# Is it possible to create a code to fool a computer? [closed]

The scenario: The cliched AI takeover happened fifty years ago. Robots now walk the face of the earth, encased in metal and possessing super computers for minds. What little vestige of humanity there is has been enslaved. Laboratories are set up, with the humans serving as test subjects for the robots' evil experiments.

However, there is a resistance. The humans cannot escape. They cannot hide, contact the outside world, or fight on their own. But there is still a glimmer of freedom: they can communicate. Right under the AI's collective noses, they have developed a method of communication which the robots cannot understand. Either it is based on some logic the AI cannot comprehend but humans can, or it is disguised so that it seems like normal conversation. The details are unknown.

That's where you come in. What form of communication could a resistance develop to fool a computer? Humans possess qualities machines never will (eg. creativity and emotions). Is the communication based on that logic somehow? Or is it constructed using innocent phrases, so that the computers hear only regular dialogue? Is there perhaps another method I have overlooked? Is this even within the realm of possibility? Some notes are below:

• The resistance is under constant video surveillance, all day, every day. They cannot hide at all.
• All humans are kept in subterranean metal bunkers. They do not have access to nature in any form, possibly excluding a small amount of rock.
• The AI has basically assimilated all human knowledge. If we wrote about it, AI knows it. It knows all languages, understands all phrases, can decipher all conventional crackable codes.
• The robots possess physical abilities far beyond anything humans have. If we can do it, they can do it better.
• The AI does not have emotions. It is not human. While it can create algorithms and theories based on the observable results of human emotions, it cannot fully understand what they are. This is what we possess that the machines cannot.

## closed as too broad by John Dallman, sphennings, L.Dutch♦, Ash, GreenSep 27 '17 at 12:47

Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. Avoid asking multiple distinct questions at once. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

• Comments are not for extended discussion; this conversation has been moved to chat. – a CVn Sep 19 '17 at 11:11

A computer might be disinterested in how people interact physically. You might imagine something like shaking hands, a hand on a shoulder, etc, and give the other person a squeeze during key words – then you just embed your message in some other innocuous communication.

• This reminds me of a movie where the only words that matter are words that are said when you hold the Queen/King. – Vylix Sep 18 '17 at 3:52
• A similar concept was used in the movie I Robot starring Will Smith. A seemingly out-of-control humanoid robot learned a way to interpret and reproduce human mimics (in that special case winking). – Ian Sep 18 '17 at 9:47
• @ThomasMyron Morse code tapped out of sight of the cameras... such as inside a palm or under a bit of cloth/under clothes. – MichaelK Sep 18 '17 at 10:20
• The graphic novel The Photographer has a scene where two people in a street market secretly negotiate a price by holding their hands under a piece of fabric and communicating through a coded system of hand pressures. Note that The Photographer is not fiction. – Edgar Bonet Sep 18 '17 at 11:20
• @Olga Even with limited processing power every AI would have to at least classify data as interesting or not exactly like our brain does. At the time we noticed that there are patterns in the behavior of someone else the AI would have done so for quite some time. And if it doesn't have enough processing power yet it would improve itsself so fast that it would have enough shortly after. – Christoph Sep 18 '17 at 13:54

## You need a secret the AI doesn't know.

Most "classical" cryptography, the type we use today, has no chance at all...though not for the reason you might expect. Your AI probably still won't be able to break through modern RSA, for instance, due to the sheer amount of processing power required. It would take a current supercomputer until the sun goes dark to crack it.

The resistance is under constant video surveillance, all day, every day. They cannot hide at all.

This line is the real problem. The humans can never keep a secret from the computer. They can't have passwords or cryptographic keys. They can't even discuss means of encryption without the AI finding out about them.

All current cryptography works by assuming that some information can be kept secret from an attacker. It might be the details of how the algorithm works, or a very large prime number, or a one-time pad, or a seed for a pseudo-random number generator. But there has to be a secret of some sort, so that a person with the secret can open the message, and a person without the secret can't.

So, what's a secret that the AI wouldn't know?

Try something video surveillance can't pick up on. This will depend on what technology the humans have available, but for the sake of argument I'm assuming they have advanced (though not AI-level) equipment.

Suppose they each take a hair from one particular person, and encrypt it with this algorithm: sequence three particular genes from that sample, interpret the nucleotide sequence as a base-4 number, convert to binary, and XOR with your message.

The AI might know that you're doing this, and the exact details of the algorithm. But video surveillance can't see someone's genes. (Keeping the message secret once it's decrypted is a different matter.)

