If I asked Stable Diffusion to produce a photo of Stability AI's server racks, I would expect it to produce a picture of a server rack, but I would not expect it to produce an actual photo of the actual server rack (assuming Stability even has one), even though it can produce a very realistic Steve Jobs.

These models are rewarded for producing realistic text and images. As almost a happy accident, this means they can often reproduce their training data. But if something isn't in their training data, they're expected to make something up. It's what they're designed to do.