Can AI do improv comedy?
Playing pun games with ChatGPT
A while back I started a Twitter account with the concept of posting an improv game every day and having Twitter followers play along. I never had the time to build a following, so it never took off. But I still love the idea of having friends with whom to play pun games—one of my favorite activities! So I thought, why don’t I try playing improv games with OpenAI’s new large language model (LLM), ChatGPT? Maybe it could even help me run my Twitter account!
What the heck is ChatGPT?
ChatGPT has been taking the Internet by storm. I’ve seen it used to write really great Python code, create children’s books, and generate viral Twitter threads, among many other things. But some of the coolest uses have been for comedy writing, from satire King James Bible passages to parody corporate memos. Some of these things are so funny, it’s hard to believe they’re real until you try them yourself. (Unfortunately, OpenAI has blocked the King James Bible prompts for being disrespectful to people’s sacred texts. Sigh.)
But what is ChatGPT? It’s a statistical LLM with a user interface that makes it work like a chatbot. It can remember things you’ve said previously in the conversation, answer follow-up questions, adjust its responses based on your feedback, and admit and correct mistakes. For now, it’s free to the public, though after an introductory research period, you’ll have to pay to subscribe.
What the heck is an improv game?
There are many different kinds of improv games—think of literally any bit from the classic TV show “Whose Line Is It Anyway?”—but my favorite are pun games. For instance, in the game “Sex With Me,” players are given a random noun and tasked to make a jokey or punny analogy to their sex life, with predictably hilarious results.
Playing pun games with ChatGPT
To get ChatGPT to play a pun game with me, I first laid out the rules. I selected a game in which one player is a customer who finds an (unlikely) random object in their soup, and the other player supplies the waiter’s response. Ideally, the respondent should make a joke that riffs on both the object and the scenario. In my instructions, I made sure to be specific about what I wanted, and to provide several sample responses. These LLM algorithms work much, much better if given examples to statistically extrapolate from.
All in all, this was a pretty successful first round! Prompted with “lamp,” the bot made some decent puns linking lamp-concepts with soup-concepts: soup can taste “light” or “bright,” and can “light up your day.” And the format of the responses was right on key. I let the bot know that its first, third, and fifth responses were really good! The other two were definite misses, but these are better than a lot of what I’ve seen from amateur improv players in comedy clubs!
On to round 2 (“there’s a car in my soup”)!
Not quite as good as the first round, but “drive-through soup” and “added a little horsepower” were solid puns. You can see how the statistical language algorithm is doing a kind of word association. It knows that “keys” and “souper” are words associated with cars, but it’s not as good at context or comedic framing. There were similar problems in round 3 (“There’s a frog in my soup!”), though a couple of the responses were pretty good:
I especially liked “the chef is hopping mad” and “the soup gives you a little croak in your throat.” All in all, a solid performance from ChatGPT!
Notably, ChatGPT can play the other side of the game, too, as I learned when I asked it to give me a prompt. Unfortunately, it can’t perform one of the most important functions of a fellow improv player: providing feedback and laughing at your jokes!
Dang it, ChatGPT. I was really proud of that “bestseller” pun! The bot is almost obsessive about reminding you of its limitations as a language model, most of which I’ve omitted from the screenshots above because it’s booooring! It also won’t listen to you if you instruct it to leave these buzz-killing disclaimers out.
ChatGPT may be able to formulate a good joke or two, but it won’t be replacing human players or audiences anytime soon!



