What is today an old sci-fi fear of technology taking control over humanity, could very well become reality one day – or at least a real threat. That’s why scientists today have started thinking ahead and developing a kill switch for artificial intelligence (AI).
Two researchers, Stuart Armstrong, from the Future of Humanity Institute at the University of Oxford, and Laurent Orseau, from Google DeepMind, have already published a paper on why and how future AI machines could be prohibited from learning to override human input.
With their solution, humans will forever remain in charge of machines. “It is sane to be concerned – but, currently, the state of our knowledge doesn’t require us to be worried,” said Dr. Orseau.
However, he also agrees that it’s smart to start working on AI safety before any kind of problem arises. AI safety is concerned with making sure the algorithms providing the machines with learning power work the way they are supposed to.
The research of Orseau and Armstrong focuses on how scientists can be sure AI machines can be interrupted by the humans in their charge – at any time, without the machines themselves learning the code for the kills switch. This way, they cannot overcome or prevent human intervention.
“Now and then it may be necessary for a human operator to press the big red button to prevent the agent from continuing a harmful sequence of actions,” they wrote in their paper, called “Safely Interruptible Agents.”
The challenge is to find a way to disable the AI machine without the learning agent learning how to disable the red button, for example.
It’s important to be prepared, the authors explained, since it’s unlike for AIs to behave “optimally” at all times. Thus, the machines we create should not “be able to disregard human attempts to stop or interrupt its functioning.”
For example, an AI created in 2013 was taught to play Tetris, but in order to avoid losing a game, the machine had learned to pause the game indefinitely.
Other scientists also agree to the importance of a kill switch, such as University of Sheffield AI expert Noel Sharkey. He also told the BBC that the ideal situation would be if the AI program could stop itself when it detects that something is wrong.
Image Source: Flickr
Deborah Campbell