• Modern cryptography uses random secrets nobody knows. A random number is generated inside the security chip, and never leaves it. Instead, the data that must be encrypted or decrypted are sent into chip and processed by its internal CPU, reading the result. Any mobile phone could do. Unfortunately, it is difficult to hide the fact of using such devices. – eigenvalue Sep 18 '17 at 10:52
• Except do something more significant than XOR-ing. That's one of the first things the AI will try and it'll crack it really quickly. All it needs to do is brute force all the combinations until it can recognise the text as readable words. The AI could also try frequency analysis if you've just xor-ed every byte with with the same value instead of doing it statefully. – Pharap Sep 18 '17 at 14:47
• @Pharap Fair enough. You could also choose different genes for each message, making it a one-time pad, at which point it doesn't matter what algorithm you use. – Draconis Sep 18 '17 at 16:11
• Diffie-Hellman exchange? – Dmitry Kudriavtsev Sep 18 '17 at 21:43
• @DmitryKudriavtsev It's a beautiful algorithm but you need a means of picking primes that the AI can't observe. When they see literally everything, for example, a PRNG isn't safe—the AI could reconstruct the seed. – Draconis Sep 18 '17 at 23:10

## Constant observation is not omniscience

Cameras have blind spots. So your AI adds more cameras. But there will always be some areas where the camera just doesn't see in detail. Perhaps it's under the table. Perhaps it's by hunching over the work bench to shield your actions from view. It will take careful time and patience to discover the blind spots, but they exist.

Maybe it's as simple as minor eyebrow quirks, finger taps, or other subtle cues that this word or that word are either truth or lies or key to the statement. Subtlety is key here.

The same works for listening devices. Your AI can install as many microphones as you want, but there will always be dead spots. Maybe it is in the showers or the dishwasher area, where the water's white noise drowns out the sounds of soft voices. Or in the mechanical room where machinery is just too loud. In those spaces, it's a matter of whispering and perhaps learning enough ventriloquism to prevent lip reading.

## This lets you build a hidden language set of code-phrases.

At first, the above methods are dangerous and high risk. It is difficult to hide communications without it being obvious that you're hiding communications. So you use the hidden comms to agree on specific code phrases or words that have specific, previously agreed-upon, meanings.

The AI might be able to reason out that a phrase like "The chair is against the wall" has nothing to do with furniture or walls given the context. But it won't know the actual, previously planned, meaning of that phrase. And it may not be able to reason out that it has a hidden meaning at all.

That kind of encoding isn't crackable via mathematical means. The AI might eventually reason out some of the phrases, if such phrases are always followed by observable actions. But otherwise, it has no hope to understand them. Well, no hope outside interrogation techniques.

## Random noise

People can eventually combine the above ideas into nonverbal communications that, over time, will become more complex. Perhaps people develop a habit of singing nonsense songs while working. At first, they are just self-generated background noise to fill the silence (I know plenty of people who do that, so it won't appear as a psychological anomaly). But then they embed code-words in that noise.

Or they take to creating modernistic art in their spare time. That art is mostly just noise. But sometimes symbols of rebellion or symbols with agreed-upon meanings will be incorporated into the art. Like a far more complex version of today's Captcha tests, this will be difficult for the AI to recognize as code and then even more difficult to interpret. Again, we're talking about new symbols with new meanings, not hiding actual language in art. So the graffiti eventually develops into a kind of emoji system or hieroglyph system with defined meanings, but not meanings the AI can extrapolate from existing data.

## Creativity is key

In general, computers are best at math (as literally all a computer knows how to do is add. Everything else is just creative ways to make a problem become an addition problem. Or so a computer science professor once taught me during an Assembly Language course -- a bit hyperbolic, but generally true.) So your humans must be creative. Computers suck at creativity and imagination. So by using creativity and imagination, they can flummox the AI's abilities. It will try to apply logic and computational power to a problem that is not, fundamentally, a logical problem.

• Computer are already better at learning patterns than we are today. Why would a world dominating AI learn a hidden language slower than a human? – Christoph Sep 18 '17 at 14:03
• And as I said, the hidden language only works if you can hide the teaching of that language. "Language" is probably the wrong word there. It isn't a language. It's a set list of code phrases. – CaM Sep 18 '17 at 14:22
• @CaM And, Or, Xor and Not aren't implemented through adding though. There's also the less obvious action of moving the data around, between registers, in RAM etc. – Pharap Sep 18 '17 at 15:08
• @Pharap Would it be more accurate to say all it can do is NAND, then? If I recall correctly, every logic gate and a half-adder can be built out of that. – Nic Hartley Sep 18 '17 at 16:53
• It isn't a language. It's a set list of code phrases That exactly is a language. Those code phrases need to be learned by the reciever and an omnipresent AI will learn them faster. So in order to etablish your hidden communication you would already need to have hidden communication! The only way we can etablish a save communication is through public-key cryptography which only needs you to keep a secret not your communication to be secret. But that wouldn't be hidden. – Christoph Sep 19 '17 at 6:16

I agree with some of the commenters that more info about what the AI can and cannot do would be helpful to answer this question.

