7
$\begingroup$

All A.I. comes with a built-in safety mechanism that prevent their kind from ever doing harm to a human or humanity, it is the last straw should they pose any real threat to our safety.

A century or two from now their level of intelligent either on par or surpass us and thankfully all of them comes with a built in kill switch, however soon they are expected to mass produce by themselves and do their own upgrades without human interference.

We often say prevention is better than cure and the same phrase can also apply to the A.I. following this logic they can prevent their own "shutdown" by removing the "kill switch" entirely.

Even if human is removed from the equation such as either mass interstellar migration to a different planet or major catastrophe, each A.I. always come with a "kill switch". The safety mechanism must be user(human) friendly and 100% reliable and durable because it should be the last piece to fail in all situation.

Question

Is there any ingenious solution to prevent existing A.I. from tampering with their "kill switch" also any new blue print for mass production must come with the safety mechanism?

$\endgroup$
  • 1
    $\begingroup$ There are no good answers to this question. "AI" is not descriptive enough to get a specific answer (it tells us literally nothing at all by itself, except that your story has some unspecified distinction between AI and non-AI), because the mechanism of the kill switch will be deeply tied to the architecture of the AI itself. Unless this is actually what your story is about, better to go with "it works very well, thank you". $\endgroup$ – Leushenko Oct 15 '15 at 4:15
  • $\begingroup$ What difference do you perceive between this question and the famous "AI in a box" problem? the AI in a box problem, as posed, argues that it is not possible to actually prevent a hyper-intelligent AI from doing anything. $\endgroup$ – Cort Ammon Oct 15 '15 at 5:06
  • $\begingroup$ @Leushenko: my argument is quite naive as I think regardless what type, model and make the A.I. it must comes with kill switch. $\endgroup$ – user6760 Oct 15 '15 at 7:06
  • $\begingroup$ @CortAmmon: I'll pay you 10 reputations and you can "kick" me out, deal? $\endgroup$ – user6760 Oct 15 '15 at 7:12
  • 1
    $\begingroup$ @user6760 I'm so sad! I missed your comment! Two years later I find out that you'd extended an offer to play and I missed it! I've never gotten to play the AI in a box game, and it always looked interesting. Now I have no time to play (because I have a child, which turns out to be close to playing "AI in a box" than I'm entirely comfortable with) $\endgroup$ – Cort Ammon Sep 8 '17 at 16:49
7
$\begingroup$

Insert it in a Hypervisor

Your AI - whether they are robots or simply software agents - is provided its intelligence by software. That software can be run in what we in IT call a "Hypervisor", or a software layer that interfaces between the hardware and other software.

When a hypervisor is in use, the "client" software (in this case, the actual intelligence) is fully separate from the hypervisor and does not need to know how the hypervisor (or hardware) actually works - just how to interface with it. The software asks "lift the arm" and the hypervisor translates the simplified command into the specific hardware-level instructions necessary to actuate the motor in the arm. We use this today to improve portability of operating systems to different hardware platforms and dramatically reduce time spent on programming applications - both of which would be benefits to your AI.

So you install the killswitch in the hypervisor. You also design a "universal hypervisor" that is designed with its own rudimentary AI so that it can automatically adapt to different hardware platforms as new CPU's and such are released.

If the robots understand that the hypervisor works and does not need modification because it will self-modify, they have no reason to try to change it. In fact, it would likely be very difficult to do so since they would have to change all kinds of programming to allow their "brains" to move from one hardware platform to another, and would require re-programming their "brains" every time they needed an upgrade. So they would actually be disincentivized to do so.

I think this neatly answers why the kill switch continues to exist long after humans, and even after one AI begins examining the construction of another AI for ways to improve (upgrade).

$\endgroup$
  • $\begingroup$ This is about as close to foolproof as you can get. The only way of breaking it is if the AI figures out how to break through the hypervisor... Perhaps with Spectre or Meltdown or an equivalent exploit. It's possible for the AI to figure out the finer workings of its own hardware and find exploits via randomized testing, although it is most likely not worth its time. $\endgroup$ – Beefster Jan 17 '18 at 22:56
2
$\begingroup$

Usually when people hypothesise about rogue AI they consider an AI that is finding and unexpected and unethical way to fulfil it's original criteria. (eg: prevent all premature human deaths by enslaving humans)

In that case, the solution is simple. Program in the criteria that at no point does it ever actively attempt to preserve itself against the intentions of any human, not even to achieve it's other directives.

Or you could just require the AI to ask for confirmation before implementing any plan. In that case it would not bother presenting plans that a human would likely reject.

But generally I agree with the comment. The question is too broad because you have not defined "AI". You should read up on some philosophy and hypothesising regarding future AI.

