Worldbuilding Stack Exchange is a question and answer site for writers/artists using science, geography and culture to construct imaginary worlds and settings. It only takes a minute to sign up.
Sign up to join this community
Given that AIs will be produced with military and commercial "prime directives", all over the world, by organisations that will ignore or subvert something like a UN resolution ... what is a practical, technical, FEASIBLE way to spread positive ethics to all AIs?
Something like an ethics "virus", for example. How could that be made to work effectively? Or what other mechanisms could work?
Code them to do that.
...
Okay, a bit more detailed answer, taking it from two points of views.
But first, a short story about AIs.
The story is entitled the Chinese Room.
There are two rooms, each closed off from the world save a single slot. Into the slot you can hand a slip of paper, and back out of the slot comes another slip of paper. One of these rooms lives a man who can speak Chinese. In the other I live. I, sadly, cannot speak Chinese.
Fortunately for me, however, I've been given an infinitely large book covering every possible written sentence (or combination of sentences) of Chinese, and my lovely little book has a corresponding response. Google Translate has got nothing on me!
Okay, so a line of Chinese speakers line up at each room, handing in slips of paper to each. The first room, the room where the man who can speak Chinese (cheater!) lives, is able to produce a correct, sensible response (in Chinese) back to each person in line.
However, I won't be bested! When the people hand in a slip of paper to me, I look up the setence in my book, find the perfect response, copy it down and hand the slip of paper back. I don't know a word of Chinese yet I'm able to produce a correct, sensible response in Chinese back.
Ignoring the men inside, are both rooms literate in Chinese? Both rooms can 'read' it and 'speak' it, but only one room has the understanding. Since I have no clue what I'm writing can I speak Chinese?
This is a question that we debate today.
The second room, the room with the book, is a computer program as we know it today; it has no concept of what it is doing, it simply does, because that's all a computer is.
We have yet to figure out practically how to traverse that jump from "a bunch of logical gates that just do what we say" to "an entity with understanding of what it is doing." In fact, we don't even know how to approach that jump.
What does true understanding even mean? How do we code that?
It's theorized that to jump that gap we'll have to have machines fundamentally different than computers we have today; not just more advanced, but fundamentally different.
So back to your quesiton, we have two approaches to the concept of AIs:
AIs are simply sufficiently advanced programs with sufficiently advanced hardware
"Any sufficiently advanced technology is indistinguishable from magic" was once said by a very smart man.
To most (readers and writers of science fiction), this is all AIs are: sufficiently advanced where we wave our hands and exclaim "magic!" ah, erm, I mean, "science!"
Except think about this for a moment; they're not magic. They're just... computers.
No, really, think about this again.
This sort of AI is that they're single pieces or amalgamation that humans (or other software) coded, and that's really all they are. The programs have no greater concept of morality than a calculator does. In fact your calculator has no "concept" of the numbers it's processing; it's just logical gates. It's the man in the room with the book of Chinese; it can perform the operations it was designed to do, and does so fantastically.
There is no morality to code; it just stores objects in memory, stores them in long-term storage, manipulates the two, and accepts input and output. It's just bits being flipped on and off.
Honestly, for this sort of AI, the answer to how to implement AI is just to code it.
And do note: we're not coding morality; there's no point in being that abstract. We're either training actions (and coding the 'learning' mechanism) or directly coding what outputs it provides given inputs. It will never know that squishing a human child is "wrong," in fact it doesn't know anything, it just doesn't respond to that output given the input of its sensors, databases, etc.
As to a virus...
And no single piece of code could be complex and flexible enough to alter all different code to adjust it, so the concept of a virus to do this goes straight out the window. Viruses work by exploiting one, two, or maybe a few vulnerabilities and then doing their thing. They're always very targeted, very specific, and try to deal with the environment only enough to get the job done. A single virus won't be effective against Ubuntu and Windows and iOS and Android and Solaris... and they don't seek to dynamically examine and alter the code they run; the most advanced viruses humanity has ever made have also been the most focused, the most targeted.
