ChatGPT and Gemini can crack jokes, but they don’t quite get your puns

We’ve all seen AI write a limerick or a dad joke that’s… passable. But a new academic study just burst that bubble. It turns out that while large language models (LLMs) like ChatGPT and Google’s Gemini are great at mimicking the structure of a joke, they generally have no idea why it’s funny – especially when it comes to puns.

The paper, cleverly titled “Pun Unintended: LLMs and the Illusion of Humor Understanding,” took a hard look at how these bots handle wordplay. The verdict? They are faking it. The study found that while AI can easily spit out a punchline it has seen before, it completely falls apart when trying to understand the subtle, double-layered meanings that make a pun actually land.

AI isn’t as humorous as it pretends to be

Here is the thing about puns: they rely on “polysemy” (words having multiple meanings) or sound-alikes to create a playful clash in your brain. Humans do this effortlessly. AI, however, is just pattern-matching.

The researchers proved this by essentially trolling the bots. They built two new test sets called PunnyPattern and PunBreak. They took real puns and slightly tweaked them – swapping out key words so the double meaning was gone but the sentence structure stayed the same.

A human would instantly know the joke was ruined. The AI? It often kept insisting the sentence was funny just because it looked like a joke it had seen during training. It’s like someone laughing at a punchline they didn’t hear just to fit in. This proves that despite their confidence, these models aren’t “getting” the humor; they are just reproducing the shape of it.

But your AI assistant isn’t a comedian yet

If you are a writer, marketer, or just someone trying to spice up a presentation with ChatGPT, you might want to double-check that copy.

This research serves as a solid warning: AI-generated wit is often hollow. Because the model doesn’t understand the intention behind the wordplay, it might give you a “pun” that makes zero sense, or miss the sarcasm and irony entirely. It’s the difference between a comedian who knows how to work a room and a parrot repeating a knock-knock joke. If you rely on it too heavily for creative writing, you risk publishing content that feels robotic or, worse, confusingly unfunny.

Can AI ever truly understand puns?

The researchers argue that feeding AI more data isn’t going to fix this. To truly “get” a pun, a system needs to understand how words sound (phonetics) and grasp the cultural context that makes a twist funny. Right now, text-based models don’t have those ears or that life experience.

Future AI might need a total architecture overhaul – hybrid systems that can combine standard language skills with phonetic reasoning – before they can actually hang with human comedians. Until then, the ability to craft a groan-worthy pun remains a uniquely human superpower.