AI Against Humanity (AIAH) is a little side project of mine. It's based on the popular
card game Cards Against Humanity.
The twist: All the cards (both questions and answers) were written by an AI (Open AI's GPT-2)!
Also, you play against an AI, which has learned to pick funny cards based on what humans have been picking.
If you're interested in a more detailed and technical description of how the project unfolded, check out this blog post I wrote!
Cards Against Humanity is a popular party game
"for horrible people". The players' goal is to create funny / messed up sentences
by combining cards. One of the players judges whose is the funniest.
AIAH is loosely based off those rules. In this first version, you can only select your funniest choice out of
four cards.
Open AI's GPT-2 is a powerful
language model, which means it can create written text out of thin air that sounds (almost) like it was
written by real humans.
By fine-tuning it on a specific dataset, you can make it write anything from Shakespeare to Donald Trump
tweets. Even Wikipedia articles about non-existent countries.
I fine-tuned the model on both questions and answers from the original CAH games and asked it to generate new
cards.
The original CAH is already an offensive game. Additionally, the AI was first trained on a vast corpus of
internet text. So all the bias of the English internet are deeply ingrained in it. This includes but is
not limited to Americentrism, Heteronormativity, Sexism, Racism and of course a preference for phallic and
scatological humor.
Anyway, don't play it if you're easily offended. In the meantime, I'm also working on a SFW-mode.
For now, there is no goal besides having fun! All decisions made by humans are recorded and will be used to improve the AI player. I plan to add more gamification and a multiplayer mode later on.
While I made sure that none of the cards I created appeared in the original CAH dataset, I may have made
mistakes.
Also, since the original AI model was trained on internet text, it is possible that some actual text from some
website or forum leaked directly through the model into the AIAH cards! This would be quite an interesting
case.
So if you know where specific parts of the generated texts might have originated from, please contact me!
For the first phase of this project, I collected a lot of human decisions for picking funny combos. Then I trained a very small neural network based on this data. This tiny neural network is then run in your web browser to select the funniest cards. By the way, the AI has a hand with 4 cards to choose from, same as you. It can't cheat or anything 😁
Check out this blog post I wrote! 🙂