GPT-4 can only do text-to-text and image-to-text. It can't generate images itself. So it will simply use an API call. Really nothing special, Bing does the same thing.