Humanity doesn’t trust robots. Really, you don’t need to look any further than our movies to see that. People want to see the Terminator far more than they want to see Johnny 5. Yet Google, supposedly made up of humans, has been slow to acknowledge this. Thankfully, though, they’ve just admitted they’re working on a way to keep the AIs they’re building in check.
In a long, math-heavy paper called Safely Interruptible Agents, Google and Oxford University’s Future of Humanity Institute lay out how to create an off-switch for an artificial intelligence or, more feasibly, an advanced intelligent agent, like a factory robot. The issue is that increasing we’re using systems that give positive feedback to robots and AIs when they do something we like, and negative feedback when they do something bad.
It’s the classic human dilemma that if something feels good, we’ll keep doing it and damn the consequences. This means that there’s a pretty good chance that at some point we will invent a junkie robot who keeps doing something that feels “pleasurable” as often as it can to the detriment of all else, like human safety. So the idea is to shut it off before it gets to that point. And yes, they’ve planned ahead for the plot twist where the AI removes the off switch so it can continue to launch the nukes. Or as they put it:
However, if the learning agent expects to receive rewards from this sequence, it may learn in the long run to avoid such interruptions, for example by disabling the red button— which is an undesirable outcome. This paper explores a way to make sure a learning agent will not learn to prevent (or seek!) being interrupted by the environment or a human operator.
The short answer is that as long as we code them properly, we can turn them off no matter what and they won’t be bothered by it. The long answer is, well, read the paper, but basically we can counterbalance this robot id with a big red button that makes it dial it back. Thankfully, this likely won’t ever apply to military robots or intelligence. Nobody wants a robot to feel awesome after it guns down a human.
(via Gizmodo)