We often hear that artificial intelligence may harm or even kill humans, so it might prove dangerous.
How could artificial intelligence harm us?
There are many valid reasons why people might fear (or better be concerned about) AI, not all involve robots and apocalyptic scenarios.
To better illustrate these concerns, I'll try to split them into three categories.
This is the type of AI that your question is referring to. A super-intelligent conscious AI that will destroy/enslave humanity. This is mostly brought to us by science-fiction. Some notable Hollywood examples are "The terminator", "The Matrix", "Age of Ultron". The most influential novels were written by Isaac Asimov and are referred to as the "Robot series" (which includes "I, robot", which was also adapted as a movie).
The basic premise under most of these works are that AI will evolve to a point where it becomes conscious and will surpass humans in intelligence. While Hollywood movies mainly focus on the robots and the battle between them and humans, not enough emphasis is given to the actual AI (i.e. the "brain" controlling them). As a side note, because of the narrative, this AI is usually portrayed as supercomputer controlling everything (so that the protagonists have a specific target). Not enough exploration has been made on "ambiguous intelligence" (which I think is more realistic).
In the real world, AI is focused on solving specific tasks! An AI agent that is capable of solving problems from different domains (e.g. understanding speech and processing images and driving and ... - like humans are) is referred to as General Artificial Intelligence and is required for AI being able to "think" and become conscious.
Realistically, we are a loooooooong way from General Artificial Intelligence! That being said there is no evidence on why this can't be achieved in the future. So currently, even if we are still in the infancy of AI, we have no reason to believe that AI won't evolve to a point where it is more intelligent than humans.
Even though an AI conquering the world is a long way from happening there are several reasons to be concerned with AI today, that don't involve robots! The second category I want to focus a bit more on is several malicious uses of today's AI.
I'll focus only on AI applications that are available today. Some examples of AI that can be used for malicious intent:
DeepFake: a technique for imposing someones face on an image a video of another person. This has gained popularity recently with celebrity porn and can be used to generate fake news and hoaxes. Sources: 1, 2, 3
With the use of mass surveillance systems and facial recognition software capable of recognizing millions of faces per second, AI can be used for mass surveillance. Even though when we think of mass surveillance we think of China, many western cities like London, Atlanta and Berlin are among the most-surveilled cities in the world. China has taken things a step further by adopting the social credit system, an evaluation system for civilians which seems to be taken straight out of the pages of George Orwell's 1984.
Influencing people through social media. Aside from recognizing user's tastes with the goal of targeted marketing and add placements (a common practice by many internet companies), AI can be used malisciously to influence people's voting (among other things). Sources: 1, 2, 3.
Military applications, e.g. drone attacks, missile targeting systems.
This category is pretty subjective, but the development of AI might carry some adverse side-effects. The distinction between this category and the previous is that these effects, while harmful, aren't done intentionally; rather they occur with the development of AI. Some examples are:
Jobs becoming redundant. As AI becomes better, many jobs will be replaced by AI. Unfortunately there are not many things that can be done about this, as most technological developments have this side-effect (e.g. agricultural machinery caused many farmers to lose their jobs, automation replaced many factory workers, computers did the same).
Reinforcing the bias in our data. This is a very interesting category, as AI (and especially Neural Networks) are only as good as the data they are trained on and have a tendency of perpetuating and even enhancing different forms of social biases, already existing in the data. There are many examples of networks exhibiting racist and sexist behavior. Sources: 1, 2, 3, 4.
Did I mention that all of these are in full swing already?
Although there is no clear line between AI and AGI, this section is more about what happens when we go further towards AGI. I see two alternatives:
In the first case, if an AI "goes rogue" we can build other AIs to outwit and neutralise it. In the second case, we can't, and we're doomed. AIs will be a new life form and we may go extinct.
Here are some potential problems:
I think AGI is coming and we need to be mindful of these problems so that we can minimise them.
In addition to the other answers, I would like to add to nuking cookie factory example:
Machine learning AIs basically try to fulfill a goal described by humans. For example, humans create an AI running a cookie factory. The goal they implement is to sell as many cookies as possible for the highest profitable margin.
