ChatGPT humour: a dissection

I love comedy; I have written comedy professionally; it is important to me that the comedy I consume is human-generated, so I try to keep on top of AI’s capabilities in that department.

A recent essay in the New York Times, ‘Why Does A.I. Write Like … That?’ analysed several hallmarks of AI-generated writing. It covered all the main giveaways like the em dash (—), ‘It’s not X, it’s Y,’ and a fixation on the rule of threes (whoopsie!).

For quite a while, I have been experimenting with ChatGPT’s ability to produce humorous writing, which has improved. It’s still definitely not consistently good, and a lot of it is nonsense, but it can sometimes, as far as it appears to me at least, produce jokes and lines that could conceivably have come from a human. As in, if I didn’t know they came from ChatGPT, I could believe they came from one of the big American publications that publish humour. And it’s not just me that thinks ChatGPT is capable of producing competent humour—a 2024 study found that ‘ChatGPT 3.5-produced jokes were rated as equally funny or funnier than human-produced jokes regardless of the comedic task and the expertise of the human comedy writer.’

Much like with general AI writing, I find there are specific quirks that appear again and again in ChatGPT’s output, both in humour essays, and in general writing when employing a humorous tone or as quips or snarky comments. Things that, when they appear in supposedly human writing, make me go 🤨.

---

I find there to be an even heavier reliance on the rule of three when ChatGPT is trying to be funny. (I hate to anthropomorphise ChatGPT, but you know what I mean.) There is a particular tendency to end on a ‘comically’ absurd third thing, and often the final item is considerably longer than the other two.

Sorry to insert a large chunk of ChatGPT text, but this is an illustrative example (it’s not always as ridiculous as this). If you are unfamiliar with the rule of three, by the time you get to the end of the following passage, I promise you won’t be. I asked ChatGPT to write from the perspective of a Christmas caroller who really wants some figgy pudding:

Yes, I’m the caroller. Yes, I’m on your doorstep. Yes, my hat has a bell on it that jingles every time I breathe in a way that suggests unresolved childhood trauma. But that’s not the point. The point is: this is the year the figgy pudding clause becomes legally binding.

For too long, you’ve treated that line like it’s just merry nonsense. A bit of whimsy. A cute outdated request from the days when people thought oranges were an acceptable Christmas present and soot was a food group.

If we, as a society, cannot honour the simple promise contained within one of our most aggressively cheerful Christmas songs, then what are we even doing? What is the point of civilisation? Why are there figs?

For too long, the line “We won’t go until we get some” has been treated as a sort of whimsical, Victorian-era bluff. A cute lyric. A festive throwaway. Something to be sung cheerfully while we politely shuffle off your porch like Dickensian Mormon missionaries.

Not this year.
This year, I have range and I have resolve.

When we reach verse three—“Oh bring us some figgy pudding”—I will not smile sweetly and tilt my head like an apologetic Elizabethan pageboy. I will not soften the consonants. I will not giggle. I will not modulate my volume for fear of upsetting your elderly terrier who already looks like it’s lived through eight wars.

No.
I will deliver that line with the full spiritual force of a person who has Googled what figgy pudding is and now wants—needs—it more than sleep.

And when we hit “We won’t go until we get some”, you will see something shift behind my eyes. A steeliness. A knowledge. A contract being invoked.

Because here’s the truth: you have been getting away with this for years.

Every Christmas, we show up. We sing. We bring tidings. We give goodwill. And you smile and say things like, “Oh, you’re wonderful!” or “Would you like some hot chocolate?” or “Please stop blocking the Ring doorbell, sir.”

So this year, consider this a formal notice.

If you do not supply figgy pudding, we will not move.
We will hold the line.
We will root ourselves into your driveway like ancient druidic oaks.

Your neighbours will come out. They will ask questions. They will wonder why we are performing the bridge of “We Wish You a Merry Christmas” with the emotional intensity of a hostage standoff.

And we will say, calmly and with dignity:
“We told them the terms.”

Listen, we don’t want trouble. We just want pudding. I’m not even asking for much. A ramekin-sized portion. A pudding adjacent to pudding. A pudding that looks like it comes with a footnote explaining the concept of “suet.”

ChatGPT is trained on existing writing, and the rule of three is a frequently used writing principle, particularly in comedy where the joke is often on the third thing in the triple. People will often incorporate this in their writing without even realising it’s an established technique because we see and hear it so often, and because it is so effective. So it’s easy to see why ChatGPT churns it out. But good god, that was a lot in succession.

---

ChatGPT will often default to motifs of contemporary culture, like vibes, tote bags, anxiety, teenage angst, the idea of being on brand.

That is the holy trinity. Shoes. Keys. Vibes.

Last night someone offered me gluten-free shortbread and said, “It’s basically the same vibe.”

I don’t even want to come inside, because I saw your dog and it does not respect boundaries.

like a cursed internal Spotify playlist.

like a child’s finger painting mixed with a sense of existential dread.

There’s a lot of [normal thing] mixed with existential dread, anxiety, angst, an unmistakable sense of loneliness, etc.

There’s often a general tone of contemporary cynicism. A kind of snarky jadedness along with approximations of the kinds of things people reference nowadays. Like someone from 50 years in the future doing a parody of what people are like today. Jokes about emotions or lack thereof, phones, modern insecurities like people not texting back.

My phone informed me it was “feels like 3°C,” which feels less like information and more like a threat. I put on a jumper. I was already wearing a jumper. I put on another jumper, just in case the first jumper was emotionally thin.

I picked up my phone to see if anyone had messaged me in the last twelve seconds. No one had, which felt rude but on brand. I replied to nothing, just in case.

That man could turn a vineyard into a midlife crisis.

I simply... compartmentalised. Like luggage. Emotional carry-ons.

