I wonder if you could make an AI that could generate likely meme ideas and then remove people from the equation completely. Seems pretty doable, if not already possible.
The problem is that memes, at least the ones that really take off, work because they are creative and funny, and so far that is probably the area AIs like GPT4 struggle the most with. It makes sense when you think about it, a technology that is built around the idea of predicting likely continuations will struggle with learning what is 'likely' in a context that asks for an element of surprise.
I.e., you'd probably need a pretty deep model if you want it to be able to understand that 'okay usually I combine likely concepts, but _now_ I need to combine unlikely concepts that still make sense together on a lower level'.