AI Can’t Tell a Joke.

Here let me prove it.

This post on FKA Twitter is about a high fidelity punchline “George Costanza in a hardcore band.” rendered by AI. The output is not actually funny, although the prompt written by a human is.

Here’s why:

Comedy is math, so it would seem AI would be good at it. But, comedy is inefficient, the most likely outcome is never the funniest, the least likely relevant outcome is. And AI as illustrated by the Paper Clip Maximizer Theory is nothing if not extremely efficient. Therefore it can’t do the hard work of being funny.

Let me break it down by pretending to be a nerd.

The Hardcore Costanza set up is funny because when they appear in a relational matrix they are FAR APART:

DATA POINT 1: [George Costanza] is contextually unlike DATA POINT 2: [hardcore band]

So, when we synthesize them using our brain computers, we identify that the loci are distant and therefore when we construct a relational matrix to connect them, they trigger maximally diverse outcomes, in other words, the setup is so wacky, it sparks our imagination and we laugh.

“Haha” says the human “that would be hilarious.”

Because,George Costanza in a hardcore band is extremely unlikely.

But, as Head Scientist of Comedy Joe Schiappa pointed out to me, the image created by AI is BORING, because it is the MOST LIKELY outcome of George Costanza + Hardcore Band.

It’s exactly what we think it would be, which is not funny it’s lame. A funny outcome would be as unexpected as the input prompt while still remaining relevant to the dataset.

None of these are good, I misspelled “George”. Midjourney prompt: “Goerge Costanza haggling with a man over money in a crowded rock and roll bar.”

For example, Joe pointed out, if George Costanza were in a hardcore band, he wouldn’t be performing. He’d be haggling about the fee for the band or explaining to a very tall woman that he was the lead singer (btw I recognize that as a heteronormative reference from the 1990’s).

“Oh” says the human “I didn’t think of that, but it’s exactly right.” That’s why it’s funny.

Joe’s ideas are funny because they require a human to evaluate ALL the possible outcomes, filter them through the contextual logic of GEORGE COSTANZA’S BEHAVIOR and also the relational data of HARDCORE BAND and then tastefully render the one that will be most UNEXPECTED.

I’ll use two famous jokes to illustrate my point. One is funny, one is not.

This “joke” and everything Jim Gaffigan has ever said is objectively not funny.

“I’ve never eaten a Hot Pocket and then afterwards been, ‘I’m glad I ate that.’ I’m always like, ‘I’m gonna die. I paid for that? Did I eat it or rub it on my face? My back hurts!’”-Jim Gaffigan

Oof.

This is just one masterpiece by comedy legend Mitch Hedberg

“My friend asked me if I wanted a frozen banana, but I said “No… but I want a regular banana later, so yeah.” — Mitch Hedberg

That Gaffigan mumblecore drivel is what I’d call a reference joke. It’s not funny. HOT POCKETS are an unlikely topic of conversation, so he’s pointing at it and saying “Recognize that data point? It’s funny.” In his weird fake voice. It’s incredibly efficient, but he hasn’t done anything human. So either he’s an AI or a bad comic, but either way he sucks.

Mitch Hedberg though, has done something extraordinary. He’s taken three data points which are entirely unfunny, “BANANAS”, “REFRIGERATION” and “TIME” and created a correlation which is both extremely logical, and wildly unexpected. He has looked at a wildly diverse data set and identified the UNLIKELY and RELEVANT thing. The funny thing. And then he went through the trouble of crafting rhythmic language to deliver his outcome for maximum impact. And because it creates so many diverse opportunities it’s hard to pinpoint exactly why it’s funny. Is it because no one ever actually plans their desire for a banana? Is it that a friend asked him if he wanted a frozen banana?

With Gaffigan’s drooling pseudo-rant it’s easy. Hot Pockets. Nothing else is funny. It’s only as funny as Hot Pockets.

So, yes, AI can write Jim Gaffigan jokes, but it can’t write a Mitch Hedberg joke.

When I entered the prompt “write a joke involving bananas, refrigeration and time in the style of Mitch Hedberg” GPT 4 produced:

“I once put a banana in the fridge to see if time would slow down for it. Two days later, it was still just a chilled banana. Time is cold, man.” — Mitch Hedberg style.

Not funny! It’s a reference joke. “Time is cold, man.” is vaguely Hedbergian, but very expected. AI is the worst.

I entered the prompt “Write a joke involving Hot Pockets in the style of Jim Gaffigan” and GPT 4 produced:

“Ever eat a Hot Pocket and think, ‘Did I just bite into the sun?’ You wait five minutes for it to cool, and then it’s ice age in the center. It’s like they have their own climate zones.”

Aside from the fact that it’s significantly funnier than burbling meatbag Jim Gaffigan, it’s just an exercise in descriptive language, there’s no actual joke. Now, for those looking to point at stuff and say “Hey, I recognize that.” the Gaffigan joke is enough to fuel a thought piece or two. But from my perspective as a writer and consumer of comedy, it’s proof that the fundamental functionality of AI doesn’t support humor.

To learn about a proprietary AI tool that forces unexpected connections and creative inspiration hit me up for a DumDum® demo!