In 2021, Netflix released what it called “the first standup comedy special written entirely by bots,” featuring an animated “comedian” performing jokes like “I am now so old my birth certificate is death certificate.” The material was purportedly written by bots that had been trained on “400,000 hours” of raw comedy-data by comedian Keaton Patti. The whole thing was a joke—the script was actually written by Patti, putting himself in a bot’s shoes so to speak—that capitalised on the then-nascent fear of the challenge that generative Artificial Intelligence could pose to the world of comedy. Two years later, with the rise of OpenAI and a host of script-writing AI bots, Patti no longer needs to imagine the sort of jokes a bot would write. All he has to do is log onto ChatGPT and give it a prompt. Here’s an example that I just generated while writing this piece:
“You ever ask an Indian engineer a simple question? You’re like, ‘Hey, how do you turn on this computer?’ And they hit you with a dissertation. It’s like, ‘Well, you see, my friend, in the grand scheme of the technological cosmos, the power button is the key that unlocks the gateway to infinite possibilities.’ I just wanted to check my email, man!”
The basic building blocks of a joke are all there, though it’s safe to say that ChatGPT is no Bill Burr, or even a Mitch Hedberg. Not yet. But the rapid advances we’ve seen in A.I. over the last two years, led by deep learning algorithms called Large Language Models, are stoking very real fears that actually funny computer-generated-comedy is no longer a pipe dream. It may even be just around the corner.
Here’s just a few troubling AI comedy data points from this year. In February, Jimmy Kimmel used ChatGPT generated jokes on his late-night talk show, even earning some guffaws from the audience. Comedy clubs in New York have been organising roasts where a human comedian faces off against an AI chatbot (with at least one ChatGPT bot, trained on Sarah Silverman’s corpus of work, actually defeating its homo sapiens opponent). Joe Toplin, an engineering graduate from Harvard who has written jokes for shows like Late Night With David Letterman, is beta-testing WitScript, his own proprietary bot that takes a prompt and spits out three jokes, and then picks the best one. “When it’s working at its best, it’s writing jokes that are good enough to be used on a late night comedy talk show, without any editing,” Toplyn said of WitScript in a recent interview.
It’s a threat that comedians and comedy insiders are taking increasingly seriously. In May, Seth Rogen told The Hollywood Reporter that “any use of AI seems terrifying.” A few months later, humorist and screenwriter Simon Rich wrote in Time magazine that, based on his experiments with more powerful LLMs that haven’t yet been released to the public, ”I think it’s only a matter of time before AI will be able to beat any writer in a blind creative taste test.” AI was also a key concern in the Writer’s Guild of America’s historic 156 day strike this year, with Hollywood screenwriters securing significant guardrails against the use of such technologies, in what many newspapers called the first workplace battle between humans and AI.
But how real and imminent is the actual threat? Are bots really poised to take over comedy, long thought of as such a quintessentially human phenomenon that it made for the ultimate Turing test? For the most part, the researchers and computer scientists most deeply engaged with these technologies think the concerns are overblown, at least in the near term. LLMs use deep-learning algorithms and huge data repositories to generate responses that seem human-like, but what they’re really doing is statistical pattern-recognition and pattern-matching. This means that they can follow standard hack joke formulae, whipping up templatised punchlines in the blink of an eye. But anything beyond that—comedy that is unique and path-breaking—still remains beyond their reach.
The first problem is that A.I. algorithms may be able to generate thousands of variations on “a man walks into a bar” jokes, but they’re still a long way away from actually understanding what makes something funny. A recent study by Cornell University researchers that set LLMs to match 14 years’ of New Yorker cartoons with their respective captions revealed that there’s a significant gap between AI and human-level “understanding” of why a cartoon is funny. The AI was only 62% accurate in a multiple-choice test of matching cartoon to caption, far behind the human average of 94%. And when it came to AI vs human-generated explanations of the jokes, humans won out two-to-one.
The funniest jokes often depend on a host of intangibles—particularly emotional intelligence—that are impenetrable to the machine logic of an algorithm.
The key stumbling block is context. Humour is about much more than assembling words into a punchline. In fact, comedy is more akin to a complex, constantly shifting language of its own. It depends heavily on shared cultural contexts, subtle references, body language and timing, all things that are incredibly difficult for algorithms to parse. The funniest jokes often depend on a host of intangibles—particularly emotional intelligence—that are impenetrable to the machine logic of an algorithm. That means edgy, innovative comedy—with its head-scratching perspective shifts and unpredictable tonal changes—is still beyond its ken. Add in the social element of standup comedy—we go to comedy shows not just for the jokes, but for the social connection, and there’s a lot of stumbling blocks for A.I. to overcome.
So Andrew Schulz or Taylor Tomlinson don’t have to worry about losing their jobs to AI just yet. But the outlook is a little bleaker for comedy writers for TV, film and social media. AI generated comedy may be decades away from winning the Mark Twain prize for American humour, but it does excel at following the sort of formulaic, rules-based topical comedy that you find on late-night talk shows, or social media reels. Even if the quality isn’t up to the mark just yet, the advantage AI does have is speed: it can churn out hundreds of jokes a minute for practically no cost. In the future, especially for studios and content producers that put commercial considerations over actual comedic brilliance (or lack thereof), it might be viable to replace writer’s rooms with A.I generating the jokes, and a couple of writers tweaking and editing them. It’s not going to win you any comedy Emmys, but it may be enough for the next Jimmy Fallon or Kapil Sharma.
What is much more likely, according to experts, is that AI becomes a tool for comedy writers, like a thesaurus or a search engine. They can use it to generate an endless variety of comedy prompts, jumping off points for their own creative work. That’s the route taken by Improbotics, an AI focused theater lab that uses an AI chatbot for its shows, feeding lines from the bots to the performers on stage via an earpiece. A non-generative AI application could be using it to analyse performance footage to see what worked and what didn’t (a sort of training aid, if you will). Maybe the most transformative AI interventions will happen in the non-creative, unsexy aspects of the comedy business: all the logistics coordination and managerial work that goes into running a successful comedy night, venue or TV show.
One example of what that might look like comes from India. For a while, the collective behind Jaipur Comedy Club (JCC) were concerned about how equitable their process for assigning open-mic comedy slots really was. They often got complaints that a particular comedian was getting more slots than the others, or that there wasn’t enough diversity on the lineups, or even that the women were being given late slots which created safety concerns. To address these issues, a couple of the more tech-savvy club stakeholders put together an AI tool that does the job for them, using a number of sorting algorithms to ensure that the process is fair, while also ensuring proper representation for women on each lineup.
“There are also some other criteria that we incorporate to make sure that it’s a good lineup for a good show, but also one that’s fair,” says Tarun Dubey, one of the JCC organisers and a key coder for the tool. “We feel like we’ve been successful on the main criteria we started with, and we’ll look to upgrade it and add more sorting algorithms as we go along.”
Dubey’s tool is easily replicable for other comedy clubs as well, and shows one way that comedians and comedy producers can use AI to make their jobs easier. And he thinks there’s plenty of potential for more. “AI models can help with the proper targeting of audiences or the dynamic pricing of tickets depending on day and time of day,” says Dubey. “This is something that will take a while to come together, but I see lots of scope for AI to make running a comedy club easier and more profitable.”