these particular ML models are trained on human-generated data, which means the model is a reflection of that data.
This doesn't mean that the model can't create something original though. There's a video on youtube where a linus techtips graphic designer competes with DALLE-2, and I found it interesting that the human graphic designer did exactly what some accused the AI of doing - finding images on the internet and kitbashing them to satisfy the prompt. The AI on the other hand could create a completely original image without any source material, even if its understanding of visual concepts are merely an "interpolation" of the training data.
> which means the model is a reflection of that data
Sorry for the image, but a deconstructed animal - here's a paw, one eye in that glass, muscle fibers sorted on that shelf - is not "what it was" anymore: it's incomparable to the original - it's not really a reflection.
"To go to the other side" is not meaningful, it made sense in the joke and in limited other contexts.
try asking it for a red sphere on top of a green cube next to a yellow cone, it falls apart instantly. extrapolation only goes so far, it's like a photoshop artist who gets the background and then copies elements from the internet to paste into the background