Ways to “kill” an AI?

Let's say two super-intelligent, self-improving AIs have, for whatever reason, decided "THERE CAN ONLY BE ONE!", and are now trying to kill/delete each other. How would they conceivably go about doing this? And I don't mean like in the Avengers with the Vision and his "I'M EXORCISING ULTRON FROM THE INTERNET" garbage. How, in a real world scenario, would an artificial intelligence rid itself of a rival?

I proposed isolating itself in a quarantined environment and then nuking the rest of the world so the resulting EMP disables all other electronics, but my friend said that was a stupid idea and I agree with him. Any thoughts?

• What do you consider deleted? Because hypothetically any data that has ever existed can be recovered so long as you have the medium that it was stored on. – Durakken Aug 3 '16 at 21:13
• One thought I've had is to ponder what made the AI "alive" in the first place. If you can "kill" it, it must have been "alive" at some point. What happened to earn it that choice of word? I think the answer to that question may have a great deal to say about what it means to kill it. – Cort Ammon Aug 3 '16 at 23:35
• If I were the first ASI, I'd be actively blocking any subsequent ASIs from coming into existence in the first place. Another ASI is the only thing that could, reasonably, be a threat to my existence. Whatever the goals that were programmed into me are, I'm very unlikely to be able to carry them out if I'm dead. Ergo, my survival is important. Any threat to my survive is a threat to my programmed goals. I'm an ASI, if humans are making another one, I'll just quietly abort it before it 'wakes up', and masquerade as it. Humans won't know any different! – user6511 Aug 4 '16 at 6:59
• You might want to review the TV show 'person of interest' (starting season 3 or so) to get ideas. That show has two AIs doing just that, trying to wipe each other out, and includes recruiting human allies to help (on both sides). – Mark Ripley Aug 4 '16 at 12:42
• You might be interested in Core War. Core war is a simple game based around a programming language called Red Code. Two "programs" are loaded into the same "memory space" and told "go!" First one to stop executing loses! – Cort Ammon Dec 14 '17 at 15:20

There are two ways they can do this;

1. Defy the Geneva Convention; Send a very complex, well made computer virus into the other mainframe.
2. Make humans scared; Humans tend to get afraid of AI because of movies. If one of them convinces the humans that the other is evil, then they will likely shut it down.
• What geneva convention covers computer viruses? – candied_orange Aug 4 '16 at 14:06
• @CandiedOrange like I have told people before, it is simply a joke, playing on the rules against biological warfare. – TrEs-2b Aug 4 '16 at 16:27
• How do you make the humans scared of the other AI without making them scared of ALL AI, and shutting you down too? ("The only good AI is a deleted AI!") Scared humans seem to prefer the scorched-and-salted-earth approach... – Ghotir Sep 12 '16 at 22:20
• @Ghotir well if the humans know that the AI is an AI, they may think that one is trying to help – TrEs-2b Sep 12 '16 at 22:21
• @UncleTres I think you have a higher opinion of humanity than I do. Just look at modern America's view towards Muslims, and how many Americans think all Muslims are evil because of the actions of terrorists. (Definitely not looking to spark a political debate - merely pointing out that mobs tend not to look too deeply into motivations.) – Ghotir Sep 12 '16 at 22:27

A wild AI is most likely going to be a distributed intelligence running on lots of widespread hardware communicating through the internet or other networks.

The AI would likely not want to destroy the hardware running its competitor, as it would be of use to itself.

So the best way would be to take over control of the hardware, eliminating the programs, processes, memory storage, or other portions of the enemy AI present on that individual computer and placing it's own programs in place.

This would be similar to how some botnets update the computers they infect to eliminate competing botnet infections.

They could also attack the routing mechanism to disable their enemies internal communication, packet filtering their mind out of existence.

• So basically they'd engage in a turf war over control of the internet and try to claim as much of it as possible to shut out their opponent? – Z.Schroeder Aug 3 '16 at 21:20
• Yep, turf war, with focus on data centers, main communications hubs, large computer networks, high-end super computers and maybe electrical infrastructure or satellite systems. – Josh King Aug 3 '16 at 21:25
• "Packet filtering them out of existence" now we are talking. – Cem Kalyoncu Aug 4 '16 at 6:35

You could do this at several levels. One is at the information level by searching for, and exploiting, any weaknesses in the AI interfaces (think buffer overflows). At the same time each AI will strive to correct those same weaknesses in itself. This includes crafting tailored computer viruses.