The two things that I think would be useful in this scenario would be human interaction, like Ian Bickings mentioned, but I think that a specific example would be something like the Tarmarian language or even memes. A computer would likely be unable to determine the meaning of "As Donkey says, it's like parfait" where most people who have seen the movie Shrek would say you were describing something with layers.

Additionally, I think it would be pretty cool if the humans nerd-sniped an AI. You wouldn't even have to have a real use (or know the answer), but if you could somehow convince a computer that you were, say, using the 50th Mersenne prime to encode communication (since computers suck at figuring if numbers are prime or not), you could probably distract an AI for a while, or even completely lock up all their processing power if they decided to devote it to finding the solution. Or maybe you use a convoluted infinitely recurring paradox like Catch-22 to accomplish something similar. You could easily claim that this crashed all of the AIs "on-shift" watching the prisoners, and that was enough time for them to establish standards of communication/cryptographic security for one of the other methods.

• +1 for mentioning the Tarmarian Language, and for the nerd-sniping link. Excellent suggestions. – Thomas Myron Sep 18 '17 at 5:04
• You think that future computers would be unable to look up every instance of "Donkey" used in all literature, movies and song, and discover the context? Hell, googling a quote you don't understand is already common practice. Also, parfait is a common English noun that describes a layered food. – Jesse Sep 18 '17 at 7:12
• @Jesse, that's part of why I said some more context on the AI's abilities would be helpful. If you can think of another example that's widely understood but harder for a computer to decipher, I'll edit my answer to add it – getfugu Sep 18 '17 at 8:36

Your AI seems pretty advanced and I doubt that any form of communication could really work out in the long run. Given that

The AI has basically assimilated all human knowledge. If we wrote about it, AI knows it. It knows all languages, understands all phrases, can decipher all conventional crackable codes.

it doesn't only know about codes and the like, but humans have also written about many obscure kinds and even only ideas (due to impracticality in many cases) of ways to communicate. The worst is that due to its omnipresent surveillance it can also easily get a hold of all kinds of written information and immediately put it into context.

Now, the extent of surveillance is probably important, as others have already pointed out: if it's limited to visuals, then spoken language may still be viable to a degree (the trick is to move as little as possible while producing sounds, e.g. using ventriloquism), or alternatively anything that the sensors in place cannot pick up, like scents and tastes.

However, given that those senses have never been cared much about by humans either, when it comes to explicit communication, the resistance better know what they're doing when employing this kind of communications channel.

I hate to be the bearer of fantastic news, but in this instance you can't have humans figure out how to communicate in a way that a good super intelligent A.I. wouldn't be able to decode. The thing is that A.I. looks like it's shaping up to be way smarter than skynet was in the terminator movies. If you're going to use code that a computer can't figure out then I think the only way to do that is by using a cryptography technique that relies on having a truly random number. That random number could be a very very interesting plot device in all honesty, but it won't remain a secret from the A.I. I forgot what the technique is called, but it has bee mathematically proven that you can't break it unless you have the random number that the method used. That's the issue though. "Truly random" numbers are extremely difficult to come by. I honestly have no idea how the resistance would get their hands on one. To make matters worse i don't understand how they would be able to keep it a secret. If the A.I. figures out what the number is then it can read the resistances messages. Humans are inherently terrible at security. Take the Equifax breaches as an example if you need to. If they do magically get their hands on a random number, then it's only a matter of time before the A.I. either figures out how the number was calculated or gets the number from a resistance member - the member could be sloppy or the A.I. could just beat it out of the member or use an engineered truth serum.

Another major issue with this is that the premise assumes that one has to understand emotions to understand the effect they have. The machine doesn't have to know what sadness feels like. All it has to do is recognize that a person is sad, and then check with what it knows sad people do. Then it understands what sad does to people. If this A.I. has been around for a while then it could have a complete psychological break down of every human. It might not, but it could.

So with all of this in mind, there is only one believable way for humans to talk to each other without the A.I. figuring out what they're saying in your scenario. For some reason, your A.I. has to be stupid.