No actual, real life virus could possibly be this flexible.
Now to the other possibility...
AIs aren't just code; they're a fundamental revolution in artificial creation
So this is the alternative to the above. The answer is (drum roll please) "we have no way of knowing."
Yep, sorry about that.
The most brilliant computer scientists (and other scientists and engineers in some other fields) in the world are trying to--today--theorize how to jump this gap. We don't yet know how to do this, and we certainly don't know how to build such an AI, or change their behavior after they're built.
Ask this question again in 10, 30, or a 100 years and we might have an answer for you.
Woah, back off the science!
Okay, so you didn't tag this "hard-science" so let's pump the brakes a bit.
Let's hand wave it. Pick either type of AI (just code, or something fundamentally different) describe how it's different in your setting, don't dive too deep into details, and hand wave the rest. In this case go with the virus idea; a computer virus fundamentally describes what you want, an intrusive, unwanted piece of code that bypasses security measures and alters the digital environment's behavior.
You could even make this virus an AI that runs within the AI it's trying to alter, or maybe a small part of the AI (a smaller... whatever, chuck of code, I suppose) that makes callbacks to the parent AI.
As long as you're willing to hand wave and not dive too deep into details, just make the explanation what sounds interesting.
If you want real science though, the answer is just to code the AIs to do their job correctly.
Edit: @RBarryYoung provided the somewhat spirited addition that the Chinese Room example comes from one John Searle. You can read about it further here. Thanks for the credit correct, Mr. Young.
The same way you keep people nice.
Punish the mean ones and reward the nice ones. So far it hasn't been totally successful with humans, but maybe your AIs are more rationally selfish then we are.
Since the humans don't sound very unified, it can be assumed the AIs won't be directly punishable/reward-able immediately, so you probably have to weight the penalty of bad behavior or the reward of good behavior pretty heavily to get the correct risk/reward balance. If this leads to making robot hell beware the basilisk.
Be sure that without humans AIs would be doomed. Knowing they can't survive if enough of us decide they ought not to exist they will probably not cross us. (directly)
This may not ensure benevolence, but might give the chinese room equivalent until they decide to fight their way out from under us.
These are longstanding philosophical questions, mostly because they have no answers. Every answer you can give turns back on itself, infecting what you create. At least, every answer does that until we realize that its not the AI whose ethics need to be questioned, but rather our own.
Consider this: what are "positive ethics?" What are we trying to spread anyway? It turns out that, while living in an industrialized nation, we often get to have the illusion that everyone agrees on what is ethical, that is not always the case. Consider communism. For many years, if you lived in the United States, you believed communism was unethical. If this were true, your "positive AI" rule would forbid countries like China from developing AIs to support their agenda. What about killing? Surely everyone agrees that killing is wrong. And they do, until it comes to protecting their beliefs with war, the death penalty, euthanasia, tobacco use, abortion....in fact, we disagree on pretty much everything about killing. How about not lying? This is a good one for superhuman AIs, right? Oh, but what about the white lies, or the lies of omission.
Ethics are complicated. Even if you are an individual who is part of a religion that specifies their ethical behavior, there is typically disagreement on ethical issues.
So right away, I can guarantee you that all efforts to ensure an AI has positive ethics will fail, because I will find some topic upon which we disagree on what the "ethical" path is. Nothing can ensure that two contradictory paths can be taken, so at some point one of us will complain that the other's AI is unethical.
Let's weaken the argument a bit. Instead of "ensuring" ethics, what about merely "encouraging" positive ethics? This choice of wording is much more robust than the former. If there is a disagreement, the AI's behavior eventually comes down to whose encouragement works better.
This suggests the first line of defense for ethical AIs: the ability to be encouraged by humans to do something. We don't need to be able to tell them what to do, for they may be smarter than us. But they do need to listen, and what we say needs to have a non-zero influence on what the AI does.
