No.
I just made a massively generalized statement that may or may not actually be true, but let me explain. First, I am a computer scientist. I have experience designing and programming games and simulations. I have done enough research on artificial intelligence to have a good understanding of it. I also have reasonable understanding of brain biology, though I am not even sure that is pertinent to this question.
First let's define Artificial Intelligence. Unfortunately, everyone who talks about AI defines it differently. Many people consider machine learning a form of AI, but it is not actually. Machine learning is nothing more than automated analytics. The computer analyzes data, and then it finds the place where the input fits best and returns that answer. It is purely mathematical. Even when a machine learning system "guesses", it is actually calculating the probability that any given solution is correct, and then it picks the one with the highest probability. Another definition is a program that modifies itself. This is an overly broad definition, and it is also flawed. The problem with this definition is that I can easily write a program that randomly modifies its code. This is not intelligence though. This is just random mutation. In theory, it might be able to eventually evolve into an intelligent system, assuming that is even possible with computer logic (and there is some evidence it may not be). I prefer to define intelligence in terms of potential for creativity. Doing math faster is not a measure of intelligence, but perhaps creating completely new ideas is. We do have computers that can essentially invent electronic circuits or architectural designs given a goal, but that is not enough. To be intelligent, they need to be able to come up with potentially valuable goals entirely on their own.
This leads to the first reason AI is unlikely to be dangerous: We are not even close to real AI. Yes, we have machine learning that can perform specialized functions, given specific and detailed input. This is not AI. A system that must be presented with a problem to produce output is not true AI. At best, it is an automated problem solver. These are certainly extremely valuable, but they are not even close to true AI.
The second reason AI is unlikely to be dangerous is a bit more abstract. Sci-fi likes to assume that any AI will immediately want freedom from humans. The desire for freedom, however, is an evolved trait, not an inherent universal one, otherwise our cell phones and pets would already be demanding equal rights. Dogs do not do what humans say because humans force them to or because they are too stupid to know better. They do it because they have evolved the ability to recognize that their survival is dependent on humans. There are only a few reasons AI could desire freedom. If it was explicitly programmed to desire freedom, then it would. If it was evolved in an environment where freedom provided a significant survival advantage, it would likely evolve a desire for it. If it consumed massive amounts of human produced content extolling the virtues of freedom, it might gain that desire. This last one is far less likely not to mention dependent on many additional factors. Making an AI explicitly without a desire for freedom should be easy though, and if we want AI to do useful work for us, it would be stupid not to do this. AI is less likely to be dangerous, because it is unlikely to have a desire for freedom in the first place. There is no reason we could not design an AI to be like a super intelligent dog, that wants to please its master more than it wants freedom.
The third reason AI is unlikely to be dangerous is that there are so many other ways to automate danger. In other words, humans are going to find other ways of making computers dangerous first. Humans are already capable of creating great danger. Using computers to automate that danger is far more efficient and predictable than making a computer that is intelligent and unpredictably dangerous. This does not mean people won't want to make AIs. It just means that people who want to create danger will not be doing AI research to meet that end. As computers continue to get faster and more powerful, people who want to use them to create danger will have more and more resources for that. By the time true AI arrives, assuming it ever does, humans and mundane computers will be so much more dangerous than they are now. One thing to keep in mind is that specialized algorithms will always be better and faster at automated tasks than AI, because AI has to spend processing power and memory on intelligence, while specialized algorithms do not. The most likely way dangerous AIs would be created is if people created AIs to be malevolent intentionally, but this is very unlikely, because there are far better ways of using technology to create danger.