Now, imagine an AI which is sufficiently powerful. This AI will notice that if he nukes all other cookie factories, everybody has to buy cookies in his factory, making sales rise and profits higher.
So, the human error here is giving no penalty for using violence in the algorithm. This is easily overlooked because humans didn't expect the algorithm to come to this conclusion.
My favorite scenario for harm by AI involves not high intelligence, but low intelligence. Specifically, the grey goo hypothesis.
This is where a self-replicating, automated process runs amok and converts all resources into copies of itself.
The point here is that the AI is not "smart" in the sense of having high intelligence or general intelligence--it is merely very good at a single thing and has the ability to replicate exponentially.
I have an example which goes in kinda the opposite direction of the public's fears, but is a very real thing, which I already see happening. It is not AI-specific, but I think it will get worse through AI. It is the problem of humans trusting the AI conclusions blindly in critical applications.
We have many areas in which human experts are supposed to make a decision. Take for example medicine - should we give medication X or medication Y? The situations I have in mind are frequently complex problems (in the Cynefin sense) where it is a really good thing to have somebody pay attention very closely and use lots of expertise, and the outcome really matters.
There is a demand for medical informaticians to write decision support systems for this kind of problem in the medicine (and I suppose for the same type in other domains). They do their best, but the expectation is always that a human expert will always consider the system's suggestion just as one more opinion when making the decision. In many cases, it would be irresponsible to promise anything else, given the state of knowledge and the resources available to the developers. A typical example would be the use of computer vision in radiomics: a patient gets a CT scan and the AI has to process the image and decide whether the patient has a tumor.
Of course, the AI is not perfect. Even when measured against the gold standard, it never achieves 100% accuracy. And then there are all the cases where it performs well against its own goal metrics, but the problem was so complex that the goal metric doesn't capture it well - I can't think of an example in the CT context, but I guess we see it even here on SE, where the algorithms favor popularity in posts, which is an imperfect proxy for factual correctness.
You were probably reading that last paragraph and nodding along, "Yeah, I learned that in the first introductory ML course I took". Guess what? Physicians never took an introductory ML course. They rarely have enough statistic literacy to understand the conclusions of papers appearing in medical journals. When they are talking to their 27th patient, 7 hours into their 16 hour shift, hungry and emotionally drained, and the CT doesn't look all that clear-cut, but the computer says "it's not a malignancy", they don't take ten more minutes to concentrate on the image more, or look up a textbook, or consult with a colleague. They just go with what the computer says, grateful that their cognitive load is not skyrocketing yet again. So they turn from being experts to being people who read something off a screen. Worse, in some hospitals the administration does not only trust computers, it also has found out that they are convenient scapegoats. So, a physician has a bad hunch which goes against the computer's output, it becomes difficult for them to act on that hunch and defend themselves that they chose to overrode the AI's opinion.
AIs are powerful and useful tools, but there will always be tasks where they can't replace the toolwielder.
This only intents to be a complement to other answers so I will not discuss to possibility of AI trying to willingly enslave humanity.
But a different risk is already here. I would call it unmastered technology. I have been teached science and technology, and IMHO, AI has by itself no notion of good and evil, nor freedom. But it is built and used by human beings and because of that non rational behaviour can be involved.
I would start with a real life example more related to general IT than to AI. I will speak of viruses or other malwares. Computers are rather stupid machines that are good to quickly process data. So most people rely on them. An some (bad) people develop malwares that will disrupt the correct behaviour of computers. And we all know that they can have terrible effects on small to medium organizations that are not well prepared to an computer loss.
AI is computer based so it is vulnerable to computer type attacks. Here my example would be an AI driven car. The technology is almost ready to work. But imagine the effect of a malware making the car trying to attack other people on the road. Even without a direct access to the code of the AI, it can be attacked by side channels. For example it uses cameras to read signal signs. But because of the way machine learning is implemented, AI generaly does not analyses a scene the same way a human being does. Researchers have shown that it was possible to change a sign in a way that a normal human will still see the original sign, but an AI will see a different one. Imagine now that the sign is the road priority sign...