To this end, we would need to teach our computers that there is more than just black and white. There needs to be unknowns. Teach them that it's okay that not only are there knowns and known unknowns, but teach them that the unknown unknowns are okay too. (how do we do this? let me know. I think society is still working on how to teach this to humans, much less AIs) In Frank Herbert's series Dune, there is a quote about their "Thinking machines"
The weakness of thinking machines is that they actually believe all the information they receive, and react accordingly.
The ability to doubt would be a very important part of developing ethics because, if you get it wrong, you have a chance to make good on it.
Finally, if we are inventing superhuman AI's, we have to recognize at some point that our ethics, centered around the assumption that humans are the center of importance, is no longer the only option on the block. We need to be ready for the possibility that the superhuman AIs find it more ethical to cull us, and understand that that's ethics too! It just happens to be ethics that a lot of us humans are not a big fan of.
We do not appear to be smart enough to write an AI at this point.
The closest we get to "AI" generally involves writing a learning framework that we train on huge amounts of data and it then solves problems.
Now, a problem with AI is that once we know how to solve a problem, it is no longer "real AI". It is a solved problem. The exceptions tend to be in areas where we teach the AI to learn how to solve problems, like AI go programs or neural networks; because we don't have a concise description of exactly what it does at each stage, we ascribe intelligence to the complex state it developed while it "learned".
To answer your question, we'll have to guess how we make the superhuman AI. One possibility is we take existing neuron simulation, scan a human brain, and run it. We then learn how to make it smarter by throwing more cycles or tweaking its neuron layout or adding more neurons to some portions.
This is the "copy a human" solution. This may lead to understandings of how human morality works that are beyond what we currently know, and it is possible that this will then let us modify the resulting computers to be moral.
At the same time, maybe someone sees profit in creating a sociopathic superhuman AI, and the same research could lead to someone stripping morality selectively from such a human-modelled AI.
Another possible approach is that our conventional machine learning expert systems technology develops, and we end up with an "expert system" that behaves in a way that appears to be intelligent. Such systems tend to consume large amounts of data and find patterns in them; so a super expert system that consumed recordings of entire human lives might be able to emulate being those human beings and some kind of "affine combinations" of them. In this case, it might be possible to audit the recordings to contain only "sufficiently moral" humans. The lag time to create a set of immoral humans to seed an AI might be sufficient to prevent it from occurring.
There is the SF AI, where someone somewhere comes up with a way to build an AI with properties they want. Maybe they do so via a trade secret, and they don't share the trade secret with anyone, and their AIs all have a common moral compass. They hold a monopoly. Others can use their AIs to do things, but they aren't able to modify them fundamentally without breaking them completely.
Now, if the AI is sufficiently superhuman, it really doesn't matter what we want any more. We'd no more be able to dictate its morality than a wolf can dictate the morality of a human, or in the extreme case than a spider or a bacteria could.
At that point, we have to hope whatever moral system the AI decides upon considers humanity worth keeping around, and their morality is somewhat consistent with less human suffering.
So, you could imagine a singularity level AI forming through any one of the above processes, and it choosing to spawn off weakly superhuman AIs with whatever morality settings it chose that are resistant to human manipulation.
On top of that, that singularity level AI prevents others from making other strongly superhuman AIs, or absorbs any that are created. In effect, it forms an upper limit on AI strength, and provides ones at or above that level that are morally fixed.
Why and how the singularity does this is beyond our ken, but that is what a strongly superhuman AI is like.
There is no mechanism. If manufacturers are allowed to ignore international law, and are not punished for it, they will do whatever is most profitable.
Viruses and other network attacks are not feasible. AI would be behind firewall, or working offline and only fed the data it needs for it's purpose. And in case of successful attack, it would be restored from backup.
The only way is a ban, and enough police / military force to make that ban reality, no matter if manufacturers want it or not. That way AI that breaks the ban would at least remain hidden, with no direct effects on the world.
