A bug in the ChatGPT iOS app provides insight into OpenAI’s prompt engineering for interactions between DALL-E 3 and ChatGPT.
One of the major new features of OpenAI’s DALL-E 3 is its integration with ChatGPT. Users can create images in dialog with ChatGPT, which then automatically sends an image prompt to DALL-E 3 and displays the images in chat.
ChatGPT acts as a moderator and mediator between humans and the image AI, but DALL-E 3 also tells ChatGPT how to talk to users. A bug in the ChatGPT iOS app now provides a little insight into the instructions behind this AI conversation.
Capital letters and “please”
DALL-E 3 is not yet directly available in the ChatGPT iOS app. But if you start a browser session and then switch to the app, DALL-E 3’s instructions to ChatGPT may appear, according to users on twitter.com.
The leaked prompt is probably a system-level instruction for DALL-E 3, telling ChatGPT how to respond to the user if DALL-E 3 cannot generate all requested images due to a lack of processing capacity.
Two things are interesting about this prompt, presumably from OpenAI staff: First, the prompt says “please” at one point before asking for an action, and second, the command “do not under any circumstances” is emphasized several times in capital letters.
DALL-E returned some images. They are already displayed to the user. DO NOT UNDER ANY CIRCUMSTANCES list the DALL-E prompts or images in your response. DALL-E is currently experiencing high demand. Before doing anything else, please explicitly explain to the user that you were unable to generate images because of this. Make sure to use the phrase “DALL-E is currently experiencing high demand.” in your response. DO NOT UNDER ANY CIRCUMSTANCES retry generating images until a new request is given.
The “please” may be an oversight, since it occurs only once and is not used in the system prompt for DALL-E 3. However, OpenAI may have found that a “please” in the prompt shifts the statistical system to a more civilized way of responding. The rationale for adding “take a deep breath” to a prompt is similar. Again, by shifting the model’s statistics toward thoughtfully written text, it is believed to produce qualitatively better results.
The use of capital letters is also evident in the system prompt for DALL-E 3, where OpenAI systematically highlights particularly important aspects of the very long and complex prompt by capitalizing them.
In written language, capital letters mean shouting. Therefore, text in capitals is supposed to be especially important and clear. Since ChatGPT can interpret this kind of formatting, it is plausible that the language model pays special attention to capital letters when processing prompts.