AI’s sense of humor is no laughing matter
When asked to complete this joke, “Why did the chatbot cross the road?” OpenAI’s ChatGPT gave this response: “As an AI language model, it doesn’t have physical presence or the ability to cross roads.”
A rather disappointing punchline, considering the chatbot’s long list of impressive capabilities.
Writers Guild of America members have raised alarms about the use of artificial intelligence in the scriptwriting process, but when it comes to killing a comedy set, these systems have a ways to go.
Marketplace’s Meghan McCarty Carino spoke to Tony Veale, an associate professor at University College Dublin in Ireland who studies computers, creativity and humor, about what it means for AI to develop its own sense of humor.
The following is an edited transcript of their conversation.
Tony Veale: Humor is one of those fundamental human qualities that we think of as being special for us as a species, especially when it comes to the contrast between us and machines. If you understand someone’s sense of humor, you understand them. It’s a bona fide of who they are.
Meghan McCarty Carino: Mastering humor has been held up as a significant milestone for AI. Why would this be such an achievement?
Veale: If you think about the things that traditionally indicated progress in artificial intelligence, such as the ability to plan, to solve problems, to play sophisticated games like Go and chess, all of these have rules associated with them, and mastery of the game involves mastery of the rules. But something like humor exploits the unwritten, tacit rules of social convention. It’s just as much about rule-breaking as rule following. If you think about a machine being humorous, it has to know which rules it can break at any one time. So, it’s not in this well-defined space anymore. Instead, it defines its own space as it goes along. And it’s a risky activity. We take a risk when we try to be funny, when we try to be provocative, when we try to play with norms and conventions. To be humorous, our machines would have to be risk takers. Ultimately humor is a sandbox for playing with ideas.
McCarty Carino: How would you characterize the current state of humor capability in large language models like ChatGPT?
Veale: If you have a low bar, then it’s impressive. These large language models weren’t designed to be humorous; they’re built to be generators of texts that are consistent with a prompt that is given. The suggestion that we might want humor and want the system to be genuinely creative is a very big ask.
If you think about what these systems do, and the facile way they work with language, they do have a certain easy wit about them. It’s not the humor of a stand-up comedian, and it’s not going to be writing jokes for one of the prime-time talk show hosts anytime soon. But on the level of everyday conversation between friends, it’s actually very good I think.
McCarty Carino: I love hearing Jerry Seinfeld talk about the mind of comedians. There’s so much that is ineffable about comedy. You try to use a formula for comedy and say this plus this equals a joke, but it’s not always funny in practice. So, I asked ChatGPT to tell me a joke in the style of Jerry Seinfeld, and it came back with this:
“Why do they call it instant coffee? I mean, if it’s so instant, why does it take longer to make than regular coffee? You stand there pouring water into a cup, stirring, waiting for that magical moment when it finally turns into something that vaguely resembles coffee. And let’s not forget the disappointment when you take that first sip and realize it tastes like someone melted a shoe in hot water. Instant coffee? More like instant regret.”
It kind of sounds like something Jerry Seinfeld would say! What is ChatGPT doing to generate that kind of imitation of a Seinfeld-style joke?
Veale: Well, Jerry Seinfeld is one of those special people that can, in this alchemical way, turn really mundane observations into comedy gold. Language models like ChatGPT have this very general-purpose approach to modeling language, which makes it good at playing language games. Its internal mechanisms are attuned to how words are used together. Speaking like Jerry Seinfeld is a language game, and speaking like Groucho Marx is a language game. If you see enough data from Jerry Seinfeld and Groucho Marx, you can play the Jerry or Groucho language games.
With these language models, words will come spilling out, and they’ll be contextually coherent and meaningful sentences. They’ll have the cadences and the rhythms of these particular comedians, but they, they won’t have the comedic brain behind them that says, “Well, if I choose this word to maximize the payoff of the joke here, then the audience will have this sense of surprise and delayed gratification.” The machine is not that strategic in its use of words to achieve a humorous goal in the way a professional comedian would be. It’s simply spilling the words out, paying no more attention to comedy than to a car manual. They are simply playing the games that language is used to play in these different contexts.
McCarty Carino: Do you anticipate these systems will become more capable of humor in the future?
Veale: Yes, they have these surprising phase transitions when they get bigger, when they have more connections and are trained on more data. They’ll certainly develop what we consider to be the components of a sense of humor. An appreciation for incongruity, for example. The funny, peculiar moments that sometimes turn into funny ha-ha moments. The machines seem to be showing good skills at detecting incongruities at the moment. It’s a skill that I wouldn’t have expected them to have by now, which suggests to me that they’re getting the baseline components in place on which one would build a sense of humor, a sense of incongruity, a sense of the absurd, an ability to turn absurdity into meaning or to extrapolate commonsense situations into the absurd in the first place.
Related links: More insight from Meghan McCarty Carino
If the intersection of AI and “Seinfeld” is your thing, then I have the perfect recommendation for you. It’s called “Nothing, Forever” and it’s an unending stream of AI-generated “Seinfeld” episodes. Or at least, “Seinfeld”-ish episodes that you can watch 24 hours a day, seven days a week on the streaming platform Twitch.
The stream was banned briefly this year after it generated transphobic remarks, but it’s back now. Apparently with more guardrails. It still has the same long, awkward silences, anatomically incorrect sitting postures and bizarrely unfunny jokes.
It reminds me of something Jerry Seinfeld — the real one — told Harvard Business Review a few years ago, when they asked him if a business consultant could have found a more efficient way for him to keep doing the show. In response, Seinfeld said, “If you’re efficient, you’re doing it the wrong way. The right way is the hard way.”
The future of this podcast starts with you.
Every day, the “Marketplace Tech” team demystifies the digital economy with stories that explore more than just Big Tech. We’re committed to covering topics that matter to you and the world around us, diving deep into how technology intersects with climate change, inequity, and disinformation.
As part of a nonprofit newsroom, we’re counting on listeners like you to keep this public service paywall-free and available to all.
Support “Marketplace Tech” in any amount today and become a partner in our mission.
