The random images from Unsplash are a godsend, but the problem is that most of the training data probably doesn’t rely on this but on code that relates to the /images folder.
So if OpenAI was able to source its images from its own default personal /images folder with a collection of like, 12 pictures of hamsters or hot air balloons. Then the code it writes will actually be visualised.
Doesn’t need to be elaborate either. Any modest set of stock images that can be provided will do just fine as placeholder.