Look, when we divorced, I was emotional. Hormonal. Postpartum, possibly premenopausal—somewhere on the female emotional timeline between “crying at an advert for tea” and “buying a castle.” And Nick was—how do I put this delicately—a man who wears denim shirts unironically.

Accidentally liking your ex’s 8-year-old post while stalking their new relationship.

---

Often, things almost make sense, but not quite.

Like it has set itself something to write, and the rest of the output it generates doesn't quite fulfil the expectation it has set. A clear example was that it generated the title ‘Things I Have Mistaken for My Phone Buzzing’ and, of the 28 items on the list, 7 of them started with ‘My phone buzzing with…’* But it’s frequently less overt than that; it will give a set-up, then the payoff doesn’t completely satisfy that set-up.

*(The list also included ‘The nation’s mood,’ ‘A sense of dread,’ ‘Capitalism,’ ‘The concept of urgency, whispering directly into my femur,’ ‘A phantom limb, except the limb is validation,’ and ‘A memory of being needed.’ See previous.)

Or things that have the cadence of contemporary humour, but are just nonsense:

The weather did not require checking, because I had already experienced it by being alive earlier today, but my phone needed to confirm this. It agreed it was cold, but in a way that felt personal. I put the keys down to find my jumper, which I was already wearing, but not enough.

Have you ever struggled to locate a jumper you were wearing because you weren’t wearing it enough????

Or things that do literally make sense—a series of words that has a discernible meaning and doesn’t directly contradict itself—but don’t make sense as a joke to someone who has experienced life as a real person:

Every year, at least one person receives something baffling and must perform gratitude like it’s a West End audition. “Oh! A… candle. That smells like… memory.”

A candle??? As a present??? Baffling!!!

---

A ‘tell’ I see out in the wild (read: on Substack) fairly frequently is that ChatGPT will replicate common tropes, but in a context that doesn’t work and thus the joke doesn’t work.

I noticed a mug I had abandoned earlier, now cold and judging me. I rinsed it. I did not put it in the dishwasher. That felt like a commitment.

Fear of commitment being a longstanding trope, but in this context it just isn’t applicable.

A lukewarm paper cup of mulled wine that tastes like heated regret.

Something tasting or smelling of regret (or other emotions like desperation) is a recognisable expression. But without relevant context, a cup of mulled wine tasting like regret, heated or otherwise, is just…nothing-y.

There’s also a tendency towards exaggeration in a way that you can tell is part of the humour generation, but it’s just not actually funny, e.g. rules being ‘enforced with the volume of a rock concert.’

---

There’s a paragraph in the NYT article I linked to above that says:

When you spend enough time around A.I.-generated text, you start to develop a novel form of paranoia. At this point, I have a pretty advanced case. Every clunky metaphor sets me off; every waffling blog post has the dead cadence of the machine. This year, I read an article in which a writer complained about A.I. tools cheapening the craft. But I could barely pay attention, because I kept encountering sentences that felt as if they’d been written by A.I. It’s becoming an increasingly wretched life. You can experience it too.

Oh my god, do I! When I read something that is phrased like a joke, but the execution doesn’t quite make sense. When someone makes a snarky quip mixed with modern malaise and existential dread. When someone uses a few too many lists of three th—never mind!!!

A post that recently did the rounds on here, which the poster later admitted was at least partly created by AI, described one classic novel as a ‘stream of consciousness modernist masterpiece that reads like someone’s fever dream.’ Another was ‘great if you have unlimited time and emotional bandwidth.’ Another ‘makes you feel like you’re reading your own anxiety disorder.’

‘Reads like someone’s fever dream,’ ‘emotional bandwidth,’ ‘like you’re reading your own anxiety disorder’—that’s that snarky humour with contemporary tropes that ChatGPT produces so much of.

Another example comes from a humorous essay with a lighthearted, silly tone. I completely thought I was reading human writing until one line read exactly like ChatGPT humour, and once I’d noticed that, suddenly tells were popping up all over the place. The line that made me raise an initial eyebrow was a quip about a woman who would be performing onstage, ‘probably wearing tights.’

Think about that image. Is that a funny or even noteworthy image? Would it occur to you to highlight that a woman onstage was wearing tights? Or is that almost literally every woman on every stage and so commonplace that it’s actually more unusual to find it worthy of comment?

The idea of being onstage in tights is often mocked in general culture. You know what the joke is. It’s a man playing Peter Pan or similar, and the ‘funny’ image is a grown man in skin tight green bottoms. So you can see how LLM output could make a joke of someone in a performance wearing tights full stop. That’s what generally jars me about unexpected AI humour. When something reads as a joke and you know you’re supposed to identify it as an amusing comment, but when you actually take the words as written, it just doesn’t really work. It’s just not something a human who lives in the world would say.

---

The reason I find the above to be easy AI tells is because they don’t work for me, humour-wise. I’m never put off reading something just because it includes a lot of em dashes, because em dashes don’t give me the AI ick. Even though they are technically one of the biggest giveaways, I believe real people still use em dashes because I am a real person and I still use em dashes. I like them so they don’t feel out of place unless they are accompanied by other obvious AI quirks.

There are probably other tells that I haven’t picked up on because they don’t strike me as non-human, or because I find them natural and smooth and contextually appropriate and therefore skim right over them. I definitely don’t believe I can always tell when writing has been generated by AI. There are almost certainly things I have read and enjoyed and found to be clever and funny that involved some amount of AI generation, either in the writing or in the editing or in the drafting. I actually think if you think you can always tell, you are doing yourself a disservice, but I’ve written about that before and I’m actively restraining myself from going on another tirade about it so that’s here if you want it.

This was long, I’m done now, bye!

If you’re feeling particularly generous, you could…