This AI Meme Generator Captures Our Broken Sense of Humor

If artificial intelligence is stealing jobs, the memelords should watch their backs.

This AI meme generator churns out weird, bizarre, often-hilarious image macros using a deep convolutional network. It was created by Dylan Wenzlau, the founder of image hosting platform and meme generator Imgflip.

In an article published on Medium, Wenzlau explained his process in detail. A deep convolutional network is a class of neural networks usually used for images, and Wenzlau trained his using machine learning platform Tensorflow and Keras to build a text generation model that would match up with 48 distinct meme formats. He also published the full code on Github for anyone to try.

“The network was trained on public images generated by users of the Imgflip Meme Generator for the top 48 most popular Meme Templates,” according to the site.

The generator is called “this meme does not exist,” a reference to earlier websites that used machine learning to generate everything from highly realistic humans to anime characters or feet with a click.

A lot of what the generator came up with made me snort-laugh out loud, and not only because they’re absurd (but sometimes only because they’re absurd). Sometimes they make sense, but most of the time they don’t—and that makes them funnier. They’re relatable and bizarre at the same time. They’re surreal. They are, in a word, dank.

Even hand-crafted memes are predictable and in a way, algorithmic: They depend on a specific format to work, and a pattern that the viewer recognizes. They’re inside jokes, but because they’re spread between strangers, they have to be formulaic enough for a lot of people to be “in” on it.

Here are some of the best ones I generated after a few clicks around the site:

“Nothing about the text generation is hardcoded, except that the maximum text length is limited for sanity,” according to the site description. “The model uses character-level prediction, so you can specify prefix text of one or more characters to influence the text generated. Using someone’s name or other short text as a prefix works best.”

In making the generator, Wenzlau didn’t filter profanity from the public data used for training, he warned on the site. A troublesome aspect of machine learning is that, without proper interventions, systems can pick up on biases and prejudices in the training data. I did encounter one slightly racist meme, but most of them are very pure and good.

And often, painfully topical for the times we’re in.