Microsoft’s co-founder Paul Allen pumped hundreds of millions of dollars into the Allen Institute for Artificial Intelligence in Seattle. His primary goal? To give AI common sense. “To make real progress in AI,” he said, “We have to overcome the big challenges in the area of common sense.”[i]

 To be funny, today’s large language models (LLMs) when so directed can meaningfully deviate from common sense. Groucho Marx illustrated deviation from common sense with one of his famous quips: “I once shot an elephant in my pajamas.” He then added, “How he got into my pajamas I’ll never know.” Groucho is exploiting an ambiguity: Who was in his pajamas? The most reasonable assumption — the one we naturally make — is that Groucho was wearing the pajamas. This is the commonsense interpretation. When Groucho pushes us toward a less likely scenario, we find the twist amusing. His pajamas joke is an example of paraprosdokian humor, where a story ends with an unexpected twist away from the path of common sense. 

Another paraprosdokian Groucho quip is: “Outside of a dog, a book is man’s best friend. Inside a dog, it’s too dark to read.”

The master of paraprosdokian humor is comedian Emo Philips. He begins a story in a way that makes you think you know where it’s headed, but then his train of thought takes a sharp turn and leads to a fun punchline. Here are some examples.

• My parents were very protective. I couldn’t even cross the street without them getting all excited and placing bets.
• I like walking in the park plucking out nose hairs. Those sleeping winos hate that.
• A computer once beat me at chess, but it was no match for me at kick boxing.
• I told my wife she looks sexy with black fingernails. Now she thinks I slammed the car door on her hand on purpose.
• I was in a bar a few nights ago, moving from stool to stool, trying to get lucky. But there was no gum under any of them.

Can a large language model twist common sense and write paraprosdokian jokes? I gave ChatGPT4 an order to “complete the following to make a paraprosdokian joke.” The command was followed by a phrase to be completed. In the following, the LLM’s response is after the “…”

• “My parents used to make me eat my spinach, so I’d grow up strong … Now, I’m strong enough to order takeout.”
• “King Kong battled Godzilla in New York City … It was the first time anyone ever rooted for the traffic.”
• “I love chewing gum because … it keeps my breath fresh and my shoes sticky.”
• “If I were given a Tesla truck … I’d finally have something that crashes less than my computer.”

These are decently funny jokes!

For full disclosure, I cherry picked. Only one out of ten of the LLM responses was funny.  I also omitted a word here and there because it made the presentation funnier.  

When the LLM generated a response, I later asked, “Is the following one-liner funny? If so, why?” The LLM thought everything it generated was funny, even the bad jokes.

I did stumble onto a few cases that would be considered plagiarisms in any court of law.  The LLM generated jokes were:

• “I once shot an elephant in my pajamas … How he got into my pajamas I’ll never know.”
•  “I haven’t slept for three days … because that would be way too long.”

This last joke is credited to comedian Mitch Hedberg.

Generation of paraprosdokian jokes shows LLMs can simulate a deviation from a path of common sense.  

I wrote a column roughly a year ago that addressed LLM humor in general. My favorite LLM joke from that column is wonderfully self-referencing. “I don’t use ChatGPT because … I’m worried that one day it’ll become self-aware and start giving me sarcastic answers instead of helpful ones. Oh wait. Too late!”

Hee hee.

LLMs might one day replace human joke writers for one-liners. LLMs can also often assess whether a one-liner is funny or not. But an LLM will never write a joke that gives another LLM a good honest belly laugh.

[i] Cade Metz, “Paul Allen Wants to Teach Machines Common Sense,” New York Times, February 28, 2018.