I’m obsessed with the question of whether machines can ever be funny.
I’ve already shared how ChatGPT just “doesn’t get it” about what a joke is. It seems to understand the structure of a joke, but never quite gets what makes jokes funny. (The last of these ChatGPT jokes is so stupid that is a little bit funny, like a 5-year’s old attempt to make a pun.)
Write a joke about a therapist on a desert island
Why did the therapist end up on a desert island?
Because they needed a little “me” time and a lot of “couch” time.
Write a joke based on the double meaning of the word “foil” — that to foil is to stop something from happening, and foil is a thin sheet of metal
Why did the chef use a foil?
To keep the fish fresh and to keep the plans of the burglars
Write a funny story about a priest, a rabbi, and a minister
Once upon a time, a priest, a rabbi, and a minister were stranded on a deserted island. They had been sailing on a boat trip together when they were caught in a storm and shipwrecked.
As they were trying to make the best of their situation, they decided to build a shelter and gather food. But they soon realized that they had vastly different ideas about how to go about it.
The priest wanted to pray for rescue and the minister wanted to have a group discussion about how they were going to survive. The rabbi, on the other hand, was all about taking action and suggested they use his tallit (prayer shawl) as a sail to catch the wind and try to reach the nearest island.
The priest and the minister were skeptical, but the rabbi convinced them to give it a try. So, they set sail on the makeshift boat, using the tallit as a sail. It worked surprisingly well, and they soon reached a nearby island where they found food and fresh water.
The trio then decided to build a raft and make their way back home. Once they arrived home, they were hailed as heroes and invited to give a talk on survival skills at the local church, temple, and mosque.
From that day on, they all had a newfound respect for each other’s beliefs and methods. And they all agreed that the rabbi had been right all along: “In times of crisis, it’s best to keep a tallit handy.”
Write a knock-knock joke about a bunny
Bunny one gonna open the door or should I just hop on through?
Write a pun on the word “celery,” which sounds like “salary”
Why did the celery turn down the job offer?
Because the salary wasn’t enough to stalk up on
Perhaps you would draw the conclusion that machines just can’t be funny. But as a counterexample, consider Horrible Therapist, a randomly generated comic from The Oatmeal. Matthew Inman has created a tool that assembles a comic from a random first, second, and third panel. A bunch of the results are dumb, but some are funny in a norm-violating oh-my-gosh kind of way. (Warning, if you try it, you’ll see that a lot of them are disgusting.)
A lot of those are not funny, either, but Matthew Inman knows something about how a joke is constructed that ChatGPT does not. His first panel asks a question, the second is shocking, and the third is an odd reaction. When the odd reaction is connected to the shocking revelation in an unusual way, it sounds a bit like a joke.
I think we could get there
On Star Trek, The Next Generation, the android Data tries (and ultimately fails) to understand humor. Star Trek’s writers apparently think a funny machine is impossible.
I’m not so sure.
When I worked with the cartoonist Mark Anderson on cartoons for a book, I was amazed at the speed with which he could generate ideas related to the chapters. He had a stock set of comics panels that included desert islands, therapists, people in the break room at an office, and, for some reason, dogs. We’d riff on ideas until something struck us as funny, then he’d turn it into a cartoon.
If you watch sitcoms (other than Seinfeld, which breaks the mold), the jokes are formulaic and it sure seems like a machine could generate hundreds of them. Maybe only 5% would be funny, but if you trained the machine on which ones were funny and why, could it learn to write sitcom jokes? (One comedy writer tried to make a joke generator.)
There is something about humor that may be beyond machines. It requires a twist in an unexpected direction, and machines may be able to understand the twist but not what an “unexpected direction” is. Matthew Inman’s randomly-generated cartoons go in unexpected directions, but often are so bizarrely incongruous that they’re just weird, not funny. ChatGPT’s jokes don’t seem to be able to make the connection to the unexpected at all.
Maybe, like Data, they’ll never get there. But we used to think natural language was impossibly complex for a machine, and now with ChatGPT and half-decent machine translations, we don’t.
Do you think humor will be forever beyond machines?
Call it the boring Turing test. Can a machine make jokes that are anything other than boring?
You can pretend to be smart. But you can’t pretend to be funny.
You’re right that A.I. hasn’t yet shown much capacity at all to write funny jokes or, more to the point, know when a joke is funny, but it seems to me that a lot of human beings trying to be funny aren’t very funny either.
The problem you pose may be harder than you’ve posited, too. I’d bet most jokes aren’t funny to a significant portion of the entire population even when a subset will find the joke funny, so creating a joke that will be funny to almost everyone may be a nearly impossible challenge. I’m guessing that a joke creator has to first determine the nature of the particular intended audience before creating a joke that is likely to resonate with that audience.