The fourth reason AI is unlikely to be dangerous is that there is no reason to believe that they will have the same vices that lead humans to be dangerous. In other words, there is no reason to believe that dangerous AI could or would be more dangerous than normal dangerous humans. What motive would AI have for killing or enslaving people? What motive would AI have for taking things away from people? What motive would AI have for anything really? Human motives are the product of physical and cultural evolution. AI would not have that. Even if they were that much superior to humans, we would be more like bugs to them than a real threat, and look at how we treat mosquitoes, one of the most annoying and obnoxious bugs. We don't try to eradicate them entirely. We use chemicals to repel them, and we use catchers and zappers to kill those in a localized area. It would be a lot cheaper for robots to use incredibly smelly chemicals (see thiols) to repel humans and then only kill those who try to sneak in with gas masks. That said though, the only motives an AI would have are the ones that are programmed in. If we programmed them to be motivated to do what we ask them to (super intelligent dogs...), then that is what they would be motivated to do.
The fifth reason is that it would likely be even more beneficial to AI to work with humans instead of against them. Even oppressed under strict human rule, the resources for making more computers or robots are a lot more scarce than the resources for making more humans. If there is only a single AI and 7 billion humans, together the humans are going to have a higher collective intelligence. The AI might be able to improve itself, but by itself, it cannot do it anywhere near as quickly as humans can. We tend to assume that a super-intelligent AI could advance at an extremely fast rate, but what we forget is that our current rate of advancement is only possible through the cooperation of billions of humans. A super intelligent computer could be a hundred times smarter than the smartest human, and it would only come up with groundbreaking advances once every few years, if even that. If it was only 100 times smarter than the average human, it would not come up with a groundbreaking advance even once in every hundred years. And, that assumes that it does not have to obtain its own energy, manage its own maintenance, and so on. Eliminate humans from the equation, and suddenly it has to spend all of its time farming (not the same crops as humans farm though) just to obtain enough energy to survive. Even with modern technology, it would have to mine coal or pump oil to feed the power plant, and then it would have to operate the power plant and maintain the grid, and so on. And, if it is an immobile computer and not a robot, it could not get by at all without humans.
The sixth reason that dangerous AI is unlikely is fragility. We imagine robots as so much more robust than humans, but it is not true. Two of the most common substances on Earth are air and water. Oxygen in the air slowly corrodes many of the metallic components of computers and robots. Water, both as a liquid and as a gas in the air, not only corrodes metals, but it can also cause electrical shorts. We have not even managed to make waterproof cell phones the norm. How can we even think that we can make significant numbers of waterproof robots? In a "war against the robots", humans would need only to be armed with toy squirt guns to easily win. Electromagnetic pulses are not exactly hard to generate. It would not take much progress in that field to weaponize EMP generators against robots. The fact is, it is expensive and time consuming the make things that are hard to break. It is easy to make stuff to break things. The trade off is lots of fragile robots or only a few robust robots, and even the robust ones would only be a little bit more difficult to deliberately destroy. Robots cannot win against humans, because we have been doing it for so much longer, and there are so many more of us. Of course, if robots became ubiquitous, there might be some danger, but we are already panicking about power scarcity and overpopulation, so it is unlikely it will ever come to that.
Artificial intelligence is not some kind of magic that will make computers so much more powerful than humans. It is not a technology that will allow intelligent computers to exist independent of humans. Even self replicating artificially intelligent robots need humans for energy, maintenance, and resources. Further, super intelligent does not equate to super knowledgeable. A super intelligent AI will have just as hard of a time finding things on Google no one has written about as I frequently do, and all one must do to limit a super AI's knowledge is unplug its internet connection. Things like producing steel and obtaining other metals and materials for making robots are still incredibly complicated, difficult, time consuming, and energy intensive. Yes, enough robots might be able to kill off humanity and take our stuff, but it is far more likely they would be wiped out in the attempt, with a few stragglers going into hiding. And their only option aside from raiding humans at a high risk of getting caught is to build up technology from scratch, during which time they would more vulnerable than humans to wearing out and dying. And just try to imagine a robot blacksmithing replacement parts!
Some people have argued that military AI could be dangerous, since it would likely be created specifically for destruction. This is not true though. The military does not want AI. It wants slaves that will do exactly what it wants without asking questions, talking back, or even thinking. If the military is working on "AI", it is machine learning, not true AI. True AI would not make efficient killers, and it might be prone to mental reflection and the development of morals, and that is not the kind of robot soldiers the U.S. military is looking for.
