Deviation Game

Defeat an AI in this co-op party game!

Deviation Game is a co-op party game for 2-6 players that pits human creativity against AI perception. Your goal: draw secret prompts like "Flamingo" in a way humans can understand, but an image recognition AI can't!

Up to 6 people can play with one copy

If just one person owns the Steam game, everyone can join by watching that player’s screen. Up to 6 players can join for free using their smartphones’ web browsers via a QR code.

How to play

Join using your own device

Using your mobile device, scan the QR or enter the room code at outdraw.ai to transform your device into the game’s controller!

The drawer picks a prompt

The drawer first picks a secret prompt from categories such as objects, actions or even concepts.

Time to deviate!

Then they just gotta draw it! But remember, the aim is to trick the AI while making sure your friends still get the picture!

Everyone (plus the AI) guesses

Once the drawer is done, all the guessers will submit their best guess, including the AI! Choose carefully, you only get one chance!

The answer is revealed!

If you fail to fool and the AI guesses right, then humans lose! If the AI fails and humans guess right, then humans win!

AI to enhance (not replace) creativity

By embracing the medium of games, Deviation Game aims to use AI as a collaborative tool that seeks not to replace human creativity but to enhance it. Through the game, players learn to express themselves in new ways to deviate from what the AI has already learned, making for a fun and brain-teasing party experience, where you’ll create things no one has ever seen before! Flipping Alan Turing’s revolutionary 1950 Imitation Game on its head, Deviation Game encourages players to fool the AI instead of the AI trying to fool them.

Keeping consent clear

At the end of each game, players can consent to their drawing being used to potentially train the AI... or not! Allowing you to decide how the game may evolve. Collection is opt-out by default, and all data without consent will be deleted. By working together, players from around the world can collectively uncover the biases in large AI models by exposing their blind spots in interpreting diverse, culturally specific drawings.

Played by humans around the world

Deviation Game was initially showcased as an art installation at Civic Creative Base Tokyo in 2023. The work has since been shown in globally renowned exhibitions, such as Ars Electronica, Taipei Digital Art Festival, and Now Play This, engaging audiences of all ages and sparking critical discussions around AI and creativity.

About the creators

Deviation Game was developed and self-published by a group of good friends consisting mainly of game designer Tomo Kihara and art-design duo Playfool (Daniel Coppen & Saki Maruyama). With a focus on play, their practice centres around designing tools that foster creativity and making artistic interventions that engage with urban space and society. Their previous work How (not) to get hit by a self-driving car received an Honorary Mention in the S+T+ARTS Prize 2024 at Ars Electronica.

Deviation Game is their first experience to self-publish an indie game.

FAQ

Q. How exactly does the AI guess?

A: In each round, the AI will return numerous guesses, depending on the number of players in the game. If one of those answers are correct, AI wins!

Q: Doesn’t the AI just have access to all the prompt words? Surely it’s cheating!

A: Not at all! While some examples are provided to the AI to give it more context, it is freely guessing each answer it submits, which is why sometimes they’re a little off-base.

Q: What AI model are you using?

A: Throughout Early Access, we are experimenting with various AI models via the service OpenRouter. In the future, we plan to introduce a feature that allows players to choose which AI model to go up against.

Q: Does the game’s AI get smarter after each game?

A: No, the AI is not learning in real-time. However, custom improvements to the AI can be made with a future update after enough data has been collected to retrain the model.

Q: How is our drawing data being collected?

A: We only collect drawing data that has been given explicit consent. Collection is opt-out by default, and all data without consent will be deleted.

Credits

Project: Tomo Kihara, Playfool (Daniel Coppen, Saki Maruyama)

Engineering: Kye Shimizu, Daiki Hashimoto, Jasper Stephenson

Music: Plot Generica

Logo: Yu Miyama