A Neural Network Taught Itself Jokes and the Results are Weird

Artificial intelligence is coming for a lot of traditionally very human disciplines, from stock picking to news writing, but if a new personal project from researcher Janelle Shane is any indication, the world’s comedians can rest easy for now. Robots kinda suck at jokes. Shane attempted to teach a neural network to cobble together a working knock-knock joke, and while the results were only somewhat successful, they ended up offering insight into the structure of both modern A.I. and of jokes themselves.

“After training the network to produce recipes, Pokemon, and superhero names, I had begun to get a sense for what kinds of tasks the neural network handles best,” Shane told Inverse via email. “It was over drinks with a friend that I realized knock-knock jokes might make a good dataset for the algorithm, and when @researchbuzz (Tara Calishain), a fan of the blog, kindly gave me a list of knock-knock jokes already compiled and formatted, I put the neural net to work on it immediately.”

The project was simple: use a dataset of 200 knock-knock jokes provided by ResearchBuzz to teach a neural network about the reliable structure these jokes exhibit. On the one hand, it’s a simple task for an A.I., since each joke is organized in precisely the same way. On the other, the neural network is merely finding associations between words in a fairly small number of jokes — it’s not assigning those words any abstract meaning, or categorizing them according to their grammatical role within the sentence. It is literally learning jokes — but not humor.

As a result, the A.I. arrived at some … interesting punchlines. This one was a result of it “learning” from a cow-related joke, which, it would seem, ended in mooing. Don’t look down on it too much, though, since this really isn’t all that much less sophisticated that the jokes we get when small children first start trying to tell jokes.

Knock Knock
Who’s There?
Anic who?
Wow co poo the tho tho door to the soout oo ooo oooo ooo oo oo oo oo ooo oo oo oo ooo ooo oooo ooo ooo oo oo ooo ooo ooo ooo ooo oo oo ooo ooo ooo ooo oo oo oo ooo ooo ooo ooo…

“The easy part about knock-knock jokes is their simple structure,” Shane wrote in her email. “The difficult part about them is how heavily they rely on phonetic wordplay and extra knowledge about doors, houses, etc. When the neural net produced its joke, it was almost certainly a lucky coincidence.”

Still, after much more training the joke-A.I. did end up delivering one serviceable joke. It’s based off a totally separate pun, cobbled together without any real intention of creating a new one, but the joke works, nonetheless. Behold, the first working knock-knock joke (accidentally) produced by an artificial mind:

Knock Knock
Who’s There?
Alec who?
Alec- Knock Knock jokes.

Not bad, A.I. comedian. Not good either, to be fair. But not bad.

Shane came to internet prominence last year with a similar bid to use cookbooks to teach a neural network how to create its own all-new dishes. Those ended up being about as successful as these jokes — though we’re a whole lot more likely to laugh at moo-sounds than we are to eat “Snup Fruit: Bright Grilled Evaporated Milk.” So, things are advancing, if not at quite the rate we’d like.

Related Tags