This makes sense because slice-of-life introduces novel experiences where there’s not always the threat of a negative outcome, but a novel experience full of surprise, and we delight in the characters’ reactions.
So that old advice of, “When in doubt, introduce a gun”? Introduce *anything*! 😋

Many theories of comedy say the unexpected twist of the punchline creates the laugh. Shock comics get laughs because of how unexpected the shock is, even if it lacks a proper punchline. A better comedian creates laughs through puns, clever incongruities, script overlaps, etc.

The big reason actual comedians dislike these folks is that they don’t really tell “jokes.”
Compare the two:
Comic A: What is up with [slur]? They’re always [stereotype]!
Comic B: Did you know jalapeños are considered berries? So, y’know, be careful if you ever ask me to make you a fruit smoothie.

I wrote that into one of my comedy shows, funnily enough. 😋

There’s one more not included, only implied, in Raskin’s theory: Context, or how it all fits together.
While all of an LLM AI’s joke outputs are painfully unfunny, a performance that makes them the butt of a joke can turn them into something so unfunny that they become funny.
So, Humans: 7, AI: 0. 😏

So that makes Humans: 6, LLM AI: 0.
This study on jokes is also why LLM’s can’t tell stories and why its creative writing output is so awful: creative writing all involves playing with language in ways AI cannot for a very, VERY long time.
Once again, I remind everyone: your creative tasks are safe.

Lastly, there’s Language, the individual words, and if we learned anything from George Carlin to Steven Wright to Patton Oswalt, playing with words is something comedians specialize in.
And despite scraping terabytes of data, quite a bit of which violate copyright laws, LLM’s somehow don’t.

Thanks to an LLM AI’s stubborn architecture, it has very little wiggle room to vary its joke strategies. It cannot do the faux mistake of Punchline A or subtle sarcasm of Punchline B.
It can only do, “I take antidepressants like lifesavers—now my psychiatrist is judging me.”

Next, Narrative Strategy has to do with ordering/meaning.
Compare the two versions of a joke by Miguel Porfirio:
Setup: “Anti-depressents are lifesavers!”
Punchline A: “Wait, I mean I’ve been taking anti-depressants like lifesavers!”
Punchline B: “Because I’m devouring them like candy!”

Next is Target, the butt of the joke, subject of praise, or simply point of weirdness. Only humans can explore this range of possibilities.
If you haven’t noticed, LLM AI can only do, “My ____ is judging me,” “jokes.” It can’t do the nuance of puns or positive incongruencies.

But LLM AI cannot utilize all these props because it cannot think semantically: it just autocompletes the sentence via the shortest possible route, resulting in, “I played ‘What’s New Pussycat?’ 27 times in a row on a diner jukebox—now everyone’s questioning my musical tastes.”
Seeing a pattern?

Next is Situation: a good joke uses all of its props in the same way a good fight scene uses its arena.
Take John Mulaney’s routine about endlessly repeating “What’s New Pussycat?” on a diner jukebox, describing a man pounding a table in rage, “And silverware flied everywhere, and it was FANTASTIC!”

Once again, the limited structure of LLM AI cannot compete with human intuition. While humans can vary their structure, AI can only do a strict setup-punchline order, often separated by dashes.
The result?
“I went to a Japanese restaurant—now I wonder of the cook is questioning my health choices.”
😑