In “The Outrageous Okona,” the fourth episode of the second season of Star Trek: The Next Generation, the Enterprise’s resident android Data attempts to learn the one skill it has previously been unable to master: Humor. Visiting the ship’s Holodeck, Data takes lessons from a holographic comedian to try and understand the business of making funny.
While the worlds of Star Trek and the real world can be far apart at times, this plotline rings true for machine intelligence here on Earth. Put simply, getting an A.I. to understand humor and then to generate its own jokes turns out to be extraordinarily tough.
How tough? Forget Go, Jeopardy!, chess, and any number of other impressive demos: According to some experts, building an artificial intelligence on the level of a top comedian may be the true measure of machine intelligence.
And, while we’re not there yet, it’s safe to say that we may be getting a whole lot closer.
Joe Toplyn is someone who doesn’t shy away from challenges. Toplyn, an engineer by training (with a large career gap in terms of actually practicing it), carved out a successful career for himself as a TV writer. A four-time Emmy winner, he’s been a head writer for the likes of David Letterman and Jay Leno. Several years ago, Toplyn became interested in the question of whether or not there is an algorithm (i.e., a process or set of rules that can be followed) that would help write genuinely funny jokes.
“People think it’s magic,” he told Digital Trends. “Some comedy writers or comedians, I think, try to portray what they do as performing magic. Well, it is like magic in the sense that a magic trick is constructed and designed, and there’s a way that it works that fools you into thinking that the magician has supernatural powers. But there’s really a logic to it.”
This belief in a steely logic to joke-telling — honed while Toplyn was trying to teach his “magic” to aspiring, would-be comedians — ultimately led him to try building an A.I. able to generate off-the-cuff quips that fit into regular conversations. Called Witscript, the results add up to an innovative A.I. system that creates improvised jokes. A chatbot that uses Witscript to ad-lib jokes could, Toplyn said, help create likable artificial companions to help solve the “huge problem” of human loneliness. Think of it like PARO the robot seal with punch lines.
“It’s context-relevant,” Toplyn said of Witscript, which was recently presented at the 12th International Conference on Computational Creativity (ICCC 2021). “This sets it apart from other joke-generating systems that generate self-contained jokes that aren’t easy to integrate into a conversation. When you’re talking with a witty friend, chances are that their jokes will be integrated into a conversation in response to something you’ve said. It’s much less likely that your friend will just start telling a stand-alone joke like, ‘A man walks into a bar with a duck on his head …’”
This spontaneous quality comes from the joke-writing algorithms Toplyn himself developed.
“Basically, the way the basic joke-writing algorithm works is this: It starts by selecting a topic for the joke, which could be a sentence that somebody says to you or the topic of a news story,” he said. “The next step is to select what I call two ‘topic handles,’ the words or phrases in the topic that are the most responsible for capturing the audience’s attention. The third step is to generate associations of the two topic handles. Associations are what the audience is likely to think of when they think about a particular subject. The fourth step is to create a punch line, which links an association of one of the two topic handles to an association of the other in a surprising way. The last step is to generate an angle between the topic and the punch line: A sentence or phrase that connects the topic to the punch line in a natural-sounding way.”
If all these handles and angles sound like hard work, the proof is — ultimately — in the pudding. Using 13 input topics, Witscript generated a series of jokes, which Toplyn then pitted against his own efforts. For a review board, he outsourced the judging to Amazon Mechanical Turk workers, who graded each freshly minted joke on a scale of one (not a joke) to four (a very good joke). One of Witscript’s best efforts garnered a 2.87 rating (“That’s pretty close to being a joke,” Toplyn said) to his own 2.80 as student beat master. The Witscript joke? Riffing on a line about the 25th anniversary of the Blue Man Group performance art company, it quipped: “Welcome to the Bluebilee.”
While perhaps not quite yet ready to displace Dave Chappelle, Toplyn believes that Witscript proves that humor can, to a degree, be automated. Even if there’s still a long way to go. “As machines get better at executing those algorithms, the jokes they generate will get better,” he said.
However, he also struck a note of caution. “To generate [truly] sophisticated jokes the way an expert human comedy writer can, machines will need the common-sense knowledge and common-sense reasoning ability of a typical human.”
This, as it turns out, may be the crux of the matter. Humor might seem frivolous, but for those who work in the fields of language, comedy, and artificial intelligence, it’s anything but.
“We use humor in a lot of different ways,” Kim Binsted, a professor in the Information and Computer Sciences Department at the University of Hawaii, told Digital Trends. “We use it to establish social rapport. We use it to define in-groups and out-groups. We use it to introduce ideas that we might not be willing to express seriously. Obviously, there’s nonlinguistic humor, but [linguistic humor] falls into a category of language use that is really powerful. It isn’t just a stand-up on stage who uses it to get a few laughs. It’s something that we use all the time [within our society.]”
“It is an enormous signifier of advanced intelligence because, in order to be truly funny, an A.I. needs to understand a whole lot about the world.”
When it comes to computational humor, Binsted is a pioneer. In the 1990s, she created one of (possibly the) first A.I. designed to generate jokes. Developed with Professor Graeme Ritchie, Binsted’s JAPE (Joke Analysis and Production Engine) was a joke-generating bot that could create question-and-answer puns. An example might be: “Q) What do you call a strange market?” “A) A bizarre bazaar.”
“It was great because it meant I could pick all the low-hanging fruit before anyone else,” she said modestly. “Which is pretty much what I did with puns.”
Since then, Binsted has developed various other computational humor bots — including one able to dream up variations on “Yo mama” jokes. While Binsted’s work has since evolved to look at long-duration human space exploration, she still views joke-telling A.I. as a sort of holy grail for machine intelligence.
“It’s not one of these things like chess, where when A.I. was starting out, people said, ‘Well, if a computer can ever really play chess, then we will know it’s fully intelligent,’” she opined. “Obviously, that’s not the case. But I do think humor is one of those things where fluent humor using a computer is going to have to be genuinely intelligent in other ways as well.”
This is why joke-telling is such an interesting challenge for machines. It’s not because making an A.I. crack wise is as useful to humanity as, say, using machine intelligence to solve cancer. But it is an enormous signifier of advanced intelligence because, in order to be truly funny, an A.I. needs to understand a whole lot about the world.
“Humor depends on many different human skills, such as world knowledge, linguistic abilities, reasoning, [and more],” Thomas Winters, a computer science Ph.D. student researching artificial intelligence and computational humor, told Digital Trends. “Even if a machine has access to that kind of information and skills, it still has to have insight into the difficulty of the joke itself. In order for something to be funny, a joke also has to be not too easy and not too hard for a human to understand. A machine generating jokes should not use too obscure knowledge, nor too obvious knowledge with predictable punch lines. This is why computational humor is usually seen as an A.I.-complete problem. [It means] we need to have A.I that has functionally similar components as a human brain to solve computational humor, due to its dependency on all these skills of the human brain.”
Think of it like a Turing Test with a laugh track. Coming soon to a superintelligence near you. Hopefully.
- I used an app to create 3D models with my iPhone, and it’s shockingly great
- Image-generating AI Dall-E is now free for anyone to play with
- Why AI will never rule the world
- Nvidia’s $200 Jetson Orin Nano minicomputer is 80 times faster than the previous version
- I pitched my ridiculous startup idea to a robot VC