the internet

Are AI-Generated Memes Actually Funny Or Are Our Brains Broken?

With the power of a neural net at their fingertips, the internet’s unreformed chuckleheads are forcing a computer to generate images like ‘Super Mario sentenced to life’ and ‘babies doing parkour’.
Screen Shot 2022-07-06 at 2

Of the long list of professions that could be automated out of existence in the cyberpunk future, you would think “comedian” ranks pretty low. There’s a certain degree of human understanding to telling a genuine thigh-slapper – surely beyond the capability of silicon transistors and artificial intelligence, no matter how sophisticated they might be.

But recent developments have had some people asking otherwise. 


The last few months has seen an explosion in the spread and use of “language models”, a type of AI that’s broadly intended to enable computers to understand and interpret natural human language.

Trained on enormous quantities of text and image data that they scour for patterns and connections, these systems can parse and output human-like language with unnerving accuracy. OpenAI’s GPT-3, first launched in 2020 but only opened for wider use earlier this year, is able to respond to human input with what at least appears to be an uncanny degree of understanding. 

Another prominent option is DALL-E 2, also created by OpenAI and currently in a very limited beta, which does the same thing with images. Type “a velociraptor in Speedos sun-baking on the sands of Majorca” into DALL-E 2 and it’ll do its best to draw on its pattern recognition ability and enormous corpus of data to make that for you. Of course, the system isn’t actually intelligent. It doesn’t know what a velociraptor looks like, or where Majorca is. But, by tapping into the internet’s deep subconscious through billions and billions of parameters, it has a vast library of the visual and language patterns usually associated with those words.

Simply imagining Super Mario being present at Dealey Plaza during the JFK assassination might not have been inherently funny twenty years ago, but it’ll reliably get some bottom of the barrel laughs today.


DALL-E 2 isn’t yet open to the public, with OpenAI insisting its trying to iron out technical problems and possible ethical issues before unleashing it onto a wider audience. But the transformer language model built to replicate it named Craiyon – originally launched with the copyright-disrespecting name DALL-E mini – is available to anyone. It’s given us a preview as to what the human culture of 2022 will do when given access to a tool like this. The answer is probably obvious: they made highly idiotic memes.

Social media accounts like Weird Dall-E Mini Generations – which has clocked nearly a million Twitter followers since its launch – have chronicled the nascent comedic explorations of the internet hivemind when given access to technology that conjures visual imagery from a simple English text prompt. 

With the power of a neural net at their fingertips, the internet’s unreformed chuckleheads forced the computer to generate images from prompts like ‘Super Mario sentenced to life’ and ‘babies doing parkour’.

GPT-3, which you can sign up for and play around with right now on its website, was put through similar paces.

The model’s great pattern-matching ability means it can be scarily good at capturing the tone and cadence of popular figures – especially if there’s plenty of transcripts of that person speaking online. Naturally, that has led to a cottage industry of fictional conversations and transcripts, like this thread of Trump roasting various figures from Greek mythology in his trademark way:


Or its unsettling ability to generate posts in the popular 4chan greentext story format:

But let’s not fool ourselves. When you ask GPT-3 to straight up write a funny joke, it’s not very good at it. Terrible, actually. What it produces tends to be either highly derivative of jokes that have already been written by humans, or make absolutely no sense whatsoever. (When I asked it to write a funny political joke, it responded: “Why don't more politicians take up surfing? Because they're afraid of getting wiped out!”) 

But the particular flavour of pattern-matching weirdness that drives these new systems happens to be very well-suited to the pop culture blender of meme culture, in which simply stitching together a few disparate things can be – and very often is – funny. Simply imagining Super Mario being present at Daley Plaza during the JFK assassination might not have been inherently funny twenty years ago, but it’ll reliably get some bottom of the barrel laughs today.

In a newsletter, writer Max Read argues that the experience of cycling through responses to comedic prompts to GPT-3 is kind of like the experience of using Twitter – “pressing a button to see what strange, funny, outlandish thing might be said next”. There’s no sophisticated punchline or concept to “the Thai cave boys hitting the club to celebrate their freedom”, and yet I just generated it with Craiyon, and the blurry output did, I’m mildly embarrassed to admit, make me laugh.


Does this simply reveal that much of online comedy is utterly mechanical now – the artificial collision of two disparate cultural concepts together, for a trained audience of people who will likely see more funnies in one afternoon scrolling their phone than a medieval peasant would experience in their entire lifetime? 

It may not be a truly intelligent machine, but it is humbling to know that something that is ultimately just very good at identifying and replicating patterns can get within a stone’s throw of replicating the contours of the meme-industrial complex of 2022.

One of the most weirdly dislocating things about this emergent culture of AI-generated comedy is that it’s so difficult to separate the content from its mode of production. It’s challenging to tell, for example, whether a distorted, impressionist image of Karl Marx being slimed at the Nickelodeon Kid’s Choice awards is actually funny in and of itself, or whether it’s just funny because a computer made it. (Which is, you have to admit, very funny.)

“You might have noticed that nearly all presentations of art produced with these models include the text prompt,” writes author Robin Sloan in a perceptive blog on AI art

“The pleasure, it seems, is not in the image; rather, it’s in the spectacle of the computer’s interpretation.” 


He’s talking about AI art, specifically, but you can see the same dynamic play out with the avalanche of generated memes that colonised social media over the past month or two. Without the text prompts identifying them as a bit of AI comedy, they probably wouldn’t be funny — just another of the trillions of images that pass us by every time we look at a screen.

It’s worth remembering, with that in mind, that the AI isn’t coming up with any of these concepts. The human users are. They’re just using the language model to bring the idea to life visually, or riffed on by the software. In this way, it’s more like a comedic tool that can be leveraged by a human comedian, rather than a possible existential threat to Instagram meme page admins.

These language models are undoubtedly going to become more and more powerful, and people will find more and more novel ways to use them beyond what their creators intended – like making cursed childrens cartoons or weird occult texts

We can be certain that the good people of the internet, given proper access, will continue to use them for their most noble purpose: some of the stupidest jokes you’ve ever seen in your life. 

I’m laughing already.

Follow James on Twitter and subscribe to his newsletter.

Read more from VICE Australia.