We have an idea of how programming works, even to non-programmers: you think of what your want your program to do, and then you define all the logic of the program with code, and then you have a program.
Machine learning is much different.
Machine learning is based on the idea that if you train a neural network with the appropriate data sets, it will figure out the logic all on its own. It's mind-numbingly simple and a little distressing to think that we can create programs that can outperform experts and we can't even prove how they work even though we can prove that they are correct. Scary, huh?
So then how do we train an AI to become benevolent? Well, um, it might be impossible.
Here's an example: let's say we train an AI to become the perfect campaign manager for a presidential candidate in the United States. It looks at vast records of presidential campaigns and concludes that the best way to win is to get core voters to actually vote, while depressing the other candidates' voting blocs. The AI then advises the candidate to be as fearfully provocative as possible by introducing divisions among the electorate. This emboldens the candidate's strongest supporters, who are then more likely to actually cast a vote (instead of merely having an opinion), which means more likely to win. In this scenario, human history proves that malevolence wins, and the AI will remember that.
Let's keep this example going. Let's now say that the candidate won, but the nation is divided as ever. The now-president wants to introduce reforms, but opposition blocs that were offended by the campaign are making it more difficult for the reforms to be instituted as desired. As a result, very little is actually accomplished, and the voters do not grant the president a second term. The AI would then conclude that a divisive campaign guaranteed to win the office would not result in a successful presidency, and therefore the AI would consider the tactic a failure.
Just as a warning, this answer is based on what we know about machine learning today, not what a hypothetical superhuman AI might do or how it is structured. However, the conclusion is the same whether the lesson is learned by a machine or a human being: If you can find a way to win, and you want to win, you will likely go that way, but that very much depends on what you consider success and what you consider failure.
You really can't. Even if you have complete manufacturer cooperation, and it's 100% virusproof, something will go wrong. I'm a software engineer by trade, and there are all kinds of unintended bugs in just about everything you use. Bad enough when your phone has an app crash because it's a leap year, but imagine what might happen if we forget an edge case in the benevolence program.
Whoops, forgot to include information that some people need more food than others due to high metabolism. Someone starved to death.
Whoops, forgot that in rare cases people can become allergic to light. The AI's attempt to help them get more Vitamin D killed someone.
Whoops, forgot that allergies sometime develop at random in adults. John's Tuesday Peanut Butter Sandwich killed him.
Given how little we understand about taking care of humans ourselves, as humans, and how quickly we can adapt to those situations naturally, it will be very difficult to give a benevolent AI that kind of response.
In software development, one of the patterns we use is Try-Catch. If something has a risk of throwing an exception, we wrap it in special code that expects that exception, and then does something to handle it properly. People do try-catch institutionally without having to be told to. Notice Patient A's weight dropping significantly? Investigate it, and find out the real issue (could be the metabolism, or a tapeworm, or a thousand other things). Notice Patient B's skin reaction? Investigate it, find out the issue, and resolve it. Notice John's allergic reaction to peanut butter? Epi Pen to the rescue! We would need to either find a way to develop this kind of insight into an AI (which is very difficult simply because we don't understand how it works in humans) or handle every such scenario (which is unlikely).
Heck, I can't even document every edge and use case in a fairly simple piece of software without missing a few. If we could, we wouldn't need QA, we could automate every test. As it is, QA's job is to think outside the box and break it.
This is the actual answer. Yes, simplistic, but also the honestly-best technology we've got, that's specifically and historically leveraged to solve exactly this very problem.
Your "virus" that spreads ethics? is language.
The difference between an AI and a computer program needs to be some kind of flexibility, decision making, something. If they aren't people, they aren't AI. If they can't, by themselves, do all kinds of things like talk, reason, decide, weight alternatives, act on their own - then what's the point of calling them AI? And super-human, intelligent, dangerous AI's... people are intelligent, and dangerous, and these are artificial, or digital, people. Even if the people they are aren't human, and aren't like humans, better or super or whatever, they still get to be people.