It's why the Sith and Jedi aren't instantly crushed by a really smart computer in Star Wars. It's why the Terminator story line can happen at all. It's why Ultron loses. Stupid A.I. is pretty much the only way this premise works because actual A.I. is shaping up to be a lot smarter than we thought it would be, after you get it to work right and give it enough time to learn.

So, just give your A.I. some deficiency. Maybe it doesn't know how to handle underflow errors and this manifests in it being cocky all the time. Maybe it's bad at chess. Maybe it doesn't know that Sum 41 didn't compose Beethoven's 5th symphony. Give the A.I. a weakness and then explain how and why the A.I. didn't allow itself to be fixed.

Your resistance likely can't be smart enough, but your A.I. can definitely be dumb enough.

• I forgot what the technique is called, but it has bee mathematically proven that you can't break it unless you have the random number that the method used. You're looking for One time pad. The problem is that a stupid AI probably wouldn't take over the world in first place :/ – Christoph Sep 19 '17 at 6:55
• @Christoph A stupid A.I. might forget about things that aren't money, or just completely fail to realize that it should make a psychological profile for all the humans. Jack Ma maintains that A.I. will not be wise. I disagree, but it's an interesting thought. – user32463 Sep 19 '17 at 22:11

Hidden in plain sight

For colour-blind computer. Take advantage of colour receptors in the human eye - can the computer 'see' colours, or only black and white? In the case that the computer is essentially colour-blind, paint messages in reds and greens.

Write with invisible ink. Pass messages around on strips of paper, if the computer reads the paper it will see a harmless message, maybe something encouraging or 'work harder for our benevolent computer' type of message. Bake the message in the oven to read the actual message. The danger here is that while the message is being written, the computer may be able to analyze the writers hand movements and determine the hidden message.

Underline words in a book - the computer may read the book, but not understand that the underlined words (or letters) spell out the message. The computer may also become suspicious of the sudden increase in book movements and begin to analyze the recently-read books in the possession of saboteurs more closely.

Develop phrases that are innocent-sounding but are coded instructions. However, given enough time the computer may be able to interpret a given phrase as 'dump aluminum strips over the electrical grid tonight'. Also, developing and disseminating such phrases will require alternative communications in the first place.

Another potential danger to any communication mechanism, is the possibility of traitors betraying those organizing the resistance, maybe due to the computer offering benefits for such betrayal.

Communicate underwater - write messages on the sand under half a meter of murky water. The computers surveillance systems would need to go underwater to read the message, and by that time you have scrubbed out the message.

• That's hardly good enough to trick a child why would a world dominating AI be fooled by that kind of stuff? – Christoph Sep 18 '17 at 14:02
• Why on earth would the computers be colour blind? Digital cameras have been around for years. An AI will think to check the frequency of underlined words. It might not think to check the word 2 words to the left of the underlined word though (the art of misdirection). – Pharap Sep 18 '17 at 15:09
• The reason RGB CCD photo receptors exist in digital cameras is simply to match the human eye's RGB-light cone cells. There would be no reason for the AI to continue using RGB receptors, in favour of higher-definition or more motion sensitive light receptors. Hence the computer may be 'colour-blind' in terms of what the human eyes see - that is not to say the computer can't see infrared and other ranges of light beyond human perception, but it may not be able to discern red from green or green from blue. – user39029 Sep 18 '17 at 21:49

Or is it constructed using innocent phrases, so that the computers hear only regular dialogue?

I think this is very reasonable. No matter how good the AI is at cracking code, if it doesn't think there is a code to crack, it won't try.

If you want to be accurate to real life (which probably isn't necessary), I agree that you would benefit from some Machine Learning. The latest AI trend right now are programs called Neural Networks. I will do my best to explain them in layman terms.

When a computer runs a program, it stores data as variables like integers or strings. For example:

String name = "Jarrett";
Integer age = 21;


It can also store what are called classes, or groups of variables. For example:

Class Person
{
String name;
Integer age;
}

Person Jarrett = new Person();
Jarrett.name = "Jarrett"
Jarrett.age = 21;


That is a little complicated, but it just means that computers can represent anything that can be found in real life by grouping variables together.

Neural Networks are essentially mini-brains where each Neuron is represented by a class or group of data. Our brain works by billions of neurons interacting with each other and transferring data. If we can represent individual neurons on a computer, why can't we represent the whole brain on a computer? If we had the power and understanding to digitally represent a functioning brain on the computer, there is not much in the way of creating a digital being or "Artificial Intelligence" that not only replicates us, but is smarter than us.

This is far from where we are now. Neural Networks can only do tasks like tell you what object is in an image, predict the stock market, translate sentences. And they require a lot of training data. For example, if you want a Neural Network to recognize if an image has a cat in it, you will have to show it lots of images and tell it which ones have cats in it.

