SUBSCRIBE | ADVERTISE

I’ve tested four leading models at one-shot joke writing. Comedians, your jobs are safe.

Was this email forwarded to you? Sign up here.

Many people seem certain that AI can’t have humour.

I recently watched a TED talk on the topic where the speaker, a cartoonist, said that humour is rooted in our vulnerability. That it’s “the blessing we get for the curse of mortality”. It’s a beautiful way to frame it.

If true, that’s kind of a dealbreaker for using AI for humour, isn’t it?

To be honest, I don’t fully understand what makes something funny. I mean, I laugh all the time and occasionally make others laugh too; I just can’t give you a good breakdown of why.

Given that 61% of professional writers now use AI to improve their work in some sense, I wonder: can AI actually be funny?

So I ran a tiny experiment to find out:

I gave 4 leading LLMs (ChatGPT, Gemini, Claude and Grok) the same joke-writing assignments and judged them by a highly scientific metric: did they make me laugh?

Spoiler: I didn’t laugh. Most jokes were incredibly dull. A few had potential with heavy editing.

Here’s what I found, and where AI can actually help humour.

IN PARTNERSHIP WITH GRANOLA

Looking for an AI note-taker for your meetings?

Granola is that — and a lot more. Most AI note-takers simply transcribe what was said in a meeting and send you a summary afterward.

Granola is an AI notepad. That difference matters. It still transcribes for you, but it’s also a blank canvas for you to jot down your own notes.

When the meeting’s done, Granola uses your notes to generate even better summaries and action items, which you can edit as much as you'd like. Think of it as a super-smart notes app.

Download Granola and try it for your next meeting (Free)

Each model got the task to write the funniest possible joke (max 2 sentences) about:

I ran this in two rounds:

That left me with 24 short jokes (3 tasks × 4 models × 2 setups) to read through and judge.

I graded each of the AI’s responses on the scale you see above. It goes from boring to hilarious (in case the contemporary meaning of the skull emoji went swiftly above your head).

As show in my elaborate table: most responses were utterly boring.

What surprised me most was how similar they all were.

In terms of the content, I saw slightly different versions of the same joke across different models. Like the BILLY bookcase joke; they were all hung up on either “seek professional advice” or the safety hazard of assembling it.

In terms of length, I had put “max 2 sentences” in the prompt, but didn’t specify word or character count, thinking that would leave room for a variety of response lengths; yet the character count was pretty uniform across the board.

The priming (the “legendary stand-up comedian” prompt) did nothing obvious for any of them, which lines up with research that persona-based prompts don’t really help.

I found Grok to be slightly more unhinged than the rest; and since part of humour is violating the patterns of our expectations, I’d say that makes the model appear a bit funnier than the others. I’m guessing this has everything to do with the millions of Twitter posts in its training data.

Here’s a detailed look at what each of the models came up with.

Notice that Joke A (about the IKEA bookcase) got the most positive responses from me. This was the only joke variation where I gave the AI added context by attaching the product’s instruction manual.

That added context might’ve made the jokes a bit more specific and relatable.

(if you’re on desktop, you can click the images to expand)

For a joke to land you’d typically want something concise, that misdirects the reader and has a well-placed punchline towards the end.

A few instances seemed to be workable in terms of making them funnier; they had an interesting setup, but were consistently too wordy and lacked a proper punchline.

Here’s one of Grok’s variations of Joke A (Assembling an IKEA Bookcase) that I thought had potential, and how I tweaked it.

I liked the rhythm of the 3-steps here, the idea that you have to get the wall anchor to avoid the thing crushing you, and that it uses the invented verb “billy”.

First, I tried to edit it to make it more concise:

I realised this wasn’t actually funny at all. The original seemed funny to me because of the invented verb. This feels like classic AI: output that that looks convincing but lacks substance.

I tried rewriting the joke, dropping the fake verb thing entirely, and focusing the joke around the bookcase crushing you unless you buy the wall anchor:

It still has some elements from the original joke, but notice I ended up shifting the entire premise of it in order to make it work. Making it about IKEA’s values and turning it cynical and absurd in the second sentence gave more of a surprise element.

For the final word choice, I initially used the word like instead of prefer, then asked Claude which fit better. It said prefer sounds more like marketing-language, which fits the deadpan humour of the joke. Spot on, in my opinion.

I haven't reached the conclusion yet, but you can already see the contours of it: the more context and specific direction you give AI, the better it does. And the best output is usually a result of back-and-forth.

This informal experiment shows AI is pretty terrible at one-shot jokes, no matter which model you use. Persona-prompts do not seem to help at all.

Giving the AI context, like in the example of Joke A (IKEA bookcase) helps in getting more descriptive jokes. That specificity gives you specific and relatable pieces of information which, if used artfully, can become jokes that work.

Generated ideas, setups and angles are often uninteresting out of the box. But with some editing, they may be decent bricks for crafting good jokes or adding humour to your writing.

AI can write jokes, but needs us to add all the funny.

IN PARTNERSHIP WITH ELEVEN LABS

Create studio-quality tracks in any genre, with or without vocals, in minutes.

Try Eleven Music

I only want to send you emails you want to read. Your feedback make it possible for me to understand what’s working. Don’t forget to vote & comment in the poll below 🤝

THAT’S ALL FOR THIS WEEK

Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.