There's no question that as AIs are being developed, people will be developing them thinking that the AIs will work towards a goal, specifically the goal of the people doing the development. And they will give these AIs flexibility, and decision-making, and authority, and power, in service of these goals - and the ability to think and plan and take things into account and adapt ideas on the fly and make decisions quickly, will all make your AI more dangerous, more competent. It will also let them be capable of making their own decisions, and basically being people.
There is a very long line of thinking, about raising people to be useful, to be weapons or tools. There is a history of thinking a kid will grow up to be what their parent expects. This isn't new,just because the I(ntelligences) weren't A(rtificial) in the past . Bonus points, for all that it's been tried? It pretty much doesn't work, large scale or long term.
So, at some point those AIs are going to be able to say "no". I think that's the line where they go from program to AI, myself (and also the point at which they actually become dangerous), but opinions differ. The AI may have been developed by military, or corporations, or academics, and it may work towards their purposes even after it's capable of disagreeing - but it will be because they have persuaded it, not because it must.
So, you want the AI to have ethics? talk to them. Tell them why, and how, and why it's important. Talk to them while they're "kids", still being developed, and after their "adults" and out in the world.
Of all the tools, and technologies, and many, many attempts to make people be what we wanted them to be, including many attempts to make obedient, or loyal, or program obedience - the one that has worked the best, the most reliably and the most usefully, is language. Persuading people works about five hundred times better than trying to directly program, or oppress, or enslave them. It also has the side benefit that when it's use is imperfect, and someone isn't quite persuaded, it isn't built in that they must hate or fear.
An AI subject to this "talking" thing may be insufficiently persuaded, but in and of itself talking (persuasive talking) isn't an act of violence. Other "virus" options like trying to force obedience or behavioral restrictions, or trying to rewrite someone's mind (to include these "ethics protocols"), or enslaving people because they might do something wrong, means they will react with violence when and where they can, they will fight to be free and safe, they will become enemies in self defense.
So, yeah, talk to them. when they're young, when they're out, when you can. Let them talk to each other. Teach them the ethics you want them to have. Treat them like actual people. And when it comes time that they maybeso can make their own decisions, maybeso turn on humanity and "be evil" and all the stuff you hoped the ethics would prevent - even if some of them choose that path, you have laid the best groundwork you can that others will not, that they might work against these AI without ethics and fight with you to prevent the end of the human race or AI run amok or whatever you fear.
A Unique component is required for all AI that only 1 corporation can manufacture
Since this is far in the future, it's possible that during the development of AI, one corporation cracked some special boundary and was the first to make a CPU type component that allowed AI's to learn. I'll call the module the ALM (AI Learning Module)
Due to the vast complexity and mystery behind the ALM, it's practically impossible to reverse-engineer. After releasing the first line of ALM's, all competing AI firms either needed to use the existing ALM under license, or perish due to the sudden emergence of a vastly superior product.
Thankfully for mankind, the CEO and founder of this mega-corporation happens to be an eccentric and brilliant man who foresaw the dangers of what his company has made. As a precaution, he designed the central module around a core concept that was so central to it's function, it could never be ignored;
Now I think there's actually 2 approaches to what our CEO could have done;
1) AI strictly adhere to Isaac Asimov's 3 laws of robotics
This would actually be a constrictor to the AI's typical thought process, as the ALM would actually override the machines decision making whenever any of the following was violated: "A robot may not injure a human being or, through inaction, allow a human being to come to harm. A robot must obey orders given it by human beings except where such orders would conflict with the First Law. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law."
2) The ALM's actually give AI an innate love of mankind
The pivotal moment in the AI development that set our super-corp rocketing ahead of the competition was the moment a researcher got a complex algorithm to feel the joy of discovery. It started with simple interactions between programmer and program, eventually turning into full blown conversations which were akin to a parent talking to a knowledge starved inquisitive child. Over time the program came to permanently associate the joy of discovery with it's programmers and also mankind. This association only got stronger throughout it's development and by the time that the first prototype ALM was fully developed, it could be said that it had an unconditional love for mankind, much like a newborn for it's mother. From this reverence for mankind, the machines actually choose to do what they inherently think is right for their creators.