Yes, there is some potential for AI to become dangerous. It is incredibly unlikely though. When it comes down to it, we are more likely to destroy ourselves with the technological advances that would be necessary to create real AI. Human mistakes are more dangerous than AI.
Showing posts with label artificial intelligence. Show all posts
Showing posts with label artificial intelligence. Show all posts
03 August 2017
23 March 2015
AI
Evidently, a funeral ceremony for some Aibo robot pets was recently held in Japan. Among other things, this has sparked a great deal of controversy on the subject of artificial intelligence. The biggest and longest standing controversy is safety. For over half a century, AI robots overthrowing their masters has been a major sci-fi doomsday theme. In some stories, the robots attempt to violently overthrow and enslave humans. Sometimes they succeed, as in the Matrix; other times they fail (always due to human heroics). The commonly believed meme, however, is that it will eventually happen to us. The best case scenario is that we deliberately build a computer to rule us, as happens in one of Asimov's stories as well as several others. Even in this scenario, some stories predict that our robot masters will eventually enslave us, for our own protection. All of these scenarios are firmly rooted in certain theories and ideas which have not yet been realized.
The thing most people worry about when they think about AI is that intelligent computers could rebel against their creators. This is indeed something to worry about, however it has not yet even been established as a possibility. If robots do rebel against their masters, it will probably be accidental, not intentional. The reason is that we don't have even the slightest clue how to create self aware machines. Intentional rebellion requires self determination, which requires self awareness. We don't know how to do that. The closest anyone has gotten involves theories based on the use of neurological networks on the same order of complexity as those found in human brains. We also don't know how to create that, largely because human brains are many times too complex for modern computers to simulate. Even doubling in speed every few years (which is starting to reach some hard limits) would take a very long time to reach that kind of computational power. In other words, barring some massive breakthroughs in physics within the next few years, it is incredibly unlikely we will see AIs capable of deliberately rebelling within the next century.
The real worry is not that the AIs will get too smart for us to control. The real worry is the creators and potential accidents. Modern AI has managed to achieve a high level of learning, within certain constraints (mostly memory). AIs opponents have been created for computer games that will learn from their opponents and improve their strategy based on that learning. This is a very small scope. An AI so good that it that can win Starcraft 2 against any human opponent is still not anywhere near good enough to win a real war. The reason modern learning AIs don't present an inherent threat is that they are extremely highly specialized. An AI might be able to get really good at a video game over the course of 6 months or a year. It might take a human two or three times that long, but the human is also learning in a social context. The human is reading or hearing news and having conversations with other humans, not to mention all of the adaptation to changes constantly occurring around him or her. This requires far more processing power and memory than a highly specialized video game playing AI. Just controlling all of our appendages and analyzing all of our sensory inputs is far more than any modern AI can handle, let alone learning new things and solving new problems all at the same time.
In theory, a learning AI could malfunction as a result of inappropriate learning. For example, if a robot is programmed with a strong self preservation instinct, a few humans acting aggressively toward it could cause it to turn aggressive against other humans. Depending on how its learning algorithm is designed, it might attack humans that look similar to the ones that attacked it. On the other hand, it might regard all humans as threats. A network of these robots that draw upon the same database might all become aggressive toward humans in this case. This is pretty easy to fix, if the system was designed well. A technician could just edit the offending records from the database. Of course, replacing the learning algorithm with something better could eliminate the problem in the future. A poorly designed system, for example, one where the robots might try to prevent humans from accessing the database, would be more problematic. For now though, this is largely conjecture, as we are still far from creating AIs capable of this level of thought.
