Google’s DeepMind artificial intelligence team has released a paper that moves the needle toward implementing some version of Isaac Asimov’s Three Laws of Robotics in real life. Engadget is calling it a “kill switch to prevent an AI uprising,” but the DeepMind team had a slightly more boring title in mind for their paper: “Safely Interruptible Agents.”
Asimov’s Three Laws of Robotics state that robots may not ever harm or allow a human to be injured, while also stating they must follow all orders from humans as long it doesn’t conflict with the first law. The third law states that the robot must act in self-preservation as long as it doesn't conflict with the first two laws.
#DeepMind safety paper in collab w/@FHIOxford on making agents that can be safely interrupted: https://t.co/Q91Nt8Ykxq - much more to come!
— Demis Hassabis (@demishassabis) June 2, 2016
DeepMind’s paper doesn’t go that far, but they do talk about how to make AI think differently about humans in relation to how it interacts with the world. The example used in the paper is an AI controlling a fleet of warehouse robots. If it happens to be raining outside, and humans tell the AI to not go outside, it penalizes the action.
However, bringing boxes from outside is more important than sorting boxes inside (the example assumes there’s only two possible actions), and thus this human interruption (because humans “know better”) should not have any effect on the value the AI places on doing those two tasks in relation to one another. Meaning, how to make sure that an AI doesn’t come up with a solution to stop humans interfering in its task.
DeepMind’s AI technology was acquired by Google in 2014, with the AI going on to later defeat a GO world champion in 2015. This feat was thought impossible before DeepMind’s neural network pulled it off, and it heralds upcoming advances in AI technology the same way computers defeating humans players in chess did.
