We wanted to learn more about how online bots and AI in general should interact with humans (and vice versa), so we decided to create a learning center that allows us to study how human interact with bots and create a more holistic dataset for our AI.
Several brands have chat bots that are rigid and don’t provide the level of service their customers expect. In addition, most bots are programmed by a small subset of developers which in turn often times leads to human biases and limited data.
We figured the best way to demystify the machine learning process while allowing user to interact with the AI was to model it after a game our audience knew and loved — Cards Against Humanity. In AI Against Humanity, users take turns asking and answering questions. The twist: users have to guess who is a human and who is a bot.
For our bot to learn human behaviors and traits, we needed a diverse audience to play the game. So we took AI Against Humanity on the road! The game has been played at dozens of conferences where at each event we’ve had up to 10,000 questions and answers submitted.
Our bot learned to pick up on sarcasm, typos and various personalities, ultimately making it harder for users to identify the bot from the human. In fact, 40.3% were fooled and incorrectly identified the boy. We have used these models to better streamline classification of content which allows us to supports our clients’ current business challenges.