The archive · Product Ideas · Product decision · 2023
OpenAI DALL·E 3 (2023): ChatGPT writes the prompts, you just describe the picture
DALL·E 3 kills prompt engineering: built natively into ChatGPT, it turns a plain-language wish into a tailored prompt, then refines the image by conversation.
OpenAI
What it had to solve
Text-to-image systems ignored wording and forced users to learn prompt engineering. OpenAI wanted a model that obeyed natural, detailed descriptions — and an interface where anyone could use it without learning that skill.
How it works
On September 20, 2023, OpenAI announced DALL·E 3, a text-to-image model built natively on ChatGPT. Instead of feeding a separate tool a hand-crafted prompt, users describe what they want to ChatGPT, which writes the detailed prompt, generates images, and refines them through ordinary conversation.
The problem it solved was prompt engineering: earlier systems tended to ignore words, so getting a good image meant learning to write precise, long prompts. OpenAI fixed the model side by training DALL·E 3 on captions generated by a state-of-the-art image captioner, producing a model that 'heeds much more attention to the user-supplied captions' and reliably renders intricate details, including text, hands and faces.
The interaction side made the model the prompt writer: ChatGPT brainstorms with the user, expands a simple sentence into a tailored prompt, and takes tweaks like 'make it warmer' without any syntax. The Verge's demo showed lead researcher Aditya Ramesh asking ChatGPT to design a logo for a mountain ramen restaurant and getting four options in return.
OpenAI paired the release with safety work — refusals for public figures and living artists, red-teaming, an early provenance classifier over 99% accurate on unmodified images, and an opt-out for creators' images — then shipped the feature to all ChatGPT Plus and Enterprise users on October 18, 2023.
Why it lands
- It removed the expert barrier: because the model writes its own prompts, a beginner's one-line wish produces output that previously required prompt-craft.
- Conversation became the interface: iterative refinement happens in natural language, not by rewriting prompt syntax between generations.
- The captioner-trained model was a technical enabler — it made the text-to-image engine obey detailed instructions, so the chat wrapper had something reliable to work with.
- Embedding the tool in ChatGPT piggybacked on an app hundreds of millions already used, taking image generation from a specialist tool to a mainstream feature.
What it did
From October 18, 2023, every ChatGPT Plus and Enterprise user could create and refine images conversationally, and the chat-native pattern became the default way mainstream image AI is used, carried forward by OpenAI's later GPT-4o image generation.
What you can take
When expert input is the barrier, put the translation inside the product: have the model itself turn plain intent into expert-level instructions, and the tool opens to everyone.
Since then
DALL·E 3 spread quickly from ChatGPT Plus and Enterprise to the API, where it became one of the most-used image models, and Microsoft added it to Bing Image Creator. The chat-native interaction it introduced — brainstorm, generate, refine by talking — became the template for mainstream image generation: OpenAI's later GPT-4o image mode kept images inside the chat, and competing products copied the conversational loop. Within OpenAI it also shifted the company's default from standalone generators toward multimodal capabilities living inside ChatGPT.
Sources
- DALL·E 3
- DALL·E 3 is now available in ChatGPT Plus and Enterprise
- OpenAI releases third version of DALL-E
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card