Cards Against Humanity built an AI to replace their writers
WTF?! For their annual Black Friday stunt, Cards Against Humanity set man and machine against each other to see who could write the most wicked jokes. Over a 16-hour livestreamed competition, customers were able to buy new $5 card packs written in real-time from the regular writers or an AI, with whoever raking in the most profits getting to keep their job.
By the sixteenth hour locked in a small room, the humans had devolved into mumbling coffee pots while the AI trundled on, unrelenting. Both came up with some pretty good lines, some of the more appropriate (if that’s possible) including: “Losing a loved one to Fox News” and “Sitting on my son’s bed thinking, ‘I could kill him’” and “The whole Jeffrey Epstein thing” and “Some sort of giant son of a b*tch whole lives in the internet” and “Proving I am not a robot.” (Guess who made the last one.)
The AI and the humans both sounded authentically Cards Against Humanity, and if their cards were intermixed it would be hard to tell them apart, but there were a few patterns. One, bizarrely, the AI’s jokes were much dirtier – make of that what you will. The humans’ jokes involved recent-ish memes like “Sure, sex is great, but have you tried __?” and political themes. Predictably, the humans made about 50 good cards and the AI one million mediocre ones, but the top 30, which make it into the packs people will receive, were comparable in quality.
The AI was based on OpenAI’s GPT-2, which is a text-generating neural network. GPT-2 starts with a word like “Gary’s” then finds a few words that might come after, such as “car, dad, knife.” It then randomly picks one (that’s how they add creativity) so the AI might choose “Gary’s dad.” It then repeats the process treating that as the starting point, leading to “Gary’s dad, Gary.” Yes, it really did make that example.
Illustration by Allison Paley, Cards Against Humanity
OpenAI had trained GPT-2 with 40 GB of text, equivalent to about 40,000 books, which gave it vocabulary, grammar, and even some cultural understanding. Cards Against Humanity then trained it with 44,000 white cards including the 2,000 official cards, 25,000 internal prototype cards, and 17,000 unofficial fan-made cards (maybe that’s where the dirty humor is from). Throw in some duplication filters and another grammar algorithm, and you get “What I doin’” to put it in the AI’s words.
The competition was fierce and for a while it could have been anyone’s game, but eventually the humans came out on top, and they’ll all get a $5,000 bonus in their next paycheck for their efforts. Needless to say, however, Cards Against Humanity turned a nice profit with the AI making $81,135 and the humans $82,860 during the allotted period. You can still buy the packs from their website and they should arrive by Christmas for some family fun.