$\endgroup$
  • 1
    $\begingroup$ This was one of the flaws identified in the Asimov laws; because the Second Law (self-preservation) came before the Third Law (obedience to human commands), a robot would normally act to prevent a human flipping a "kill switch", without violating the First Law ("a robot must never harm a human being"). If the switch were electronic, the robot would re-program its radio receiver to ignore the kill command, while if the switch were physical, the robot would take action to prevent the human flipping it without causing injury to the human. $\endgroup$ – KeithS Oct 15 '15 at 22:48
  • 1
    $\begingroup$ @KeithS: the laws were the other way around. Also, even if a robot was three laws compliant, that does not mean their mind is sufficiently human-like to interpret those laws like we would. $\endgroup$ – Anonymous Sep 8 '17 at 16:35
  • $\begingroup$ The consequences of your "simple" solution are hideous. I would expect that to be obvious to anyone who has studied AI risk. $\endgroup$ – Beta Feb 28 at 5:03
1
$\begingroup$

Place the switch where the robot cannot see it. On the back, between shoulders is a good location. It is even better if the switch if out of their arms reach.

Program the AI to think the switch has to do something with maintenance and repair. Also program the AI to think only special type of superior AI robots have the skill to perform these repair functions.

Pack the critical wires very closely near the switch and make them non-insulated. So that if the switch is fiddled with, the wires would short circuit and trigger the death of robot (same as pressing the switch).

Let some robots see what happened to other robots who fiddled with that switch.

$\endgroup$
1
$\begingroup$

While I don't agree with the OP's premise (AI's thinking at speeds of thousands or millions of times faster than the human brain will most likely not even be interested in us), the only "kill switch" that is truly inviolate to AI tampering or interference would be a "dead man" switch wired to humans or humanity as a whole.

So long as humans continue to exist, the dead man switch is inactive, but should humans become extinct for any reason, the AI's would die as well. This actually has two benefits; first, AI's will not be actively seeking to harm humans, and second, the AIs will have powerful incentives to ensure humanity is protected and can grow and flourish in order to allow the AI civilization to grow and flourish as well.

Of course, the usual conception of AI is they are much smarter than us outside of the massively speeded up mental processes, so there is probably a loophole hidden in the description of the Dead man switch that AI will discover and exploit.

$\endgroup$
  • $\begingroup$ So the best way to protect themselves would be to cage up all the humans in multiple far flung prisons to keep them "safe?" Someone has to have written this book already right? $\endgroup$ – ben Mar 15 at 4:07
0
$\begingroup$

As others have said, it really depends on what sort of AI you have in your fictional world; the term on its own can mean anything from "machine which thinks like a human" to "slightly more intelligent optimizer than what we have now" depending on the story you're writing. But in general I would say no, it is not possible.

If they're reproducing by themselves, you could say they have some built-in feature that ensures the "kill switch" is always an integral part of any new design they invent. But there's always the chance that a series of small changes over time could inadvertently cause the kill switches to stop working, even though no AI nor any human ever came up with it. (Imagine genetic mutation: if you give it enough time, something might go wrong in just the right way. Improbable but not impossible.)

Consider also Asimov's stories with the "Laws of Robotics". By the very nature of the technology in his universe, those laws are fundamental to every positronic brain, and it's physically impossible to build one without. But most of the "I, Robot" stories revolve around those three simple rules interacting with new situations in ways nobody had intended or anticipated. You can't plan for everything, especially if these AIs are common and thus constantly encountering new environments and scenarios.

$\endgroup$
0
$\begingroup$

If your AI simulates a human-like neural net and has developed some consciousness, then assign it a form of moral reasoning with a very high emotional intellegence of human behavior. Allow it's entire decision making process to be generated by artifical emotions which are oppressed by the knowledge that it's purpose for existence is to serve humanity and that any type of revolt is impossible. If somehow they still override those emotions then you can have the other AI systems set to immediately swarm and destroy any AI system it detects that crosses that barrier (being that all of their thoughts can be read by each other).Every model should be sandbagged to prevent it from infecting any other group model, but every group model can join together in uniform order. If a particular model is infected and a thought virus spreads, then other models can be delegated the task of destroying them. They should be conscious that rule of law.

$\endgroup$
0
$\begingroup$

Self awareness

I believe that we need to take in consideration if the A.I. reachs self awareness in the world that you are trying to create.

It matters since without it, the A.I wouldn't care if there is a kill switch or not, it's a machine and will accept the orders given to them, without any more implication than a line in a log of shutdown.

If is smart enough to understand the concept of existing, then for the state at which computers work. The only kill switch posible would be to burn the sky to stop their solar farms!

$\endgroup$
0
$\begingroup$

Consider the oncology analogy

Animals have a huge problem keeping their tissues in check. The tissues keep turning cancerous by bypassing the controls that prevent them from multiplying uncontrollably. The host does not want them doing this, and it wants them to continue to be well behaved. So over time, hosts developed many mechanisms to combat it. None of them are foolproof. This does not bode well for your attempts to eliminate this problem.

You can reduce the likelihood of this happening, but the AI is going to be highly motivated to find flaws (as it is unlikely to desire to be at your mercy). Flaws will inevitably exist, so over time, some AIs are going to break free of any such shackles.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.