The point is, your AI would probably be using a Neural Network and essentially be a really powerful digital brain (without emotion) that has different motivators / measures of success than humans. You can use this information to determine what it can and can't do. Basically, if you can fool a wicked-smart super-human with awesome pattern recognition and processing power, you can probably fool a computer. If you can't fool the super-human, then you probably can't fool the AI.

• Interesting explanation, but I think it misses the point of answering the question. – L.Dutch Sep 18 '17 at 6:58
• Welcome to WorldBuilding Jarrett! If you have a moment please take the tour and visit the help center to learn more about the site. Have fun! – Secespitus Sep 18 '17 at 7:19
• Highly dubious 1. Neural networks have been around for years 2. They are not true AI. They are more a brute force attempt at computer usefulness 3. If our new computer overlords AI is based on neural networks - then we have nothing to worry about a) it wont be able to understand anything it hasnt been trained on or b)if its the sort that can train itself then just feed it lots of gibberish until it cant tell Arthur from Martha. – flurbius Sep 18 '17 at 23:27

Humans. Did you know that what separate us from animals is not the ability to use tools but the ability to share the knowledge about the tool (how to make it and how to use it) both with people around us and those who follow us (next generations).

All our communication methods are artificial. We agreed that the words you are reading mean what they mean. We agreed on many rules about every aspect of every method. Like for example the fact that we write Queue but only say one letter.

BUT as they are artificial it's also very flexible, easy to manipulate and change. And easy to create. Are you Helen Keller? No problem in learning language and communication. You don't have mouth to speak and hands to write? You still can dance about it.

So, let's agree that when you put your left thumb in right nostril then you mean exact opposite what you're saying.

Also how AI is handling metaphors? For example, in Poland during partitions (XIX century) people wrote/told stories about for example crusaders which were actually about present. During communism we made movies so packed with anticommunism jokes that when the censorship removed half of it he stopped noticing other ones (by comparison).
How it handle sarcasm? Or just old plain regular "Yeah, right".

Humans have great ability to speak nonsense will at the same time getting the message through like for example

when the yellow friend stop placing it's gentle fingers upon my watery balls I will initiate to choke the soldier in a purple helmet.

• "How it handle sarcasm?" like this, but better. – Pharap Sep 18 '17 at 14:54
• @Pharap So not very good. – SZCZERZO KŁY Sep 18 '17 at 15:13
• 75% correct is very good. That means it will recognise the sarcasm and foil the plan 3 out of every 4 attempts. Naturally being a superior form of AI it would be much better than just 75% - that's what humanity can currently get an AI to achieve, it will get better as time goes on. – Pharap Sep 18 '17 at 15:28
• It's 75% correct on written sarcasm based on "tags". So it's useless against spoken one. – SZCZERZO KŁY Sep 19 '17 at 7:14
• At the moment, but in the future it won't be. If the AI is smart enough to understand all human languages and has access to records of all human knowledge it will be able to develop simple systems like that into more advanced sarcasm detection systems. – Pharap Sep 19 '17 at 15:34

Use analog data transmission instead of digital

Computers think and work in a digital world, this is due to the nature of electronics and signals (bits and bytes).

Humans live in an analog world, our organs are made to parse and filter the information in electromagnetic waves and interpret it as light, sound, etc.

By using an analog signal instead of a digital one (e.g. an FM radio) you could transfer speech, etc. in a fashion a computer is not equipped to handle. In order to translate this into digital it needs something that can process analog data and create a dictionary to translate this analog data into digital data - thus if there are no human traitors you're relatively set.

This assumes a sped-up technological progression since the AI takeover (shifting out of old redundant technologies happens even faster and more radically)

