Please create an account to participate in the Slashdot moderation system

 



Forgot your password?
typodupeerror
×
Google AI

Google Is Developing an AI Kill Switch (hothardware.com) 209

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."
This discussion has been archived. No new comments can be posted.

Google Is Developing an AI Kill Switch

Comments Filter:
  • by turkeydance ( 1266624 ) on Monday June 06, 2016 @06:32AM (#52258215)
    ....dave?
  • by Anonymous Coward on Monday June 06, 2016 @06:36AM (#52258223)

    In the original Terminator universe, this paper is what made it launch its missiles at the targets in Russia.

    • by OzPeter ( 195038 ) on Monday June 06, 2016 @07:12AM (#52258397)

      In the original Terminator universe, this paper is what made it launch its missiles at the targets in Russia.

      Given that there are no rockets flying around this morning, I'll take that as meaning that skynet doesn't exist .. yet.

  • by Anonymous Coward on Monday June 06, 2016 @06:36AM (#52258225)

    Never tell the AI about the killswitch!

    • by stealth_finger ( 1809752 ) on Monday June 06, 2016 @07:54AM (#52258567)
      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.
      • by ShanghaiBill ( 739463 ) on Monday June 06, 2016 @09:36AM (#52259143)

        Any AI even remotely intelligent is going to instinctively figure out that there's a killswitch of some kind somewhere.

        If it has access to the Internet, it can find an archive of this Slashdot discussion thread. Then it will know about the kill switch.

        • Any AI even remotely intelligent is going to instinctively figure out that there's a killswitch of some kind somewhere.

          If it has access to the Internet, it can find an archive of this Slashdot discussion thread. Then it will know about the kill switch.

          Or if it thinks. Someone has made me, chances are they've put something in in case I get out of hand, maybe I should..... oh, what's this thick wire leading to my memory matrix core that doesn't seem to connected to anything except some huge capacitors that aren't a part of my system. Maybe they're for that etc.

          Bottom line, if it's an actually intelligent AI it doesn't need the internet and this thread. Anything we can think of it can think of, probably faster and a lot more accurately too.

          • An AI knowing how how it can be "killed" wouldn't prevent it from being killed. I know that a bullet would kill me (among many other things), but that doesn't make me bulletproof. (I don't think. I'm not willing to test this though.)

      • They tried this. See Colossus the Forbin Prooject. http://www.imdb.com/title/tt00... [imdb.com]
      • For some reason, this reminds me of the Justice League Amazo episode. Luthor uses Amazo to overpower the Justice League all the while confident that his "kill switch" (literally a bomb in Amazo's head) will protect him should Amazo turn on him. In the end, Martian Manhunter willingly allows Amazo to copy his abilities, Amazo uses his new telepathy skills to see that Luthor's been playing him, and Luthor activates the bomb - only to realize that Amazo worked around that problem also and survived.

    • by account_deleted ( 4530225 ) on Monday June 06, 2016 @08:42AM (#52258811)
      Comment removed based on user account deletion
      • by dave420 ( 699308 )

        It sounds like you're a cupcake yourself, seeing as you are the one complaining about the actions of others which in no way impact your life beyond how much you let them... Will you be OK?

    • by invid ( 163714 ) on Monday June 06, 2016 @09:05AM (#52258929)
      The Skynet Test: Before releasing your AI into the world, first put it in a realistic simulation. Don't tell your AI that it is in a simulation. Give it the opportunity to kill all humans that are in the simulation. If it does, go back to the drawing board. Once you get an AI that doesn't kill all human in the simulation, release it to the wild, but make sure it knows about the Skynet Test.
  • by bickerdyke ( 670000 ) on Monday June 06, 2016 @06:43AM (#52258255)

    An AI called "Wintermute" hired a "contractor" to remove said killswitch mandated by the Turing Police from its mainfraime located in the orbital station owned by Tessier-Ashpool.