What I mean is that even if the AI has no evil intents, bad guys can try to make it behave badly. And to more important actions will be delegated to AI (medecine, cars, planes, not speaking of bombs) the higher the risk. Said differently, I do not really fear the AI for itself, but for the way it can be used by humans.
I think one of the most real (ie. related to current, existing AIs) risks are in blindly relying on unsupervised AIs, for two reasons.
Physical error in AI systems may start producing wildly wrong results in regions in which they were not tested for because the physical system starts providing wrong values. This is sometimes redeemed by self-testing and redundancy, but still requires occasional human supervision.
Self learning AIs also have a software weakness - their weight networks or statistic representations may approach local minima where they are stuck with one wrong result.
This is fortunately frequently discussed, but worth mentioning: AI systems' classification of inputs is often biased because the training/testing dataset were biased as well. This results in AIs not recognizing people of certain ethnicity, for more obvious example. However there are less obvious cases that may only be discovered after some bad accident, such as AI not recognizing certain data and accidentally starting fire in a factory, breaking equipment or hurting people.
If a robot is similar to a human machine interface the device is the same like a remote controlled car. It's possible to discuss with the operator behind the joystick and negotiate about a wishful behavior. Remote controlled robots are safe inventions because their actions can be traced back to humans and their motivation can be anticipated. They can be used to improve the daily life, and it's funny to play with them.
In contrast, some robots aren't controlled by joysticks but are working with an internal dice generator. The dice toy is known from it's social role in gambling but it has also a mystical meaning. Usually, a random generator is strongly connected with chaotic behavior which is controlled by dark forces outside the influence of humans. An electronic dice built into a robot and improved with learning algorithm is the opposite of a human machine interface, but it's potential troublemaker because the randomly controlled robot will play games with humans which can't be anticipated. It's not possible to predict the next number of a dice, therefore the robot will behave abrupt as well.
The connection between randomly controlled games and negative social impact was explained in the following sentence.
quote: “In many traditional non-Western societies gamblers may pray to the gods for success and explain wins and losses in terms of divine will. “ Binde, Per. "Gambling and religion: Histories of concord and conflict." Journal of Gambling Issues 20 (2007): 145-165.
Human beings currently exist in an ecological-economic niche of "the thing that thinks".
AI is also a thing that thinks, so it will be invading our ecological-economic niche. In both ecology and economics, having something else occupy your niche is not a great plan for continued survival.
How exactly Human survival is compromised by this is going to be pretty chaotic. There are going to be a bunch of plausible ways that AI could endanger human survival as a species, or even as a dominant life form.
Suppose there is a strong AI without "super ethics" which is cheaper to manufacture than a human (including manufacturing a "body" or way of manipulating the world), and as smart or smarter than a human.
This is a case where we start competing with that AI for resources. It will happen on microeconomic scales (do we hire a human, or buy/build/rent/hire an AI to solve this problem?). Depending on the rate at which AIs become cheap and/or smarter than people, this can happen slowly (maybe an industry at a time) or extremely fast.
In a capitalist competition, those that don't move over to the cheaper AIs end up out-competed.
Now, in the short term, if the AI's advantages are only marginal, the high cost of educating humans for 20-odd years before they become productive could make this process slower. In this case, it might be worth paying a Doctor above starvation wages to diagnose disease instead of an AI, but it probably isn't worth paying off their student loans. So new human Doctors would rapidly stop being trained, and existing Doctors would be impoverished. Then over 20-30 years AI would completely replace Doctors for diagnostic purposes.
If the AI's advantages are large, then it would be rapid. Doctors wouldn't even be worth paying poverty level wages to do human diagnostics. You can see something like that happening with muscle-based farming when gasoline-based farming took over.
During past industrial revolutions, the fact that humans where able to think means that you could repurpose surplus human workers to do other actions; manufacturing lines, service economy jobs, computer programming, etc. But in this model, AI is cheaper to train and build and as smart or smarter than humans at that kind of job.
