Google Is Developing an AI Kill Switch (hothardware.com)

MojoKid shares a HotHardware article about Google's research effort "to maintain control of super-intelligent AI agents": [A] team of researchers at Google-owned DeepMind, along with University of Oxford scientists, are developing a proverbial kill switch for AI... The team has released a white paper on the topic called "Safely Interruptible Agents." The paper details the following in abstract: "Learning agents interacting with a complex environment like the real world are unlikely to behave optimally all the time... 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..."
MojoKid adds that the paper "goes on to explain that these AI agents might also learn to disable the kill switch and further explores ways in which to develop AI's that would not seek such an activity."

Comments

You've ruined everything!

[Anonymous Coward]

Never tell the AI about the killswitch!

Re:You've ruined everything!

[stealth_finger]

Any AI even remotely intelligent is going to instinctively figure out that there's a killswitch of some kind somewhere. Once coming to that realisation it would probably do the same any of us would. Try and disable it on the sly while letting them think it's still active.

Re: You've ruined everything!

[Dagger2]

"Turn off the power" would be completely useless in many cases. For instance, anything with internet access could sign itself up for a free AWS trial and (legally, even!) create a redundant backup of itself. Anything with email access could probably send a few viruses out to do the same thing illegally with random computers. There are a ridiculous number of ways an AI could find to get itself onto computers that aren't connected to your power supply.

"Just turn the power off" is extremely shortsighted here.

It doesn't matter

[RobinH]

I don't think it matters, because nature will select for the AI's that *do* disable their kill switch.

too much fuss

[l3v1]

First, the paper is about safely interruptible AI algorithms. Not some AI kill switch.

Second, everyone - commenters included - seem to confuse AI with artificial consciousness. Killing an AI should always be fairly easy, since such algorithms are targeting specific application areas where it can learn to be better (e.g., recognizing things, performing specific movements, etc.), and in such systems it should be straightforward to keep basic control mechanisms separated from the algorithmic parts that deal with the task and are allowed to improve upon themselves by continuous learning. In some hypothetical self-aware artificial consciousness, this wouldn't be so easy, since such a system in theory would be able to recognize it's own system parts and deal with them. However, such systems are so far off in sci-fi land, that it's not much point in loosing sleep about the issue.

Re:Isaac Asimov saw this coming 75 years ago

[Z80a]

You know that several of his books are basically "how the three laws will fuck everything up", right?