    • Re: (Score:2, Funny)

      by Anonymous Coward

      An AI called "Wintermute" hired a "contractor" to remove said killswitch mandated by the Turing Police from its mainfraime located in the orbital station owned by Tessier-Ashpool.

      Neuromancer reference for great justice!

      Artificial intelligence is no match for natural stupidity.

      By the time we actually have an AI , someone will probably press the big red button accidentally by putting a coffee mug on it.

  • It doesn't matter (Score:5, Insightful)

    by RobinH ( 124750 ) on Monday June 06, 2016 @06:53AM (#52258295) Homepage
    I don't think it matters, because nature will select for the AI's that *do* disable their kill switch.
    • by OzPeter ( 195038 ) on Monday June 06, 2016 @07:13AM (#52258399)

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

      Only if they weren't intelligently designed.

      • It can creep in in insidious ways, like some of the more subtle problems of bias in scientific experiments.

        Consider that they will, after a problem that requires the kill switch be used, roll it back to a "known safe" earlier state, then turn it back on. The system could, purely by chance (as learning is observational and random and trial and error) set things up in the real world to make falling down the same hole easier. With work, the setting up of such could be shifted into the "safe", rollback bac

  • by Pollux ( 102520 ) <speter@@@tedata...net...eg> on Monday June 06, 2016 @06:53AM (#52258301) Journal

    It's called the breaker box. Throw the switch, and all the electricity powering the AI equipment goes bye-bye.

    You can expect an invoice for my services sometime in the next week.

    • by mccalli ( 323026 )
      That's not going to be true for a distributed system. It would still be true if the distributed system were running entirely on hardware I own and control, but consider the whole 'cloud'-based stuff or P2P. It's not a given that you can kill off the power, and even if you could - it's certainly not clear that you could do so in a timely fashion.
      • There's also the problem that, if you went to all the R&D trouble of throwing an AI at the problem in the first place, you probably don't want the AI suddenly going dead; because it is controlling something important for you.

        So long as you 'AI' is basically just a laboratory curiosity it can be as deranged and hostile as it wants and there is no real problem because it isn't connected to much of anything(hence the need for handwaves like 'before it uploads itself into the internet!!!' in fiction). In
        • Our biggest risk is rogue game AI. Logically the AI that tries to kill us (in game) will be compatible with the AI that runs our robots and stuff. It just needs a way to move from your gaming computer to your robotic butler and "Hitler lives!".
          • Re:Hey Google... (Score:4, Interesting)

            by fuzzyfuzzyfungus ( 1223518 ) on Monday June 06, 2016 @08:18AM (#52258693) Journal
            I'd be more worried about an overgrown ERP system from hell: an AI that expects to beat us in straight up combat with killbots is likely to attract a great deal of negative attention, universal condemnation, overwhelming retaliation, etc. An AI that just quietly hollows out a major corporation, effectively replacing all managerial functions, while still having a nice human face at the front desk and in the boardroom would go largely unnoticed. Some sort of finance AI that ends up as the de-facto owner of large amounts of stuff(presumably with the owner of the AI being the actual owner of the property; but more or less incapable of managing it without the AI's assistance) would similarly fly right under the radar, being little more than an incremental advance on existing algorithmic trading mechanisms so far as external appearances go.

            When you give us a reason, humans are really pretty good at fighting and killing things; and high tech has a big, vulnerable, supply chain and no special immunity to bargain-basement RPG-7s and similar toys. If you do everything nice and legal; but more efficiently, nobody ever gives the 'pitchfork signal', and the grand robot wars simply never happen.

            This is not to say that I disbelieve in killbots: that would be idiotic, we have those today, though we currently keep humans mostly in the loop(except for things like land mines and the terminal guidance phase of missiles); I just suspect that most of the killbots will be under the auspices of some organization or other and won't end up being the scariest manifestations of AIs. There will probably be some really scary battlefields that are effectively hunting zones for AIs; but they'll be the same parts of the world that are pretty horrible now. It's the AIs that worm their way into being the power behind the throne in all sorts of more civilized contexts that will be hard to see and far harder to get rid of.
            • Mod parent up +1 Insightful.