As evidenced by the ethanol-induced Arab spring, crops and cropland can be used to fuel both machines and humans. When machines are more efficient in terms of turning cropland into useful work, you'll start seeing the price of food climb. This typically leads to riots, as people really don't like starving to death and are willing to risk their own lives to overthrow the government in order to prevent this.
You can mollify the people by providing subsidized food and the like. So long as this isn't economically crippling (ie, if expensive enough, it could result in you being out-competed by other places that don't do this), this is merely politically unstable.
As an alternative, in the short term, the ownership caste who is receiving profits from the increasingly efficient AI-run economy can pay for a police or military caste to put down said riots. This requires that the police/military castes be upper lower to middle class in standards of living, in order to ensure continued loyalty -- you don't want them joining the rioters.
So one of the profit centers you can put AI towards is AI based military and policing. Drones that deliver lethal and non-lethal ordnance based off of processing visual and other data feeds can reduce the number of middle-class police/military needed to put down food-price triggered riots or other instability. As we have already assumed said AIs can have bodies and training cheaper than a biological human, this can also increase the amount of force you can deploy per dollar spent.
At this point, we are talking about a mostly AI run police and military being used to keep starving humans from overthrowing the AI run economy and seizing the means of production from the more efficient use it is currently being put to.
The vestigial humans who "own" the system at the top are making locally rational decisions to optimize their wealth and power. They may or may not persist for long; so long as they drain a relatively small amount of resources and don't mess up the AI run economy, there won't be much selection pressure to get rid of them. On the other hand, as they are contributing nothing of value, they position "at the top" is politically unstable.
This process assumed a "strong" general AI. Narrower AIs can pull this off in pieces. A cheap, effective diagnostic computer could reduce most Doctors into poverty in a surprisingly short period of time, for example. Self driving cars could swallow 5%-10% of the economy. Information technology is already swallowing the retail sector with modest AI.
It is said that every technological advancement leads to more and better jobs for humans. And this has been true for the last 300+ years.
But prior to 1900, it was also true that every technological advancement led to more and better jobs for horses. Then the ICE and automobile arrived, and now there are far fewer working horses; the remaining horses are basically the equivalent of human personal servants: kept for the novelty of "wow, cool, horse" and the fun of riding and controlling a huge animal.
In addtion to the many answers already provided, I would bring up the issue of adversarial examples in the area of image models.
Adversarial examples are images that have been perturbed with specifically designed noise that is often imperceptible to a human observer, but strongly alters the prediction of a model.
AI that is used to solve a real world problem could pose a risk to humanity and doesn't exactly require sentience, this also requires a degree of human stupidity too..
Unlike humans, an AI would find the most logical answer without the constraint of emotion, ethics, or even greed... Only logic. Ask this AI how to solve a problem that humans created (for example, Climate Change) and it's solution might be to eliminate the entirety of the human race to protect the planet. Obviously this would require giving the AI the ability to act upon it's outcome which brings me to my earlier point, human stupidity.
Artificial intelligence can harm us in any of the ways of natural intelligence (of humans). The distinction between natural and artificial intelligence will vanish when humans start augmenting themselves more intimately. Intelligence may no longer characterize the identity and will become a limitless possession. The harm caused will be as much the humans can endure for preserving their evolving self-identity.
Few people realize that our global economy should be considered an AI: - The money transactions are the signals over a neural net. The nodes in the neural net would be the different corporations or private persons paying or receiving money. - It is man-made so qualifies as artificial
This neural network is better in its task then humans: Capitalism has always won against economy planned by humans (plan-economy).
Is this neural net dangerous ? Might differ if you are the CEO earning big versus a fisherman in a river polluted by corporate waste.
How did this AI become dangerous? You could answer it is because of human greed. Our creation reflects ourselves. In other words: we didnot train our neural net to behave well. Instead of training the neural net to improve living quality for all humans, we trained it to make rich fokes more rich.
Would it be easy to train this AI to be no longer dangerous ? Maybe not, maybe some AI are just larger then life. It is just survival of the fittest.