Abel Peirson and Meltem Tolunay, two researchers at Stanford University, recently posted a paper to the arXiv preprint server detailing a new machine learning model they created that’s capable of creating pretty convincing memes:
Peirson and Tolunay trained their machine learning algorithm with a dataset of more than 400 types of memes with multiple captions that they pulled from memegenerator.com using a Python script. To simplify the task, they focused only on “advice animal” style memes, the kind where an image of a specific character like “socially awkward penguin” is overlaid with a caption of text that represents the traits of that character, usually with a humorous observation.
“This allows for relatively simple collection of datasets,” the researchers wrote. “In this paper, we specifically refer to meme generation as the task of generating a humorous caption in a manner that is relevant to the initially provided image, which can be a meme template or otherwise.”
After training the algorithm on what kinds of captions go with which images, they were able to test it out by having it generate memes over and over. Some of them didn’t make a lot of sense:
But others were pretty solid:
The pair measured how successful the neural net was by having other people view the memes and judge them on whether they thought they were real or machine-produced, and how funny they were. Testers were able to identify the computer-generated memes 70 percent of the time, which is better than chance but still left some uncertainty, according to the paper.
“The average meme produced is difficult to differentiate from a real meme and both variants scored close to the same hilarity rating as real memes,” they wrote. “Although this is a fairly subjective metric.”
Taking our jobs is one thing, but now computers are beating us at our meme game? I’m starting to think we’ve gone too far.
Get six of our favorite Motherboard stories every day by signing up for our newsletter .