The Twitch ‘Seinfeld’ Show Proves AI Shouldn’t Write Comedy
After viewers flagged the segment for the transphobic remarks, Twitch banned Nothing, Forever for two weeks. Hartle, embarrassed and apologetic, said he’d be more careful about filtering topics for offensiveness. But he made no promises that Larry would get funnier material.
Nanyun Peng, an assistant professor of computer science at UCLA, studies AI’s ability to be creative and cowrote the papers “A Unified Framework for Pun Generation with Humor Principles” and “Pun Generation with Surprise.” She says AI’s unfunniness stems from the fact that it uses a probabilistic model to determine the most expected idea, and humor is based on unexpected responses.
That seems like a poor excuse. I’ve worked in sitcom writers’ rooms and a lot of joke writing is just math. In theory, one should be able to use machine learning to do this. In fact, Peng did once teach an AI model the rules of humor as comedy writers have tried to explain it. Theories such as the rule of three and the incongruity theory. “Our machine was able to generate ‘The greyhound stopped to get a hare-cut,’” she told me. I felt bad for not laughing, which seemed to deflate her. “It’s not that funny. But it is somewhat. When we saw this result, we were excited.”
The challenge, Peng says, is huge. “I don’t think humans really understand jokes. There aren’t theories where you can use them and then you become a standup comic. Some of it really is the talent,” she says.
Comedian Whitney Cummings, who had a robot made in her likeness for her 2019 Netflix special, wasn’t surprised that an AI told awful jokes. “Why are people shocked that robots aren’t funny? Most humans aren’t funny. The only funny robots are Roombas when they get stuck under the couch,” she says.
Cummings is pro-robot in general; she even keeps the robot version of herself in her house. But she doesn’t expect it to make her laugh. “Comedy is one of the few things that’s so specific to the essence of a human being,” she says. “Comedy is about the trauma that comes from a human’s lived experience and how they cope with it. Robots can’t be traumatized.”
When I asked Spike Feresten, who wrote for Seinfeld from 1996 to 1998 why he thought Nothing, Forever wasn’t hilarious, he suggested asking AI why it isn’t funny. But when I logged on to ChatGPT, it said it was unavailable because it was at capacity. Oddly, though, the left side of the page provided an explanation of its failure by way of an AI chat prompted by “Write a standup comedy routine about the status of ChatGPT.” The closest thing to a joke in it was:
When I sent that to Feresten, he replied, “It’s like asking why Spock isn’t funny.”
In fact, when I was able to log on to ChatGPT later and asked why it wasn’t funny, the bot basically said the same thing Feresten did, only with less humor: “While AI can recognize patterns and generate responses based on them, it doesn’t have a sense of humor in the same way that humans do. It doesn’t experience emotions, understand context, or grasp the nuances of language in the same way that humans do.”
Still, it was able to summarize this article in a few seconds: I should have thrown in more jokes.