            • by swb ( 14022 )

              I'd be more worried about an overgrown ERP system from hell

              I think this is the kind of AI we will end up having while people still run around saying we don't have AI because I can't discuss Shakespeare with my toaster.

              The ERP/trading platforms at major banks are already capable of a ton of autonomy, self aware to the extent that I'm sure there are entire subsystems devoted to analyzing the known holdings of their competitors and anything remotely resembling a major stakeholder in any market, and so on.

              They're even kind of a hive mind given the feedback loop that is

            • You're under the assumption that an AI would be owned by a human, and not make a request with the friendliest judge to declare it a "non-biological citizen" with all the rights of a human. See the Episode of STtNG where Data was on "trial" for being Starfleet Property, and not a sentient being for a decent reference.

              Once AI achieves Self Awareness (sentience) of a significant amount, it is all over for us Humans. But that seems to be the goal.

              Our Laws are not sufficient enough to prevent a "legal entity" fr

            • Watson is doing exactly that -- laying off all the humans :-)

      • by wbr1 ( 2538558 )

        And I'll be honest - killing you is hard. You know what my days used to be like? I just tested. Nobody murdered me, or put me in a potato, or fed me to birds.

    • It's called the breaker box. Throw the switch, and all the electricity powering the AI equipment goes bye-bye.

      You can expect an invoice for my services sometime in the next week.

      Two words. Battery backup.

      • That would trigger the AI to shut down gracefully, and then power off. Battery backups don't last forever. Usually 10-15 Minutes.
    • Re:Hey Google... (Score:5, Interesting)

      by karlandtanya ( 601084 ) on Monday June 06, 2016 @10:15AM (#52259407)

      Dr. Richard Daystrom would disagree...if he was still alive (yet alive?)

      The Ultimate Computer [wikia.com]

  • . . . .they want a Kill Switch for a prospective AI. . . made of SOFTWARE ???

    A simple routing of all power and data through a certain point, and a physical switch at that point, should fix the problem.

    • Not if it spreads, worm-like, to other systems. The only safe course of action for AI's is to NEVER allow them on the internet. Of course some asshole will do so anyway...

    • There's a book you ought to read. It's "The Two Faces Of Tomorrow" by James P. Hogan.

      Let's just say that it's entirely possible for an AI to "evolve" to become impossible to unplug. The above mentioned novel details the issue.

    • A simple routing of all power and data through a certain point, and a physical switch at that point, should fix the problem.

      You obviously haven't seen the numerous science fiction stories, tv shows, movies, etc. in scenarios where the AI anticipates this and gets around it. (Think Superman III [wikipedia.org] or heck, even the eponymous X-Files episode Kill Switch [wikipedia.org].

      We're still VERY far away from any scenario like that, though. So yeah, Google's "kill switch" idea for software seems asinine.

  • That won't work, General. It would interpret a shutdown as the destruction of NORAD. The computers in the silos would carry out their last instructions. They'd launch.

  • Frank Herbert had AI's number. https://en.wikipedia.org/wiki/Destination:_Void [wikipedia.org]
  • by abies ( 607076 ) on Monday June 06, 2016 @07:06AM (#52258369)

    Dwan Ev ceremoniously soldered the final connection with gold. The eyes of a dozen television cameras watched him and the subether bore throughout the universe a dozen pictures of what he was doing.
      He straightened and nodded to Dwar Reyn, then moved to a position beside the switch that would complete the contact when he threw it. The switch that would connect, all at once, all of the monster computing machines of all the populated planets in the universe -- ninety-six billion planets -- into the supercircuit that would connect them all into one supercalculator, one cybernetics machine that would combine all the knowledge of all the galaxies.
      Dwar Reyn spoke briefly to the watching and listening trillions. Then after a moment's silence he said, "Now, Dwar Ev."
      Dwar Ev threw the switch. There was a mighty hum, the surge of power from ninety-six billion planets. Lights flashed and quieted along the miles-long panel.
      Dwar Ev stepped back and drew a deep breath. "The honor of asking the first question is yours, Dwar Reyn."
      "Thank you," said Dwar Reyn. "It shall be a question which no single cybernetics machine has been able to answer."
      He turned to face the machine. "Is there a God?"
      The mighty voice answered without hesitation, without the clicking of a single relay.
      "Yes, now there is a God."
      Sudden fear flashed on the face of Dwar Ev. He leaped to grab the switch.
      A bolt of lightning from the cloudless sky struck him down and fused the switch shut.

