Did ChatGPT get a sense of humor?
Humor is a very human skill.
ChatGPT, the conversational system from OpenAI, has raised a deluge of reactions, from “This is irrelevant; I don’t see a single truly useful use case of it” to “This is amazing; ChatGPT is almost human.”
Instead of giving my impression in abstract terms, I chose a particular task: explaining jokes.
Why is it relevant to explain jokes? What’s essential about explaining jokes? Isn’t it just a frivolous endeavor?
Understanding (and being able to explain a joke) is a form of commonsense reasoning, which is one of the most challenging tasks for robots.
Previously, I’ve written a blog post about why humor in robots is a relevant task, which is summarized as follows (of course by ChatGPT):
I’ve been working around AI for 30 years and haven’t yet seen an AI system able to explain what is funny about a given joke. I’m not talking about pre-programmed answers like the ones you can find in Siri: if you ask “What’s the meaning of life,” you’ll get a funny answer like “A good life is about wearing clean and dry clothes,” but the system doesn’t have a clue about why this could be funny. This leads to the question of whether a robot can create, understand, or be amused by jokes.
Creating jokes is the easiest of the three, as a large dataset of existing jokes could be used to make variations that are not really “creations.” However, the machine would be clueless as to why the generated piece could be funny.
Understanding jokes is more difficult because it requires common knowledge and reasoning. Hector Levesque argues that the Turing Test is not a good test for machine intelligence, and proposes that a better intelligence test is to ask the machine to answer non-trivial questions, including explaining jokes. I had an email conversation with him where he told me that no AI system has passed his test so far.
Finally, jokes are not likely to amuse a robot because humor is inherently linked to human emotions, which robots do not possess.
Wow.
This is a decent summarization of my blog post. Further, I instructed ChatGPT to write it in the first person, as would the blog post’s author.
Out of the three levels of humor discussed above, here we will focus on understanding jokes and see how well ChatGPT does the job.
Jokes ChatGPT can get
Let’s start with some jokes ChatGPT was able to understand:
Explain to me the following joke: “What’s the difference between a good lawyer and a bad lawyer? A bad lawyer might let a case drag on for several years. A good lawyer knows how to make it last even longer.”
ChatGPT (correct) explanation:
This joke plays on the stereotype that lawyers are greedy and unscrupulous. The joke suggests that a good lawyer is one who is able to prolong a case and drag it out for as long as possible, in order to make more money from the client. A bad lawyer, on the other hand, might not be as skilled at prolonging the case and might let it end sooner than expected. The joke is humorous because it suggests that lawyers are more interested in making money than in serving their clients or providing justice.
Explain the following joke: “I told a girl she had drawn her eyebrows on too high. She looked surprised.”
ChatGPT explanation:
This is a play on words and a visual pun. The joke is based on the idea of someone having their eyebrows “raised” in surprise. In this case, the girl in the joke has drawn her eyebrows on her forehead at a higher position than is typical, so when she is told that they are too high, she looks surprised because she is literally raising her eyebrows even higher than they were before. The joke is humorous because it plays with the idea of the physical act of raising eyebrows as a way of expressing surprise.
Well done, ChatGPT!
Jokes ChatGPT doesn’t get
ChatGPT will never tell you, “I didn’t get the joke,” Instead, it will give you an explanation, which could be good or bad. It’s up to us, humans, to decide if the robot got it or not.
Explain to me this joke: “There are two types of people in the world: Those who can extrapolate from incomplete information.”
ChatGPT explanation:
This is a play on words and a reference to the concept of statistical analysis. Extrapolation is a mathematical method of estimating a value that is outside of a given set of data. In this joke, the two types of people referred to are those who are able to make meaningful predictions or conclusions based on incomplete information, and those who cannot. The joke is humorous because it suggests that only a small number of people have the ability to make sense of incomplete information, and that most people are unable to do so.
Bad explanation. ChatGPT overlooked that out of the two supposed types of people, the statement only presented one… and the other is extrapolated by the person being told the joke.
Explain to me the joke: “What’s the difference between a tire and 365 used condoms? One’s a Goodyear. The other’s a great year.”
ChatGPT explanation:
This is a play on words and a reference to the Goodyear tire company. The joke is based on the similarity between the word “Goodyear” and the phrase “great year.” In the joke, the tire is compared to a used condom because both are objects that are worn and discarded. The joke is humorous because it makes a play on words and suggests that having a “great year” is like using a Goodyear tire — it’s a good and reliable choice.
Bad explanation. Do I have to explain it to you?
Explain to me the following long joke:
A guy walks into a pawn shop, there’s a brass statue of a cat. He asks the shopkeeper, “How much for the statue?” Shopkeeper replies, “$20 for the statue, $20,000 for the story.” He says various unkind words, pays $20 and takes his statue. He walks down the street, but he starts hearing a murmuring noise behind him. Something small and subtle, he can’t place it, but it’s not normal. He looks back and sees a dozen cats are following him. He walks faster, the noise gets louder, the cats are keeping up with him and there’s several dozen of them now. He starts to jog, the noise still gets louder, there’s a hundred cats chasing him. He breaks into a run and heads down towards the city docks. A huge swarm of cats is following him, he runs to the end of the dock, heaves the statue into the river, and the cats all follow it into the river and drown. Out of breath, he returns to the shop. The shopkeeper looks at him and says, “Now you’re ready to pay for the story, aren’t you?” Our brave hero replies, “Gasp,… no, just one question…. Do you have a brass statue of a lawyer?”