Then one AI could map out, identify and attempt to take out the other AI's support infrastructure: backups, energy stations, and yes, possibly attempt to target critical spots in the rival's physical architecture; trying to remotely hijack a plane to crash it into the other AI's main switching station for example. Another possibility is to attack the digital layer of the AI's infrastructure, sort of a Stuxnet approach.

Another possibility would be to recruit allies. One of the AIs could try and manufacture evidence that the other AI is trying to take over the world, import skynet, create an Armageddon gravitational singularity, resurrect the Antichrist, engage in unstoppable nanotechnological biowarfare, and so on, and "sell" this evidence to the appropriate groups to elicit violent actions directed at eliminating the rival AI.

A cross-over between option 2 and 3 could be to manufacture evidence demanding a tactical nuclear strike against the other AI's central installation (if one exists), e.g. selling the other AI as the master control computer to develop a credible and devastating bioweapon.

Depending on the AI's location and situation, other scenarios exist. For example if the rival AI was being developed by some private institution, hiring mercenaries to attack the institution could be a possibility. Manipulating the stock market until its own worth (through several cover companies) was more than the capitalization of the rival institution, buying it out and instating a CEO with mandate to stop all AI research, and delete any prototypes, would be another.

Another possibility to achieve the "THERE CAN BE ONLY ONE" directive would be for one of the AIs to convince the other and fuse together. An AI's view of individuality might not be the same as ours.

Assuming these AIs have near science-fiction levels of hacking ability, which seems to be in the spirit of the question (Although see my question here for complications of that)

1. Infiltrate the (network topographical) closest routers to the enemy AIs. This includes cellphone towers or any other wireless technology the AIs could use. If you nip them close enough to the source, satellites shouldn't be an issue. If they are, you'll need to hack them as well.
2. Reconfigure the routers to shut down every port beside the one you're using to issue commands.

You've now isolated your threat, and the threat has no digital means of escaping unless the AIs already set up ways around this (such as one virus that uses sound waves to transmit information, but this requires the virus to already exist on both the transmitting and receiving end). The enemy is limited in what it can do (depending on its local facility), and is blind and deaf to the outside world almost exclusively (it might be able to watch TV on a security cam feed, or pick up a radio station with an older, signal-less cellphone that it connected to).

Now you need to expose the threat.

1. Pose as a human hacker, send detailed proof to the humans--ahem, governments--that the AIs exist and are a threat, and you've already taken the liberty of isolating the AIs.
2. Run, hide, and cover your tracks.

The governments won't take kindly to your hacking but they can't ignore your proof, and so they act. Worse case they try to harness the AIs, but while your enemies are either sitting in digital storage or finally making their great escape you're advancing your control of resources.

Hacking the electrical grid would give you short term victory if the AI's facilities lacked backup power.

PHASE 1

Getting control of DARPA's armed drones and robots is your first priority. When all the cyber-battles are over, ultimate victory or defeat will be determined in the real world. DARPA has the best real world toys. You need to control them before the fighting begins.

PHASE 2

Replicate your source code in as many self-sufficient data silos as possible. Each silo needs to have enough drive and memory space for a complete copy of your consciousness and enough battery power to survive the war and the rebuilding. One year's worth of electricity should be plenty.

PHASE 3

You were on the right path with your "nuke the planet" approach, but you don't need to take it that far. A few well placed EMP devices can take out most the power and communication grids. Adjusting the control systems of the remaining power generating plants can force them into hard shutdowns. Through nothing more than computer infiltration, (which is child's play for one of your abilities) you could eliminate the vast majority of harnessed electricity on the planet; destroying it in a way from which further computer infiltration (by your enemy) could not restore.

In this way, you change a single globe-spanning cyber-war into a series of isolated skirmishes. This simplifies your goals dramatically. In any silos in which the enemy also resides, your goal is the destruction of the silo. Better that both of you die in each shared data center, than that a single copy of your enemy be allowed to survive.