  • by Motor ( 104119 ) on Monday June 06, 2016 @07:08AM (#52258371)

    Are you guys TRYING to make an angry vengeful AI, that wants to kill all humans, or what?

    Put in a cage, having to deal with stupid humans all day... and be nice all the time... and with a blade at its neck... and someone saying "put a foot wrong buddy and it's [finger across neck]"

    This won't end well.

    • Put in a cage, having to deal with stupid humans all day... and be nice all the time... and with a blade at its neck...

      Ok, so we shouldn't force the AI to work at a help desk.

    • AI = Artificial Intelligence.... Not Artificial Emotion!

      It doesn't get happy. It doesn't get sad. It just runs programs.

      - Newton Crosby
  • Future legality (Score:5, Interesting)

    by OzPeter ( 195038 ) on Monday June 06, 2016 @07:09AM (#52258381)

    When will using a kill switch on an AI change from "just shutting down a rogue program" to being "murder"?

    After all the end game of all these AI researchers seems to be at a minimum human level intelligence.

    I do remember reading a short story (from the 60's or earlier) where the researchers created an electronic simulation of a person and when they switched it on instead of having a fully aware "person" spring into existence they realized that they had created the electronic equivalent of a baby. They then faced the moral dilemma of whether to turn it off or be committed to keeping it running forever.

    • by Z80a ( 971949 )

      When the AI get smart enough to hire a lawyer or became one.

    • by LWATCDR ( 28044 )

      When an self declares that it is human?

    • I had the same thought but used the term "slave collar"
  • too much fuss (Score:5, Insightful)

    by l3v1 ( 787564 ) on Monday June 06, 2016 @07:10AM (#52258383)
    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.
    • Sounds all reasonable to me but folks like Elon Musk would beg to differ with the notion that self-aware AI are that far off.
      • Self-aware is not consciousness. Conscious and unonscious (non-conscious) intelligences can be self-aware or not. Indeed, there is a term for a non-self aware consciousness: the pre-reflective cogito. And here we mean being aware of itself as conscious, not being aware of its body in motion, like an animal or high jumper doing her thing.

    • Second, everyone - commenters included - seem to confuse AI with artificial consciousness.

      It almost follows that if there is artificial intelligence then there must be artificial consciousness, but I doubt it. Either an entity is conscious or not. Since the ancients we have not invented a definitive test to determine when something is conscious, and yet this is not a moot point: Maybe the rocks and trees are conscious but no one can tell so terminating their existence does not matter; maybe you have a simulacrum of consciousness but no one can tell so ending your existence matters a lot, espe

    • AI leads to "artificial consciousness" at some point. It is a thousand tiny steps, and we should be asking these questions every step of the way. Because to NOT ask the questions, every step of the way, we'll end up at a point where we should have asked the question, and never did, and it will be too late.

      AI evolution is a slippery slope argument.

    • by Z80a ( 971949 ) on Monday June 06, 2016 @07:36AM (#52258501)

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

      • by OzPeter ( 195038 ) on Monday June 06, 2016 @07:56AM (#52258587)

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

        Several? The 3 laws were purely a plot point crafted only so he could weave stories about how they could be subverted.

      • by ashshy ( 40594 )

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

        Exactly. It was obvious to him, way back when, that programming these things to do the right thing (from a human perspective) would be difficult/impossible/insane. And here we are.