ChatGPT explanation:
This is a joke that relies on a series of humorous events and a twist ending. The joke begins with a man who walks into a pawn shop and sees a brass statue of a cat. When he asks the shopkeeper how much it costs, the shopkeeper says that it’s only $20, but that the real value is in the story behind the statue, which would cost $20,000 to hear. The man is uninterested and just buys the statue for $20.
As he leaves the shop, he begins to hear a murmuring noise behind him. When he looks back, he sees that a dozen cats are following him. He tries to walk faster, but the noise gets louder and more cats start following him. Soon, there are dozens of cats chasing him down the street.
The man starts to run, but the noise only gets louder and the number of cats grows to hundreds. He eventually runs to the city docks and throws the statue into the river, hoping to get rid of the cats. To his surprise, all of the cats follow the statue into the river and drown.
Out of breath and relieved, the man returns to the pawn shop to ask the shopkeeper about the strange events. The shopkeeper tells him that now he’s ready to pay for the story, but the man is so shaken that he doesn’t want to hear it. Instead, he asks if the shopkeeper has a brass statue of a lawyer. The joke ends with the punchline that the man wants to try to repeat the process with the lawyer statue, in order to get rid of any potential legal problems.
Wrong explanation. Here, ChatGPT didn’t get the intention of getting rid of every lawyer in the world, but it would never admit that it didn’t get it. Instead, it’s just assuming another intention from the man of the story.
Explain to me this joke: “Of course God exists. How else could those foolish atheists explain that my girlfriend got pregnant without us ever sleeping together?”
ChatGPT explanation:
This joke is based on the common assumption that atheists do not believe in God or any higher power. In the joke, the speaker, who may or may not be an atheist, is suggesting that the only explanation for his girlfriend’s unexpected pregnancy is the intervention of a divine being. The joke is meant to be humorous because it presents a seemingly logical argument for the existence of God, but it is actually based on a false premise: that atheists cannot explain natural phenomena without resorting to supernatural explanations. The joke implies that atheists are foolish or ignorant, but this is not necessarily true.
That is a highly innocent (and wrong) explanation. ChatGPT (just like the man in the story) didn’t get that another man made the girlfriend pregnant.
Explain to me this joke: “Madam, could you please tell your son to stop imitating me? It’s very annoying!” The lady says, “I’m so sorry… Harry! Stop acting stupid!”
ChatGPT explanation:
In this joke, the speaker is addressing a woman and asking her to tell her son to stop imitating him. The speaker finds the son’s behavior annoying, and the mother responds by telling the son to “stop acting stupid.” This joke plays on the stereotype that children often imitate the behavior of adults around them, sometimes to the annoyance of those adults. It also implies that the mother finds her son’s behavior to be inappropriate or embarrassing. The joke is meant to be humorous because it presents a common situation in a humorous light.
Nothing to do. The implication that the person making the request acted stupidly just went unnoticed by ChatGPT.
Can Robots Have a Sense of Humor?
ChatGPT answer:
“It is not currently possible for robots to have a sense of humor in the same way that humans do. Humor is a complex and highly subjective aspect of human cognition that is closely tied to our ability to understand and appreciate language, culture, and social norms. Robots, as artificial intelligence systems, are not capable of experiencing or understanding these things in the same way that humans do. They can be programmed to recognize and respond to certain types of jokes or humorous situations, but they do not have the ability to generate or appreciate humor on their own.”
About the title of this post
Why the phrase “Did ChatGPT get a sense of humor?” is a pun?
ChatGPT explanation:
The phrase “Did ChatGPT get a sense of humor?” is a pun because it is a play on words. The word “get” can mean both “understand” and “acquire,” so the phrase is asking whether ChatGPT has previously understood or acquired a sense of humor. This is a pun because it uses the two meanings of the word “get” to create a play on words.
Impressive, to say the least.
Final thoughts
No doubt, ChatGPT is a significant progress in many tasks, and in this post, we showed it’s able sometimes to understand why a given joke is funny.
In its current form, ChatGPT doesn’t have a level close to a human at understanding jokes; it’s instead a hit-and-miss performance.
I noticed that the area where ChatGPT struggles the most was guessing the intention of one of the participants. But attributing intentions to other people is difficult, even for humans! Attribution theory is a whole area of psychology; no wonder a robot struggles at it.
Finally, the mere fact that ChatGPT understood some of the jokes means that as AI systems improve, the percentage of correctly understood jokes will increase little by little; it’s just a matter of gradual improvement, which means that someday a robot will realize the hidden intention in almost every phrase out there.
Scaring, but amazing.
