super fun seeing people love images in ChatGPT", but that "our GPUs are melting", referring to the graphics chips that run the model, and he announced temporary limits on how many images people could make. The launch for free users was delayed for a few days while OpenAI caught up.

The trend also set off a sharp debate. Many fans and artists recirculated a 2016 documentary clip in which Hayao Miyazaki, Studio Ghibli's co-founder, reacted to a demonstration of an early AI animation, calling it "an insult to life itself". That comment was about a different, much older piece of technology, but many people felt it applied. Critics argued that a company was profiting from a style built by human artists over decades. OpenAI's position, in a statement reported by the Associated Press, was that it refused requests in the style of an individual living artist but permitted "broader studio styles", which it said people had used to create "truly delightful and inspired original fan creations". Lawyers quoted in the same report noted that a style in general is not protected by copyright, though specific characters and scenes can be. The episode is worth remembering because it showed, in one week, both why people love these tools and why many artists feel wronged by them.

From blanket refusals to an opt-out: the rules on real people

OpenAI's rules on drawing real people have moved a long way. DALL-E 2 was built to block photorealistic faces of real individuals, including public figures. At DALL-E 3's launch in 2023, OpenAI said it had taken steps to limit images of public figures and pictures in the style of living artists, and it let artists ask for their work to be left out of the training of future models.

With the March 2025 launch, OpenAI changed course. Joanne Jang, who led model behaviour at OpenAI, wrote that the company was "shifting from blanket refusals in sensitive areas to a more precise approach focused on preventing real-world harm". ChatGPT would now draw well-known people, such as politicians, and anyone who didn't want to be depicted could ask to be put on an opt-out list. As reported by TechCrunch, the update also relaxed other rules, for example allowing some edits to a person's appearance that it had refused before. OpenAI said it kept tighter limits on images of children and stronger safeguards on nudity and violence whenever a real person was in the picture.

Supporters saw this as treating adults like adults and allowing satire and commentary. Critics worried it made misleading images of real people easier to produce. The rules that apply at the time of writing are in OpenAI's usage policies (effective 29 October 2025), which forbid using someone's likeness, "including their photorealistic image or voice", without their consent in ways that could confuse people about what is real. Policies like these change often, so check the current version before you make images of anyone real.

Labels you cannot see: C2PA and SynthID

How can anyone tell that a picture came from ChatGPT? OpenAI's answer has come in layers. From early 2024 it began attaching C2PA data to DALL-E 3 images. C2PA is an industry standard for "content credentials", a kind of digital label stored inside the image file that records where it came from. The GPT-4o launch post in 2025 said every image it generated carried this label.

The weakness is that such a label lives in the file's hidden data, and that data is easy to lose. OpenAI's own help pages have said that most social media sites strip it on upload and that taking a screenshot removes it too. So on 19 May 2026 OpenAI announced a second layer: an invisible watermark woven into the pixels themselves using SynthID, a technology developed by Google DeepMind, applied to images made through ChatGPT, Codex and the API. It also previewed a public tool for checking whether an image came from OpenAI. OpenAI was careful to say that no method is foolproof. If the tool finds no label or watermark, it won't declare the image human-made, because the signals can sometimes be removed.

An everyday scenario: a cover concept and a launch graphic

Picture Priya, a self-published author finishing a cosy mystery set in a seaside bookshop. She is not ready to pay a designer for mock-ups before she knows what she wants. In ChatGPT she asks for "a book cover concept, portrait shape, a small bookshop on a rainy harbour street at dusk, warm light in the window, a black cat on the step, space at the top for a title". She gets a charming image, then asks for the cat to face the other way and the rain to be lighter. Next she asks it to add the title and her name in a classic serif font. The words come out almost right, but her surname has one letter wrong, which she only notices on the third look.

She uses the result the way a sensible author would: as a brief to show a cover designer, not as the finished cover. Then she asks ChatGPT to make a square social media graphic announcing her pre-order date, checks the date and the spelling letter by letter, and posts it. When she eventually publishes, she remembers that her retailer asks authors to disclose AI-generated images, including covers.

Try it yourself

Menus and names change often, so treat these steps as a guide to the process. At the time of writing, free ChatGPT accounts can make a small number of images, and paid plans get more, plus extra features such as the "thinking" image mode introduced with Images 2.0. OpenAI does not publish fixed numbers and adjusts limits with demand, so check the message ChatGPT shows you when you reach a limit.