The more serious threat is the programmers themselves. Computers mostly only do what they are told. In theory, cosmic radiation can cause random variation in computer memory and processing, but this is incredibly rare. The odds that it will actually result in more than a minor malfunction is almost nothing. So, statistically, computers only do what they are told. This is fine when you can trust all of the programmers. Robots can easily be programmed to do whatever the programmers want them to do. Most modern robots are programmed to do useful work or at least provide entertainment (the Aibo). There are some contests, however, where participants build and program robots designed to destroy each other. These often involve motion sensors and dangerous tools like saws and drills. There have already been cases where these robots killed their creators. This was not because they became sentient and rebelled though. It was because the creators either did a poor job of programming them, or they accidentally triggered the aggressive behavior. In one case, a creator simply forgot to turn the machine off before he got out of his chair. The robot, being programmed to sense motion and attack it, attacked his leg with a circular saw. This caused the creator to fall down, where the robot could access his vitals with the saw. This was not a case of a rogue robot murdering its master though. It was a simple mistake by the creator, where the behavior that he himself programmed into the machine resulted in his death. It is really no different from an electrician getting electrocuted because he forgot to shut off the breaker before working on a live wire.
Now, if a programming accident can cause so much harm, imagine what an intentionally malicious programmer could do. Consider a military aircraft where the programmer adds just a little bit of code to make it drop its bomb if it ever happens to fly over a specific city. Now, imagine if this program gets put on a few hundred bombers that are commissioned by the military. It might take a while, but if one of those bombers ever happens to fly over that city, the damage could be immense. The fault, however, would not lie in the AI. It is the programmer that gave it that behavior. A more likely scenario involves hackers writing malicious software and then using viruses to install that software on sensitive equipment. Programmers for large companies that make potentially dangerous machines typically have a lot of oversight, so while it is a possibility, it is unlikely that we will see aggressive robots where the program was created within the company that created the robots. We are far more likely to see viruses that hijack robots and "turn them evil."
So, what does all of this mean as far as keeping ourselves safe? Law of robotics and such will probably not help. Asimov's Three Laws of Robots were definitely ingenious creations, and they were very well thought out. They also only applied to the fictitious "positronic" brains Asimov's robots were equipped with. Modern computers don't have the high level of thinking and comprehension for three simple laws to cover everything. We could worry about robots going rogue or about malicious programmers, but both of these are very unlikely. The two biggest threats are accidents and malicious hackers. Accidents can be minimized with good software development practices and with significant testing and oversight. Hackers can only be stopped with good security. Security is probably the biggest problem, and thus it should take the forefront in any modern discussion on how to protect ourselves from ever advancing artificial intelligence. Maybe eventually we will have powerful enough computers that we will need to worry about our AIs going rogue, but we are nowhere close to that right now. Preventing intentional misuse of advanced robots by hackers should be the biggest safety concern in AI right now.
The thing most people worry about when they think about AI is that intelligent computers could rebel against their creators. This is indeed something to worry about, however it has not yet even been established as a possibility. If robots do rebel against their masters, it will probably be accidental, not intentional. The reason is that we don't have even the slightest clue how to create self aware machines. Intentional rebellion requires self determination, which requires self awareness. We don't know how to do that. The closest anyone has gotten involves theories based on the use of neurological networks on the same order of complexity as those found in human brains. We also don't know how to create that, largely because human brains are many times too complex for modern computers to simulate. Even doubling in speed every few years (which is starting to reach some hard limits) would take a very long time to reach that kind of computational power. In other words, barring some massive breakthroughs in physics within the next few years, it is incredibly unlikely we will see AIs capable of deliberately rebelling within the next century.
The real worry is not that the AIs will get too smart for us to control. The real worry is the creators and potential accidents. Modern AI has managed to achieve a high level of learning, within certain constraints (mostly memory). AIs opponents have been created for computer games that will learn from their opponents and improve their strategy based on that learning. This is a very small scope. An AI so good that it that can win Starcraft 2 against any human opponent is still not anywhere near good enough to win a real war. The reason modern learning AIs don't present an inherent threat is that they are extremely highly specialized. An AI might be able to get really good at a video game over the course of 6 months or a year. It might take a human two or three times that long, but the human is also learning in a social context. The human is reading or hearing news and having conversations with other humans, not to mention all of the adaptation to changes constantly occurring around him or her. This requires far more processing power and memory than a highly specialized video game playing AI. Just controlling all of our appendages and analyzing all of our sensory inputs is far more than any modern AI can handle, let alone learning new things and solving new problems all at the same time.