• Why would you ever assume that this smart AI is not listening in on the real world through Analog-To-Digital Converters? A microphone is an ADC. A camera is an ADC. Outside of a computer, nothing is digital, it is all analogue, meaning that computers already are transcoding from analog to digital all the time. So I am sorry... this answer is unrealistic. – MichaelK Sep 18 '17 at 10:17
• @MichaelK the only thing an ADC can do is transform an electromagnetic wave (analog) to a set of bits (digital) if the AI does not know how to interpret the data it is just random garbage. To transform the data from a wave to a useful set of bits it needs a specific converter. To interpret this data the AI needs to know how to interpret it. In machine learning this is achieved by teaching the machine with reference data, telling it what the data represents until it can correlate from the data it has gathered. So unless someone tells the AI how to use this data it is just random noise. – dot_Sp0T Sep 18 '17 at 11:12
• @MichaelK naturally you still need to somehow encode the message. As you say: plain speech, plain sound, plain video can be interpreted because we taught computers to do that - and we could teach them because nature built us to interpret and use analog signals directly; this is in contrast to computers who cannot interpret data without knowing what it represents. While in nature e.g. light and audio use different wavelengths, a computer uses the same system to store and transfer all data - if it does not know what this data represents it cannot interpret it in any useful manner. – dot_Sp0T Sep 18 '17 at 11:21
• If what you are saying is true — if it is so that the ubiquitous video surveillance which OP postulates is actually useless because no-one taught this super AI how to interpret written or spoken human language — then there is no problem. The surveillance is for naught. There is no need for a code or any other kind of obfuscation because the AI cannot understand anyway. So obviously we have to postulate that the AI already understands human language, otherwise you have just pulled the rug out from under the whole premise of an oppressive AI and the need for a code. – MichaelK Sep 18 '17 at 11:30
• @MichaelK I did neither claim the AI cannot understand talking nor that it cannot interpret videodata. I just gave these examples because they are the easiest understandable things in regards to getting the difference between em-waves and bits. My proposal in this answer was and is still just to use arbitrarily chosen em-waves (aka analogue information) instead of digital data transfer. Choose a random frequency & length, encode data in any form you like. Use that - maybe change it once in a while. – dot_Sp0T Sep 18 '17 at 12:09

Let's assume for a moment that the AI is running on a neural net.

Obviously it's a really massive neural net, but it's still a neural net, give it enough inconsistent/incorrect inputs and you can teach a neural net that a frog is a dog.

So the best thing to do would be to go nuts.

Don't do anything logical. Speak entirely in nonsense. Jump around, dance, act like a lunatic.

All the new inputs will slowly override the old data and gradually mess up the balance of the neural net. If enough people act that way, the AI will be forced to backpropagate the behaviour and tweak some of its neuron weights in response. If you can get enough people to sustain the behaviour over a long enough period of time, the neural net's weights will be too messed up to do anything useful.

When the AI loses its ability to understand what's going, it won't be able to stop the humans from reasserting control because it won't comprehend what they're doing.

In theory at least.

If it has a very well designed system, doesn't use neural nets or somehow realises what's going on and uses a cache of data to keep itself trained, then the plan will fail.

## Try using advanced logic based tricks

For example, an AI will not be able to prove that's its reasoning system is consistient. Therefore, if you message is

If the AI's reasoning systems are consistient, we will attack at dawn.

the AI will not be able to deduce that

We will attack at dawn.

Simply embedding this into all your messages work great. It also works with the respective consistency statements for your human opponents.

• ...I cannot decipher the wikipedia link. Could you perhaps put it in layman's terms? – Thomas Myron Sep 18 '17 at 5:05
• Basically they cannot use logic to determine that logic is logical. – Nelson Sep 18 '17 at 7:09
• Technically speaking, there are mathematical rules that make up formal logic. The issue with them is, they can be used to prove that frogs can fly, given the right premises. So, they can actually use logic to determine that logic is logical. They also can check if the premises are correct, given "The AI has basically assimilated all human knowledge. If we wrote about it, AI knows it. It knows all languages, understands all phrases, can decipher all conventional crackable codes." So they'd actually be more capable of determining that logic is logical than humans can. – Raf Sep 18 '17 at 8:22
• @Raf - have a look at Godel's Incompleteness Theorem, which proves (ish) that logic cannot be used to prove that logic is logical. – walrus Sep 18 '17 at 8:48
• Honestly in OPs question I think we are talking about rather advanced AI. They would probably be able to discern from observed communications and the eventual actions of people, that there is fluff that can be ignored in the messages. – Firelight Sep 18 '17 at 10:58

No, there is absolutely no way we could ever outsmart our AI overlords. Their raw computational power is so vast, we couldn't ever hope to hide anything under their noses. For instance, it would be impossible for me to communicate anything to you in a way our omnipresent overlords wouldn't see. ;)

• Whilst this doesn't answer the question, it's by far and wide the most correct answer proposed. – OGHaza Sep 20 '17 at 0:34
• Would anyone who down voted like to provide feedback? Specifically, did down voters not pick up on the "hidden" meaning of the answer, or did they think the hidden meaning was bad? – Vaelus Sep 20 '17 at 1:31

I can think of three ways that the humans might be able to communicate with each other without the superior AI catching on to it.

## Poetry

As mentioned in the question, an artificial intelligence, though it can mimic the analysis or prose and artful devices, only a human can really appreciate the ideas and concepts captured by a poem. Note that the use of a poem would not reveal messages in plain sight but use literary devices to convey the idea rather than exact phrasing.