  1. Sign in to ChatGPT at chatgpt.com, or in the ChatGPT app. A free account is enough for this exercise.
  2. Ask for a simple image with words in it. Copy and paste this prompt, then change the details to suit you: Create a square social media graphic announcing a free online workshop called "Start Your First Newsletter". Date: Saturday 14 November, 10am. Clean modern style, soft green and cream colours, a simple illustration of an envelope. Leave plenty of empty space around the text.
  3. Check every word and number. Read the text in the image slowly, letter by letter. Image models still misspell words, invent extra text, or change a date. This is the image-world version of hallucination.
  4. Ask for one change at a time. For example: Keep everything the same, but make the title larger and move the date to the bottom. Small, specific requests usually work better than rewriting the whole description.
  5. Try an edit of your own photo. Upload a photo of an object you own, such as a mug or a book, and ask: Put this on a wooden table next to a window with morning light, product photo style. Compare how well it keeps the object's real details.
  6. Change the shape. Ask for the same image as a wide banner for the top of a website or as a tall image for a phone screen, and notice what it adds or crops to fit.
  7. Try a cover concept. If you write, ask for a cover mock-up of your book in portrait shape with room for a title, as in the scenario above. Treat it as a sketch to discuss, not a finished product.
  8. Check the provenance label. Download one of your images and upload it to OpenAI's verification tool (openai.com/verify) or the Content Credentials site (contentcredentials.org/verify). Then take a screenshot of the same image and check the screenshot. Seeing the difference teaches you more about AI labels than any article.
  9. Read the rules before you publish. Check OpenAI's current terms and usage policies, and the rules of wherever you plan to publish, such as a book retailer or print-on-demand shop.

Common mistakes and misunderstandings

"OpenAI says I own it, so it's protected by copyright." These are two different things. At the time of writing, OpenAI's terms of use (effective 1 January 2026) say that, as between you and OpenAI, you own the output, and OpenAI assigns to you whatever rights it has in it. That is a promise from OpenAI, not a guarantee from the law. In the United States, the Copyright Office's position is that images produced by AI from prompts alone are not the work of a human author, although your own additions, edits and creative arrangement can be protected. In March 2026 the US Supreme Court declined to hear an appeal from Stephen Thaler, a computer scientist who had tried to register an image made entirely by his AI system, leaving in place the rule that copyright needs a human author. In practice this means someone else may be free to copy a purely AI-made image you use. Rules differ between countries, so check your own. The diffusion model page tells the story of the prize-winning AI image that was refused registration.

"My image is unique." OpenAI's terms say plainly that output may not be unique and other users may receive similar output. A cover or logo made from a common prompt could look a lot like someone else's.

Trusting the text. Text rendering is much better than it was, and OpenAI has kept improving it, but it is not perfect. OpenAI's own launch posts list small text, dense information and text in some languages other than English among the known weaknesses. Proofread every word, date, price and web address.

Forgetting platform rules. At the time of writing, Amazon's Kindle Direct Publishing asks authors to disclose AI-generated content when they publish, and says this includes cover and interior images. Other retailers, stock sites and print-on-demand shops have their own rules. Check before you upload.

Using real people without thinking. Being able to draw a public figure doesn't make every image of them acceptable. OpenAI's usage policies forbid using someone's likeness without consent in ways that could mislead, and laws on image rights, defamation and deepfakes apply whatever the tool allows. Never make a realistic image of a private person without their permission.

Believing the label proves anything on its own. A missing content label doesn't mean a picture is real, because labels are often stripped when images are shared. A present label only tells you where the file came from, not whether what it shows is true.

Asking for a living artist's style. In 2025 OpenAI said it refused prompts in the style of an individual living artist. Even where a tool allows it, imitating a working artist's look for commercial use is ethically questionable and may create legal risk. Describe the qualities you want instead, such as "loose watercolour, muted colours, visible brush strokes".

Thinking "DALL-E" is still the current tool. Many older guides and books still talk about DALL-E 3. In ChatGPT it was replaced as the default in March 2025, and the models have been updated several times since. If instructions don't match what you see, the guide is probably out of date.

Expecting it to replace a designer for finished work. It is excellent for concepts, mock-ups, social posts and quick graphics. For a professional book cover, genre expectations, typography, print resolution and the legal position of the artwork all matter, and a human designer still brings a great deal.

Related elements