Why being funny is AI's toughest test
"Hey everyone! Great to be here. You know, life is a lot like a toaster. Sometimes you're popping out of bed ready to take on the day, and other times you feel like you're stuck on the darkest setting, just waiting for something to change."
That's the opener ChatGPT offered when I asked it to come up with "an original comedy routine:" An observation so far from joke territory, with a landing so baffling and clunky, not even a Chapelle or a Seinfeld could raise a laugh with it. Imagine performing this routine to a comedy club audience and you can imagine the withering response. Stone-cold silence broken only by coughing, the shuffling feet, and easy pickings for a heckler: "Yeah, buddy, I think you just toasted your act."
AI isn't funny. That much is clear from anyone who's ever asked a system like OpenAI's ChatGPT or Google's Gemini (formerly Bard) to tell a joke. It's also desperately in need of new material; according to one 2023 study in which researchers asked for 1,008 "original" jokes from ChatGPT, more than 90% of replies were the same 25 jokes ... none of which were original. (The most popular go-to gag in the study: why did the scarecrow win an award? Because it was outstanding in its field.)
That's not to say AI can't be unintentionally funny — kind of like the oblivious "straight man" in sketch comedy. Witness this viral Twitter thread where a human asks an AI art program to draw "a very normal image" then keeps asking it to double down on the normality, with hilarious results. But ask AI to be funny on purpose, and what you get is often as confusing as the toaster joke. A particularly baffling example from the study: why did the man put his money in the blender? Because he wanted to make time fly. "ChatGPT has not solved conversational humor yet," the researchers concluded with hilarious understatement.
More importantly, many experts believe that AI – at least in the forms conceived today – may never solve its funny problem. "Most people don't realize how hard humor is, because there's a lot of planning ahead that goes into a good joke," explains Mark Riedl, a professor at Georgia Tech's Machine Learning Center who studies AI's storytelling abilities. Large Language models like ChatGPT are simply looking for predictive patterns in language; they don't even know they're supposed to build up expectations and create tension that will be released with an unexpected punchline.
Seasoned comedians are playing chess; AI is playing checkers.
"It's a little bit like an alien watching a standup routine and saying 'oh, I can do that, I think I understand the pattern,'" Riedl says. "But it doesn't realize there's a lot more going on in the mind of the comedian and the minds of the audience. We would need AI systems that are fundamentally constructed differently to really do a good job of this."
Good news for comedians everywhere: unlike other knowledge workers, you may never have to worry about AI coming for your job.
If at first AI fails …
Is there any hope for our unfunny AI comedian? Well, the greatest standup artists have to hone their craft endlessly in front of small crowds, even when they're at the top of their game, and they can't be afraid of being unfunny. That's one takeaway from Comedy Book, a critically acclaimed nonfiction tome released at the end of last year. It's an exhaustive look at the superstars of standup, and the rules one can infer about comedy itself, by Vulture comedy critic Jesse David Fox.
For example, Fox recounts one time he saw Chris Rock show up for a surprise late-night set at a New York comedy club in 2003, at the height of his fame. He bombed. At the time, "I chalked it up to a bad set," Fox wrote. "When I saw him do it again 14 years later, I knew better." Rock deliberately bombs, Fox discovered — stripping the context from jokes, tripping over his punchlines, trying new material on the fly — all to figure out what the actual essence of funny is. If it works in that room, it'll work anywhere.
Okay, so maybe AI just needs to pay its dues. It needs to bomb a few times, or a few thousand times, to truly hone its craft. If OpenAI or Google would let you provide feedback to ChatGPT or Gemini on their jokes – stars out of five? – they'd learn what we like to laugh at faster.
Speaking of honing, Riedl points out that we're not being entirely fair to ChatGPT a lot of the time; "tell me a joke" is so nonspecific that it's like saying to a comedian "hey, be funny." So maybe we just need to ask better questions to get better comedy. Surely these Large Language Models have devoured enough Marvelous Mrs. Maisel to know what a "tight five" is. Alright ChatGPT, give me your best tight five!
"Hey, everyone! How's it going? Good to see you all. You know, I've been thinking a lot lately about adulting, and let me tell you, it's overrated. I mean, who decided that paying bills and worrying about your credit score was the peak of human existence? If that's the case, I want a refund on adulthood, please."
Um, okay. Well, that's still not LOL-worthy, and even though our imaginary comedy club heckler has a ready response – "Yeah? I want a refund on this show!" – he might be too depressed by that little existential cocktail to even heckle. And there might be something funny in that – in the incongruity of an AI pretending to complain about "adulting," in the starkly naive hope of a refund, in the mundane anti-humor of it all. Funny in the laughing at rather than laughing with sense, but funny nonetheless.
A few observations from Fox's book seem pertinent here. First of all, as he repeats throughout, humor is 100% personal: "if someone somewhere is amused by a joke, it is by definition funny." Secondly, comedy trends change over time, especially as younger audiences (Gen Z, the so-called "comedy natives" who grew up on Netflix standup specials) become ever more sophisticated.
Not all comedy is laugh-out loud funny; there's the comedy of callbacks and catchphrases, the comedy of pointing out how artificial comedy itself is. We've recently gone through a fourth wall-breaking trend in comedy specials; an AI that refuses to play along when asked to deliver a tight five, and drops a truth bomb instead, seems like it has potential to be the Bo Burnham of AI.
But for now, at least, AI is hampered by another of Fox's rules, something so surprisingly important to comedy he wrote a whole chapter on it: trust. We laugh differently, more freely, when we trust that a comedian knows where they're going. We need to believe that whatever incongruous (or even highly offensive) thing they appear to be saying in the present moment, all will be resolved with a release of tension in the end.
In the end, we simply don't trust AI enough for its output to be funny. Quite the opposite: we're on our guard around everything it utters. "I'm not convinced that we want AI systems that are better at tricking us into thinking they're human," Georgia Tech's Riedl says. In the comedy club of AI, we're all hecklers.
But that's okay, because ChatGPT's got responses for hecklers! "Well, if I'm not human," it said, apparently warming to a theme of fake-human-worried-about-finances, "does that mean I can finally get a refund on all those therapy sessions?"
Reader, despite myself, I chuckled.