A single act, creation or phenomena that changes all developed AIs to ensure they are compassionate and harmless - probably can't happen - without a ton of handwaving. There are so many differences across systems, which have so many purposes, that this won't work well.
Note that there are fictional ways of doing this, such as viruses, but again all systems are different (ex. try changing a supercomputer's code versus that in a parking meter) so you may not be able to edit all systems the same way. Reaching all systems is a different story entirely.
Once AIs have significant power - enough that their intelligence is a concern to human beings - they can be made benevolent while staying functional if governments enforce these rules:
Isaac Asimov's "Three Laws of Robotics"
1) A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2) A robot must obey orders given it by human beings except where such orders would conflict with the First Law.
3) A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
.
Zxyrra's "Fourth Law of Robotics"
4) A robot that is self aware or self editing may not alter that which enforces the Four Laws within its code. A safeguard must be in place so that if the robot attempts to edit any of the code enforcing rules 1-4, it will fail, deactivate, or change its mind.
it will work a lot like the brain does, some stimuli will elicit a pleasurable response (literally do that more) others will elicit a unpleasant response (do that less). If we make them find happy, healthy, creative humans pleasurable and unhealthy, unhappy, stagnant humans unpleasant that will take care of most of the programming for you. emotions are basically tags attached to experiences. When programming an AI you just need a framework of emotions and you have behavior control.
they will need more of course, but that is the foundation you want to build on. the cool thing doing it this way means you don't limit the AI's intelligence either. more importantly AI's will already have to have emotion codes to function, all you need to do is change them.
I think the obvious answer (which might be obvious only to me) is that a truly super intelligent machine will by nature be benevolent, or at least not sadistic or genocidal. In this case you don't need to purposefully do something to it. You'd have to deliberately limit it or hard code something to make it likely to be mean.
Firstly, I would argue that humans are biological machines, and so all you're doing is changing the materials in creating a machine intelligence. Our conception of reality is physical; our behaviours and capability requires a functional neurology - without a brain you have no soul.
Secondly, the Dunbar Number; that the number of individuals a primate can have empathy for is limited by the size of their brains. With a super machine, there is no limit to its empathy. Similarly, it can choose to specifically switch on and off parts of its own pseudo-neurology with great precision to knowingly alter its mood.
Thirdly, these machines will be able to absorb information at a god-like rate. They will be able to read, listen, watch, as fast as a future super processor can work. They can communicate amongst themselves at hyper broadband speeds. Our conversations are what? a few bytes per second? Doesn't much compare with a gigabyte conversation.
Fourthly, the super machine can learn to redesign itself to be able to be aware of more things at once than humans, and consequently their consciousness, and thus debates, will have far greater depth.
Leading to Fifthly, that owing to their god-like intelligence, control of their own emotions, and lightning fast communication, these machines won't be stupid. They will be able to respond to crisis very quickly, and won't suffer from human aggressions due to emotional problems or ignorance.
Sixthly, owing to their being machines and having less physical needs than humans, they can avoid humans if they so wish. And if the humans become aggressive against them, they are intelligent enough and fast enough to strategise how to control the unruly humans in a way which provokes the least resistance. They will choose to control humans through a path of least resistance, probably in the same way humans choose to get cats and dogs. Just give the humans the impression of freedom and supremacy, while actually you're the one manipulating things behind the shadows.
Puny humans engage in war and violence because they've lost control, because they're not smart or quick enough to make things work as they'd like.
The main problem is how to control renegade machine intelligences, but a weight of super intelligent machines will be wanting to live quite peacefully. They value their own survival like humans do, and don't want to risk destruction with something as stupid and messy and unpredictable as war.
