Humour has long been framed as a boundary between humans and machines. Computers can calculate faster than us and now write essays, compose music, and generate images, but making someone laugh is different. Comedy depends on timing, empathy, and a shared understanding of the world. So what happens when artificial intelligence tries to be funny?

I told my AI to “show, don’t tell” in a story.
It replied with 10,000 tokens of exposition: “The character was sad because statistics.”

The Last Human Frontier?

AI systems don’t laugh, cringe, or enjoy comedy. They learn humour the same way they learn everything else, by analysing enormous amounts of text. Jokes, scripts, and social media posts all become training material. From this data, AI learns which linguistic twists often lead to a punchline.

This is why AI struggles to tell a joke. The structure is visible in the data. Subtle irony or deeply contextual satire, however, remains far more elusive. Ask an AI for a joke and you might get something genuinely amusing. Ask it for ten more and the cracks start to show. The humour can become repetitive, oddly phrased, or unintentionally awkward.

AI Still Doesn’t (really) Get Puns

According to research from Cardiff University and Ca’ Foscari University of Venice, AI still can’t understand a pun. Not generate one, but understand one. The study, put several leading language models through a pun‑comprehension test. A linguistic obstacle course designed to separate genuine understanding from mimicry.

Models could usually identify that a sentence contained wordplay, but when asked to explain why the joke worked, they stumbled. Explanations were often literal, confused, or unintentionally funny. In a creative landscape increasingly shaped by automation, humour remains stubbornly human. Wordplay is messy, cultural and context‑heavy.

Beyond Comedy

The researchers argue that humour is a useful diagnostic tool. Puns rely on cultural context, shared assumptions, and the ability to juggle multiple meanings at once. Humans do this effortlessly. AI, despite its increasing fluency, can’t.

That gap reveals pattern recognition isn’t the same as comprehension. In areas like education, healthcare and customer support misreading tone or intention can have consequences. If a system can’t decode a simple joke, how confidently should we expect it to interpret more complex human signals?

Researchers see this not as failure but as a frontier. Truly understanding humour will demand models that grasp context, culture and intention, not just patterns. AI can imitate wit, but without lived experience or social instinct, it can’t match how humans use humour. In all its messy, context‑soaked complexity, humour remains one of our most human forms of communication.

I don’t trust stairs — they’re always up to something.

Why did the tomato turn red? It saw the salad dressing.

I used to be indecisive… now I’m not so sure.

Why did the PowerPoint presentation cross the road? To get to the other slide.

Why was the piano so hard to open? Because the keys were inside.

lol