Just remember that this is generated by an algorithm, not written, and that algorithm has no capability of understanding or even explaining why what they've generated is funny.
My daughter was 3 or 4 when I told her a knock-knock joke.
Knock knock.
Who's there?
Banana.
Banana who?
Banana split!
Knock knock
Who's there?
Cow.
Cow who?
Cow SPLIT!!! hahahaha
After having kids, I definitely think copying and random tweaking comes before understanding.
The pinpoint of the boundary between the buildup and the punchline was the point you said "split". This is hilarious to whoever has adequate mental development/laterally orthogonal context for there to be so-called "benign violation", *but*, the elements of societal inclusion and social/conversational flow that were also present and framed the temporal boundary of "this is fun!" were (apparently) developmentally much more accessible around 3-4, and picked up on and cued off of - this was a novel new point of social engagement connection! However the only bits of extrapolable context (in the sense of "n bits of information") available was only the word "split", and since this word did not gel with any currently known contextually-local reference points, it was thus considered "new" and given a sort of independent standing in much the same way we encode brand names and taglines without questioning or parsing them, and also similar to how the radiation symbol was strategically designed so as to be meaningless (dissimilar to any existing and widely known symbols) so "this is dangerous" could be given a globally unique mnemonic. *However* what's really cool is that this sense of "this is its own thing with this additional meaning" was not set in stone in the same way information tends to be later in life, and was likely completely backed out and deleted over the following few days after empirical testing/experimentation found no further correlation (and the whole thing consciously surfaced as boring). Bonus points would be achieved if this event was consciously remembered later on in life and properly recalibrated, but the chances are the info was just subconsciously archived (if that's actually a thing that happens) and disassociated.
TL;DR, associative learning go brrrrt
My potentially-incorrect conclusion is that the classification process you were seeing was filtering information based on an already dizzyingly complex and amazing set of pre-existing intelligent processing and catalyzation, as opposed to an arbitrary fuzzy process that "figured it out".
Source: high-functioning autism + late development that's left me with a stick-shift brain instead of an automatic one :P - I'm still figuring out where my biases and misunderstandings are, hence the repeated caveats above
Your brain "understanding" things is just an illusion.
I would be willing to bet $1000 that you can't articulate exactly what makes a joke "funny".
A deep NN can simulate any function. One that's big enough can generalise input data in any way, including in the same ways the hims mind works.
The universe of possible knowledge of large neutral network is much bigger than the ones of any human brain. The challenge is making it become a superset of the other.