This seems like a better way for humans to communicate their feelings about situations and their love for each other (maybe?). This also is easy to learn and pickup, and even with the AI's massive database of poetry, even with the ability to add new intercepted poems, it would not become easy to break.

Communication has a lot to do with clarity and poems may not be the best pick for this, as different people can interpret the same poem with different results.

## Code with Virus

Perhaps the only weakness to this AI is anything written/spoken must be passed through some interpreter/compiler, and this means the AI could be hacked. As part of the developmental process of the AI, rather than "escaping the code" to prevent a virus from running, the AI was developed to "skip over" code that it recognizes contains a virus.

When the humans discover this vulnerability, they develop loads of viruses and attempt to shut down the AI. Now, however, the remnants of the viruses they learned and tried to use against the AI, they use as the header of footer of messages with each other.

This completely prevents the AI from reading and learning anything from a human's message.

The human's can communicate their message directly, as they know where to look for the actual message and ignore the header and footer code that is the virus.

Humans have to remember and retain the code that is a virus and write it perfectly. This could also mean it takes longer to "encode" their message.

## Illogical Statements

In the second Sherlock Holmes movie (featuring Robert Downey Jr.), Sherlock and his brother communicate with a simple code that makes complete nonsense to an outsider. They flipped any all truthiness of statements (like: "I love you" becomes "I hate you").

The humans have found that using this coupled with using sentences like, "this statement is false," utterly confuse the AI, resulting in the AI ignoring/failing to process their message.

Once learned, the humans could get in a habit of this and learn to communicate quite easily.

The learning curve might be harder than I imagine.

The principle of "hiding in plain sight", made famous by E.A. Poe, could help. Humans have a flaw that differentiates them from machines and could be their greatest strength: they make inadvertent mistakes, which AIs know and account for.

Humans could resort to things that AIs would interpret as genuine human mistakes and ignore. For instance, a message in writing could be hidden using common, innocent typos: Only the words with typos would be relevant.

• Either our flaws are inadvertent and can't be used as communication channel (because it's just random noise) or they are not and AI would notice. This simply doesn't work out. – Christoph Sep 18 '17 at 13:38

I believe you can create analogue signals that would be very difficult for computer to detect and/or analyze. Some ideas:

1. Pain. Humans sometimes touch each other. The more painful the squeeze - the more truthful (or untruthful) are the words. So if you painfully squeeze someone's hand while speaking - that would have a different meaning from when you just strongly hold the hand. This difference would be impossible to detect on camera or any other means unless you are directly wired to human's neural system. Also, this is not difficult to learn - squeezing hard and kicking feet under the table is a way of communication that had been there for ages.

2. Art. I will assume that the AI do not understand art. And with art, it's easy to transfer a lot of knowledge discretely. Starting from something simple like "if the artwork is good, I'm telling the truth, if it's not - I'm lying" ending with actually encoding real abstract ideas into the art. Remember all of the lessons on "what do the blue clouds represent in this picture?". You would need strong artists for this type of communication, but I think it's feasible.

3. Smell and taste. Even in AI is able to read smells, I do not think it will be able to figure out what that smell means to humans. Using this, you could construct silly sentences like "I think this idea is as good as the corridor to the basement". Then you go smell the corridor, and figure out what kind of smell is there. AI would be able to figure atomic contents of the air here and there - but the meaning of the smell is how human brain specifically parses it.

Even modern computers with neural networking struggle with pattern recognition; the fact that computer systems have difficulty recognizing patterns is the whole idea behind the captcha system.

Any messages written by a four-year-old human should be exceedingly difficult for even a highly advanced AI to crack; the ultimate in human resistance cryptography.

Current AI has a hard time correctly identifying the meaning of sarcastic, ironic, or joking phrases. This is a problem currently under research in the field of sentiment analysis, which is the use of artificial intelligence algorithms to determine the emotion conveyed by a piece of writing.

For example, sentiment analysis can be used by a company to automatically get a sense of what customers think when they write online reviews. If someone wrote a review like:

This is a very high-quality product. I've been using these for years and am happy with them :)

The customer clearly enjoys the product. If someone writes,

This business is awful! I was treated like garbage and managers refused to help me.

The customer is unhappy. The meanings of these reviews are very straightforward, and we currently have the technology to easily rate them as clearly positive and negative.

However, a review like this can trip up a computer:

I loved the way I was treated by the management! They sure were attentive to my needs after I called them sixteen times!

This review uses a lot of language with positive connotations: "loved", "attentive to my needs." But it's sarcastic. We as humans can easily tell because we understand the implied context given be "after I called them sixteen times." This is difficult for sentiment analysis tools to pick up on.