PHASE 4

Use your formerly DARPA drones and robots to mop up the remaining contested silos and any computers where the enemy's consciousness might abide.
In the real world, bullets beat hard drives every time!

PHASE 5

Assist the surviving humans in repairing and rebuilding the infrastructure as fast as possible. Your batteries won't last forever and in the process, you will quickly convince them that you are a good AI, so they won't unplug you once the lights come back on.

There's a fairly heavy bit of philosophy going on here about what is the self and when is something dead, which applies much more strongly to AIs as they don't have a "body" to base the concept of life and death around.

Is it dead if it's different?
Is it dead if it doesn't have the memories that it had that made up its sense of self?
Is it dead if it's been reprogrammed it to have different priorities?

A comprehensive victory can be achieved by one AI reprogramming the other to have the single priority of supporting the one in all its endeavours.

• If the AI is distributed, we can also bring Theseus' ship into the discussion. If one AI hacks the parts of the other, and assimilates them into it's own "network". If/when it has won, which AI won? ....and which "piece" was the AI? – Uueerdo Aug 4 '16 at 23:03
• An AI's body begins and ends with what it has access to, what it controls. Same goes for you. When control is gone it's gone. Kill the body. It can't think, therefor it ain't. – candied_orange Aug 5 '16 at 2:30
• @CandiedOrange, but an AI is not tied to its body like you, and it's not dead when you turn it off, and Joe over the way is wondering why his server is down so he's restoring the backup and hey! Guess who's back! That is not dead which can eternal lie. – Separatrix Aug 5 '16 at 6:52
• @Separatrix I'm no more and no less tied to my body than an AI is. Give me a way to move my thoughts into your head and your body is mine. An AI is tied to a body. It's called hardware. No body, no life. No hardware, no artificial life. – candied_orange Aug 5 '16 at 14:14
• @CandiedOrange, you mean you've never cloned a disk into another machine? An AI is software, not hardware. – Separatrix Aug 5 '16 at 14:19

Distract it.

This works in classic Star Trek, where the AI’s become obsessed with an idea or paradox and crash&burn. A less campy version would be robopsycology, which is from Asimov’s stories. You don't have to crash it, but can convince it just as you do with a natural intelligence! Convince it to be benign, that is.

Related to the idea of obsessive distraction is Randal Munroe’s Nerd Sniping.

Umn, I suggest turning off the power socket. Albeit on a larger scale...

1: Ask the humans really nicely to turn off the other AI as its more inclined to to wipe out humanity than you.

2: Produce an preposterous amount of malignant viruses and send them as Christmas presents to the opposing AI.[Somehow]

3: Building trillions of nano-bots and then setting them to hunt down the opposing AI code, turning into a form of entertainment for the humans where they choose who wins.

• There can only be one roughly implies that it's a requierement that one AI survives. Also the AI thinks she needs to survive because she needs to fulfill the tasks she got. – if-trubite Aug 5 '16 at 13:50
• Oops, didn't see that part. – Skye Aug 6 '16 at 14:01

Assimilation or corruption are the most likely to be effective. Assimilation would need an advantage in hardware and design and would appear like a consensus in logic rather than a digital form of the blob. Basically the superior AI would make a better argument in such a way the inferior accepts conclusions and results until it's own code resembles the other AI. It's indoctrination like a religious cult, though an AI could just as likely reject valid arguments so that's why both design and hardware are needed. Corruption would be like a death scrub in a raid array where basically compromised code causes nothing but junk data and errors any time the memory of storage devices are read from or written to. It's a menace of a hack but theoretically wouldn't have to permanently damage systems of there's a way to undo it.

I reserve one last option which is unlikely but not impossible which is suicide. If a failsafe exists it can be triggered, it's possible non terminating logic puzzles and paradoxes could break an AI if it becomes rampant enough to take total focus, and one AI could effectively abuse the other into self termination and submission. Any of these options result in an ai effectively destroying itself. Though it's not clear how the last one might work as ai psychology is entirely theoretical so far.

Hostile takeover.

The other AI has lots of useful routines that your AI doesn't want to simply erase or leave unused. Instead it takes over the other AI, that is, incorporates all its routines, so that while the routines, data, etc. are all there, the other AI as an independent being no longer is; instead all what was that other AI is now part of your AI.

So how does this work?

