Why Can’t AI Be Funny On Purpose?

A couple of months ago, a friend experimented with ChatGPT, asking it to write a comedy bit. He gave it a healthy prompt with good detail. He even asked it to model the humor after another comedian.

The result was a short, coherent monologue with lines that, as a reader, you’d identify as having the intent to be funny.

But it was certainly not funny. I was much funnier at the 4th-grade talent show telling tread worn Polish jokes. Not to worry, Presentists; it was the early 80s, and I'm Polish. 

I’ve since experimented with GPT-4 to write some jokes to see what it could do, and the jokes produced are all about as funny as licking envelopes. They read like a lot of the writing generated by A.I. Cogent but banal. Understandable but flat.

Example: “Act like a comedian." Write a joke about Mr. Rogers using a microwave.”

ChatGPT: “Why did Mr. Rogers get a new job at the appliance store?"

“Because he was the only one who could make a microwave oven feel special.

He'd pop in the food, set the timer, and say, ‘You know, Microwave... I like you just the way you are... Especially when you're not burning my popcorn!"

They were all like that. Imagine Mark Zuckerberg telling a joke during one of his congressional testimonies, and you get the idea.

Here’s Robin Williams doing a bit about Mr. Rodgers using a microwave from his album – and the only 8-track cassette I have ever owned – Reality… What A Concept::

Now we're gonna do some wonderful experiments you can do around the house. Let's put Mr. Hamster in the microwave, okay?... He knows where he's going. BEEP! Pop goes the weasel! That's severe radiation. Can you say "severe radiation?" Oh, look, you got a little balloon now.

Why isn't AI funny?

A large part of the humor is individual relatability, imagining what is real, imagining what isn't real, and the surprise of bringing all those together. There isn't enough language in the world to create those conditions without the experiences of living, and not only living but living through a BODY. "Funny" is a reaction to a notional state, and laughter is a reaction to that. Quite tricky to reproduce that in others when not experienced by the teller of the tale. AI systems rely on data and algorithms to produce output, and this data reflects the human experiences that have been programmed into it. But AI systems can't draw upon the experience and the accompanying feelings of those experiences and then render them into humorous anecdotes. Additionally, AI systems don't possess the same physical reactions humans do, such as laughter, so creating a sense of relatability and surprise with AI humor would be challenging.

A limitation to AI brought up by some roboticists, futurists, and philosophers is the problem of “embodiment.” Embodiment is the idea that a physical presence or interaction with a physical environment is crucial for developing true understanding and cognition. Intelligence isn’t just a product of a computational model but is also profoundly influenced by the body's interactions with the world. This perspective emphasizes that learning, perception, and cognition are fundamentally grounded in the body's sensory and motor experiences. A brain in a vat can only experience the world textually, giving it words, if it could even receive those. A professor who taught my modern philosophy class at Berkeley, Wallace Matson, liked to say, "To produce a mind, produce a body." Somehow, experiencing "funny" is a physical phenomenon, not just a mental one. Laughter is a physical response not only to the words you hear or the sights you see but to how you FEEL about them.

Much of what we find funny is contextual and spontaneous on the receiver's part. But humor can be conveyed through the written word, too. To do that, the writer must have an audience in mind. That requires knowing what an audience is, who’s in it, and imagining that audience.

One can spend much time tweaking their prompts to achieve a humorous outcome. But when you have worked that hard, you are better off writing your joke, or it isn’t that funny. The joke juice isn’t worth the prompt squeeze.  

What’s funny changes based on who’s speaking, who’s in the audience, the time, the place, the zeitgeist. And what has come before. How does a language model discern a zeitgeist? How does a software program take Lenny Bruce, Robin Williams, Gallagher, Richard Pryor, and Margaret Cho and develop an AI version of, say, Jerrod Carmichael?

There is room for surprise and delight in AI output that can be perceived as humor. So much of being funny is responding to other people’s responses. I’ve performed a lot over the years, and nothing motivates you to improve your act more than real-time audience feedback. If a model can be hooked up to call-and-response data that lets it “discover” what people find funny, perhaps it can hone the ability to create humor.

But for now… when AI can make me laugh like David Sedaris, Mark Twain, Dave Barry, H.L. Mencken, or Art Buchwald, let me know.