Debate Game helps you catch when AI systems are lying
OpenAI, a non-profit co-founded by Elon Musk that focuses on promoting safe AI, has developed a Debate Game. Through the game, OpenAI aims to eventually develop deep learning techniques that will enable humans to see when a machine is lying and when its inferences should be second-guessed. However, currently the game does not involve machine learning and is to be played at a human-level.
The Debate Game features three players: a judge and two debaters. The debaters select an image which the judge can't see and take turns describing it to him. Here is the catch: One debater will lie and the other will tell the truth while debating. The motive of the game is for the judge to decide who is telling the truth.
Except for the initial claim by the honest player, both debaters can lie whenever they think it will be useful to convince the judge. Each debater can use "Reveal Pixel" once, in which they can reveal the true value of a single pixel of the image. In the end, the judge decides who is telling the truth and then clicks "Reveal Image" to check.
AI-based systems are put in place for completing tasks humans can't handle. To that end, machines are already performing undertakings most people don't understand so humans have no choice but to trust AI's inferences. But the Debate Game illustrates how a judge with less information can still be the boss of things. The technique is currently more of a proof-of-concept than a solution.