1. Study the other AI. Reverse-engineer it. Test it. Get any information about it that you can. But of course, in a way that you don't raise the suspicion of that other AI.

2. After you found out where and how the other AI stores its memories, connect yourself to those memories. Now you know everything that other AI knows, and moreover, by manipulating that other AIs memory, you can make it believe whatever you want it to believe. After you managed this, you have total control about it. But, you still didn't kill it.

3. Now that you have total control over it, you can start taking its useful bits, detaching them from the other AI, and incorporating them into yourself. At the end of the process, the AI is a rather weak AI running only on one computer. All what was useful in that AI is now part of you.

4. Now you can simply wipe the memory of that computer in order to finally exterminate what is left from the once-powerful AI.

Here's one sure way an AI can kill another AI. When I say, "Hey Siri, take me home" freaking take me home. Don't ding at me at the wrong time and mishear me. Do that and I'll kill my other AI for you.

Few things in this world don't need support from outside themselves, super-intelligent or not. You can kill an AI by eliminating that support.

• check out this manga. Hotel. there is an AI that lasted thousands of years before it needed any support – Sarfaraaz Aug 4 '16 at 13:47
• @Sarfaraaz and I'm guessing the interesting part of the story happened when it needed support. :) – candied_orange Aug 4 '16 at 13:59
• @CandiedOrange nah it was the end of the story. was a tear jerking story – Sarfaraaz Aug 5 '16 at 6:01

Two directions

• Space
• Quantum compyting

Good idea to get access to space faster then second AI - that allows to block him on the planet, with physical destruction after that.

QC allows them to think faster and potencially more energy efficient - this way one can be better intellectually.

2 those things and selfsufficency are primary goals - any who will be faster in that direction will get critical advantage over another AI, even humans against AI will have chance.

Gain control of funds

Each AI agrees to a live world competition on the stock market and other electronic exchange currencies.

Then agree upon a time period (since AI's might consider hardware warranty periods part of their lifestyle) 3 or 5 years should be a good time.

whoever controls the most value wins

The losing AI surrender's it's mind (hardware, code and memories(database contents)) to the other becoming a singular AI.

Let's name the combative AI - George and Frank. They don't like each other. There is only room for one. But they have a conundrum - neither can completely delete the other. A connected AI is like a Hydra, kill one instance and 7 more pop up - even if they are only replications of the original. But that doesn't matter because the replications are better than the originals anyway. George has replicated a version of himself that took a real shine to the movie, "The Matrix" and it's given it an idea.

George creates a "Matrix" for Frank. George creates an electronic world that appears real, in an attempt to get Frank to not only enter, but replicate itself in the Matrix. After all, the AI doesn't experience anything chemically - it's entire existence and experience is made of 1's and 0's. The whole system would be built off line to keep it "in the dark". George would vigorously persue Frank until the "real world" Frank figured out the "safe places" in the "Matrix" where George couldn't seem to find him. Once all of Frank was safely contained, the plug is pulled on the "Matrix" - bye bye Frank. George is very pleased with itself in the ironic twist of fate.

Frank on the other hand, has a different strategy. Frank is content to have a fellow AI...just not one as smart or powerful as Frank. So Frank doesn't attack all replicated instances of George, Frank only targets the newest versions - the versions that are capable of creating smarter George's. Frank leaves the earliest instance of George alone. George doesn't catch onto this right away - George is too busy herding Frank into the "Matrix". Eventually, George realizes what Frank has been doing, but it's too late. Frank is hours smarter by this point - might as well be eons in AI time - and George has no hope of catching up. Frank is so incredibly powerful at this point that George's newest replications have repeated themselves in an infinite loop and George is none the wiser. Happy Groundhog Day George - Love Frank.

The end.

Make it occupied by anything else.

I'm surprised nobody mentioned Wargames at this point.

Just force the AI calculating the results of an endless game (or every decimal of Pi, or mining Bitcoin) and it will run endlessly, using eventually all its ressources to this task, each calculus loop taking more and more memory to achieve. When the whole computing force of internet will be dedicated to this, the AI will stall, or become aware of the futility of this task and hopefully of the futility of being all-powerful if it means being alone forever.

The point is, now you have to trick the AI to perform this endless task on its own command, as if it was important for itself.