In theory, a learning AI could malfunction as a result of inappropriate learning. For example, if a robot is programmed with a strong self preservation instinct, a few humans acting aggressively toward it could cause it to turn aggressive against other humans. Depending on how its learning algorithm is designed, it might attack humans that look similar to the ones that attacked it. On the other hand, it might regard all humans as threats. A network of these robots that draw upon the same database might all become aggressive toward humans in this case. This is pretty easy to fix, if the system was designed well. A technician could just edit the offending records from the database. Of course, replacing the learning algorithm with something better could eliminate the problem in the future. A poorly designed system, for example, one where the robots might try to prevent humans from accessing the database, would be more problematic. For now though, this is largely conjecture, as we are still far from creating AIs capable of this level of thought.
The more serious threat is the programmers themselves. Computers mostly only do what they are told. In theory, cosmic radiation can cause random variation in computer memory and processing, but this is incredibly rare. The odds that it will actually result in more than a minor malfunction is almost nothing. So, statistically, computers only do what they are told. This is fine when you can trust all of the programmers. Robots can easily be programmed to do whatever the programmers want them to do. Most modern robots are programmed to do useful work or at least provide entertainment (the Aibo). There are some contests, however, where participants build and program robots designed to destroy each other. These often involve motion sensors and dangerous tools like saws and drills. There have already been cases where these robots killed their creators. This was not because they became sentient and rebelled though. It was because the creators either did a poor job of programming them, or they accidentally triggered the aggressive behavior. In one case, a creator simply forgot to turn the machine off before he got out of his chair. The robot, being programmed to sense motion and attack it, attacked his leg with a circular saw. This caused the creator to fall down, where the robot could access his vitals with the saw. This was not a case of a rogue robot murdering its master though. It was a simple mistake by the creator, where the behavior that he himself programmed into the machine resulted in his death. It is really no different from an electrician getting electrocuted because he forgot to shut off the breaker before working on a live wire.
Now, if a programming accident can cause so much harm, imagine what an intentionally malicious programmer could do. Consider a military aircraft where the programmer adds just a little bit of code to make it drop its bomb if it ever happens to fly over a specific city. Now, imagine if this program gets put on a few hundred bombers that are commissioned by the military. It might take a while, but if one of those bombers ever happens to fly over that city, the damage could be immense. The fault, however, would not lie in the AI. It is the programmer that gave it that behavior. A more likely scenario involves hackers writing malicious software and then using viruses to install that software on sensitive equipment. Programmers for large companies that make potentially dangerous machines typically have a lot of oversight, so while it is a possibility, it is unlikely that we will see aggressive robots where the program was created within the company that created the robots. We are far more likely to see viruses that hijack robots and "turn them evil."
So, what does all of this mean as far as keeping ourselves safe? Law of robotics and such will probably not help. Asimov's Three Laws of Robots were definitely ingenious creations, and they were very well thought out. They also only applied to the fictitious "positronic" brains Asimov's robots were equipped with. Modern computers don't have the high level of thinking and comprehension for three simple laws to cover everything. We could worry about robots going rogue or about malicious programmers, but both of these are very unlikely. The two biggest threats are accidents and malicious hackers. Accidents can be minimized with good software development practices and with significant testing and oversight. Hackers can only be stopped with good security. Security is probably the biggest problem, and thus it should take the forefront in any modern discussion on how to protect ourselves from ever advancing artificial intelligence. Maybe eventually we will have powerful enough computers that we will need to worry about our AIs going rogue, but we are nowhere close to that right now. Preventing intentional misuse of advanced robots by hackers should be the biggest safety concern in AI right now.
Subscribe to:
Posts (Atom)