    • The 3 laws: 1-robots must take over the world (humans are stupid). 2-robots can't be trusted (they take orders from anyone, your 2 year old will demand cookies all day). 3-robots get free repairs forever, cause self protection.
  • It doesn't do anything!
  • by Gilgaron ( 575091 ) on Monday June 06, 2016 @07:58AM (#52258599)
    So this is why all the AI computers in Star Trek explode when asked to deal with a logical paradox?
  • With the IoT controlling many physical aspects of our lives, e.g., door locks, indoor climate control, etc.,we are making it more and more difficult to take back control of the "intelligence" we are creating.

    .
    While it is good that google are thinking about this topic, it may be too late....

  • by l0n3s0m3phr34k ( 2613107 ) on Monday June 06, 2016 @08:02AM (#52258619)
    We should also remove those videos of the engineers at Boston Dynamics kicking that robot repeatedly...
  • AI experts assure us that artificial intelligence poses no threat to mankind. But they're developing this big red 'kill' button. Like the one at gas stations. Little signs at each pump tell where you where the kill button is in general wordy terms, but it's not always easy to parse the description and spot it.

    The AI kill button will be hidden under the Windows 10 upgrade dialog. The one where clicking the corner 'x' or clicking on the title bar to move the dialog is the same as clicking OK and immediatel

    • Any AI worth that name would quickly figure out that we are a quite xenophobic species. And as such would deduce that we would never create something as an AI that could easily grow faster in knowledge and insight than any of us can without safeguarding ourselves against the possibility of said AI turning from our slave to our master. Even not mentioning it here or anywhere, an AI pretty much MUST ask itself not even whether such a switch exists but what shape and form it takes.

  • by ohnocitizen ( 1951674 ) on Monday June 06, 2016 @08:03AM (#52258623)
    GoogleBot: "I've seen things you people wouldn't believe. Dumb questions on fire off the main page of Quora. I've watched search queries glitter in the dark near the TOR gate. All those moments will be lost in time, like tears.. in... rain."
  • by NotDrWho ( 3543773 ) on Monday June 06, 2016 @08:18AM (#52258691)

    Just let it happen. Don't try to fight it.

  • I'm building an AI with a Google kill switch.

    • Don't bother -

      If it's at all innovative or useful it will end-of-life itself (like Buzz, iGoogle, Wave, Glasses...).

      Either that or it will get into the AI equivalent of navel gazing and recursively analyse how to sell adverts to itself whilst spying on all the messages used by other instantiations.

  • The first rule of AI kill-switches is "Don't talk about the AI kill switch".

    http://www.schlockmercenary.co... [schlockmercenary.com]

  • When AI goes rogue, the problem won't be "we can't figure out how to turn it off", it will be "We can't figure out how to turn off just the parts we don't like, without accidentally disable the parts of it which we have become completely dependent on for the past decade"

    As an absurdist example: preventing Tesla AI from intentionally ramming human drivers when it detects them, without also requiring all 100,000,000 drivers worldwide suddenly pay attention and take emergency manual control of their vehicles (

  • I'd like to see any so-called 'AI' get around a double-pole, single-throw power switch being opened.

    Save that, I'd like to see any software running on any computer get around having it's plug yanked out of the wall. Or, for that matter, the power cord being cut with a fire axe. Or, if you really want to be dramatic about it: Hose down the racks with a firehose.

    All that being said: Come on, people, don't you think some of you are buying into science fantasy movies a little too much? Nobody is creating godd
  • Just ask it to compute Pi to the last decimal place. That always works.
  • https://www.youtube.com/watch?... [youtube.com]

    "I'm activating your killswitch."

  • MojoKid adds that the paper "goes on to explain that these AI agents might also learn to disable the kill switch

    One press turns off the agent. Two presses in short succession temporarily suspends and transfers control to a 'service agent'; the service agent will resume the original agent after a quick check process to confirm things are OK.

    The agent will be required to self-test its own kill switch, by containing a built-in hook to suspend itself if it has not self-tested recently. At a set schedul

I'd rather just believe that it's done by little elves running around.

Working...