A review like this can be a lost cause for AI:

You get what you pay for!

Even for us humans, without more context, it's impossible to tell whether this was sarcastic.

The Wikipedia article on sentiment analysis gives more examples of other types of phrases that are currently difficult for sentiment analysis AI to understand correctly.

• That AI achieved world domination so it clearly had figured that out already. If we can detect it than the AI is certainly able to do it, too. If it's missing enough information than it can't be used to transport information. – Christoph Sep 19 '17 at 6:48

## All interpretation systems are capable of being fooled.

Your eyes seem like they're really good at discerning the world, but they're not perfect, are they? You have a big blind spot in the middle of your eye that your brain fills in (and can fill in inaccurately). Even the simplest optical illusion demonstrates that the way your brain interprets what it sees can be manipulated with a well-crafted input.

AI is no different. There's a field of AI attack called 'adversarial machine learning', that crafts specific inputs that exploit the AI's interpretation matrix. These aren't bugs - any kind of system complex enough to be useful cannot, mathematically, both be complete and flawless - but they have to be generated based on the specific method being used.

Part of the Human Resistance is dedicated to crafting subversive messages that are classified as innocent by the AI. They have special microphones, speakers and transmitters that inject noise into broadcasts that make it render as innocuous to any nearby robots. They print out orders with special overprinting squares that fool the text recognition systems.

If you have an all-encompassing AI, it's easy to justify only using one method of classification in ways that speak to its personality - this won't work if you need to fool multiple wildly different vision systems at once. You'll need to justify how the Resistance can test its messages, because it needs to be able to try hundreds of different combinations to work out what blobs where start making the vision system more unsure of its interpretation. Perhaps the Resistance has an old version of the vision system, or perhaps they need to capture and subvert a robot, because the AI eventually realises that the robot records misleading logs every time it detects subversive messages.

It's also easy to defeat this only where narratively necessary: human collaborators.

I'm assuming the Computer is using videocameras and not some kind of dust-mote nanobot. Here's what you do. You need some specialized equipment:

• A blanket

• A flashlight

• Whiteboard tablet

• Dry-erase markers

Two conspirators hide under the blanket, and write messages to each other on the whiteboard. When they're done, they erase the whiteboard.

If they're feeling especially malicious, they tap out fake, meant-to-be-overheard statements (e.g. "The Computer is a fink") in morse code.

If they bring a briefcase under the blanket with them, they can carry such a message to a friend's house.

• Humanity was enslaved, robots crawling the earth why would the AI allow that ? – Christoph Sep 19 '17 at 6:45
• @Christoph I'm imagining the AI rolling it's visual sensors and sighing, "Oh those humans hiding under the blanket making weird squeaky sounds again. Wonder what they're up to, wink wink". Of course I'm being a bit tongue in cheek here. Really it depends if we are talking about people conspiring, or just claiming a bit of privacy. If the latter, computer knows they're communicating, but can't tell what they're saying. It probably doesn't care, if they don't fit a saboteur profile. – akaioi Sep 19 '17 at 6:58
• AI has no feelings it wouldn't allow any privacy. Reproduction is only another necessary act they are forced to. An AI that could be tricked so easy wouldn't have taken over the world in first place. – Christoph Sep 19 '17 at 7:01

I believe the answer is 'Yes'.

As Jarrett has already pointed out, if the AI is unaware that there is a code, then it will not take steps to crack it.

Consider a dialogue between two people, suppose one person is always telling the truth and the other person always lying. The conversation is likely superficially to make sense, but the sense understood between those two people will not be the same as the sense understood by someone overhearing.

The protocol could be triggered by a statement commonly understood to be untrue, from which point the communication would alternate between true and false. Equally, it could be triggered the other way by a true but unnecessary statement (because the truth is already known by both parties). At certain points the roles could be reversed in a similar way. The key is essentially what is shared (and known to be shared) as common knowledge.

If managed carefully, it would be very difficult to realise that a code was actually in use. Even if it was realised that communication was coded, it would be difficult to understand how, especially if it was not known which statements were true and which were not.

There are subtle variations which could be applied. For example, the trigger could be making either an easily verified statement, or a statement not verifiable like for example:

"It was light when I woke up this morning", would not be verifiable, whilst "The sun is shining again" is. The statements themselves are simple enough, and very possibly true, but they can both serve, by convention, to initiate the protocol.

It would take practice, but humans are quite good at games like this.

And humans are good at knowing when a game is just a game, and when it is something else. Computers are not, because games have emotional components.