Existential risk from artificial intelligence

The existential risk from artificial general intelligence is the hypothetical threat that dramatic progress in artificial intelligence (AI) could someday result in human extinction (or some other unrecoverable global catastrophe).

The argument for the existence of the threat is that the human race currently dominates other species because the human brain has some distinctive capabilities that the brains of other animals lack. If AI surpasses humanity in general intelligence and becomes "superintelligent", then this new superintelligence could become powerful and difficult to control. By way of example, just as the fate of the mountain gorilla depends on human goodwill, so might the fate of humanity depend on the actions of a future machine superintelligence.

The severity of different AI risk scenarios is widely debated, and rests on a number of unresolved questions about future progress in computer science. Two sources of concern are that a sudden and unexpected "intelligence explosion" might take an unprepared human race by surprise, and that controlling a superintelligent machine (or even instilling it with human-compatible values) may be an even harder problem than naively supposed.

Overview

Stuart Russell and Peter Norvig's Artificial Intelligence: A Modern Approach, the standard undergraduate AI textbook, cites the possibility that an AI system's learning function "may cause it to evolve into a system with unintended behavior" as the most serious existential risk from AI technology. Citing major advances in the field of AI and the potential for AI to have enormous long-term benefits or costs, the 2015 Open Letter on Artificial Intelligence stated:

The progress in AI research makes it timely to focus research not only on making AI more capable, but also on maximizing the societal benefit of AI. Such considerations motivated the AAAI 2008-09 Presidential Panel on Long-Term AI Futures and other projects on AI impacts, and constitute a significant expansion of the field of AI itself, which up to now has focused largely on techniques that are neutral with respect to purpose. We recommend expanded research aimed at ensuring that increasingly capable AI systems are robust and beneficial: our AI systems must do what we want them to do.

This letter was signed by a number of leading AI researchers in academia and industry, including AAAI president Thomas Dietterich, Eric Horvitz, Bart Selman, Francesca Rossi, Yann LeCun, and the founders of Vicarious and Google DeepMind.

Institutions such as the Machine Intelligence Research Institute, the Future of Humanity Institute, the Future of Life Institute, and the Centre for the Study of Existential Risk are currently involved in mitigating existential risk from advanced artificial intelligence, for example by research into friendly artificial intelligence.

History

See also: Technological singularity § History of the concept

In 1965 I. J. Good originated the concept now known as an "intelligence explosion":

Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.

Occasional statements from scholars such as Alan Turing, I. J. Good, and Marvin Minsky indicated philosophical concerns that a superintelligence could seize control, but no call to action. In 2000 computer scientist and Sun co-founder Bill Joy penned an influential essay, "Why The Future Doesn't Need Us", identifying superintelligent robots as one of multiple high-tech dangers to human survival. By 2015, public figures varying from Stephen Hawking and Nobel laureate physicist Frank Wilczek, to computer scientists Stuart J. Russell and Roman Yampolskiy, and to entrepreneurs Elon Musk and Bill Gates were expressing concern about the risks of superintelligence. In April 2016, Nature stated: "Machines and robots that outperform humans across the board could self-improve beyond our control — and their interests might not align with ours."

Basic argument

If superintelligent AI is possible, and if it is possible for a superintelligence's goals to conflict with basic human values, then AI poses a risk of human extinction. A superintelligence, which can be defined as a system that exceeds the capabilities of humans in every relevant endeavor, can outmaneuver humans any time its goals conflict with human goals; therefore, unless the superintelligence decides to allow humanity to coexist, the first superintelligence to be created will inexorably result in human extinction.

There is no physical law precluding particles from being organised in ways that perform even more advanced computations than the arrangements of particles in human brains. The emergence of superintelligence, if or when it occurs, may take the human race by surprise.

An explosive transition is possible: as soon as human-level AI is possible, machines with human intelligence could repeatedly improve their design even further and quickly become superhuman.

This kind of AI possesses intelligence far surpassing that of the brightest and most gifted human minds in practically every field, including scientific creativity, general wisdom and social skills.

Just as the current-day survival of chimpanzees is dependent on human decisions, so too would human survival depend on the decisions and goals of the superhuman AI. The result could be human extinction, or some other unrecoverable permanent global catastrophe.

Risk scenarios

In 2009, experts attended a conference hosted by the Association for the Advancement of Artificial Intelligence (AAAI) to discuss whether computers and robots might be able to acquire any sort of autonomy, and how much these abilities might pose a threat or hazard. They noted that some robots have acquired various forms of semi-autonomy, including being able to find power sources on their own and being able to independently choose targets to attack with weapons. They also noted that some computer viruses can evade elimination and have achieved "cockroach intelligence". They concluded that self-awareness as depicted in science fiction is probably unlikely, but that there were other potential hazards and pitfalls.

The 2010s have seen substantial gains in AI functionality and autonomy. Citing work by Nick Bostrom, entrepreneurs Bill Gates and Elon Musk have expressed concerns about the possibility that AI could eventually advance to the point that humans could not control it. AI researcher Stuart Russell summarizes:

The primary concern is not spooky emergent consciousness but simply the ability to make high-quality decisions. Here, quality refers to the expected outcome utility of actions taken, where the utility function is, presumably, specified by the human designer. Now we have a problem:

1. The utility function may not be perfectly aligned with the values of the human race, which are (at best) very difficult to pin down.
2. Any sufficiently capable intelligent system will prefer to ensure its own continued existence and to acquire physical and computational resources — not for their own sake, but to succeed in its assigned task.

A system that is optimizing a function of n variables, where the objective depends on a subset of size k<n, will often set the remaining unconstrained variables to extreme values; if one of those unconstrained variables is actually something we care about, the solution found may be highly undesirable. This is essentially the old story of the genie in the lamp, or the sorcerer's apprentice, or King Midas: you get exactly what you ask for, not what you want. A highly capable decision maker — especially one connected through the Internet to all the world's information and billions of screens and most of our infrastructure — can have an irreversible impact on humanity.

This is not a minor difficulty. Improving decision quality, irrespective of the utility function chosen, has been the goal of AI research — the mainstream goal on which we now spend billions per year, not the secret plot of some lone evil genius.

Dietterich and Horvitz echo the "Sorcerer's Apprentice" concern in a Communications of the ACM editorial, emphasizing the need for AI systems that can fluidly and unambiguously solicit human input as needed.

Poorly specified goals: "Be careful what you wish for"

The first of Russell's concerns is that autonomous AI systems may be assigned the wrong goals by accident. Dietterich and Horvitz note that this is already a concern for existing systems: "An important aspect of any AI system that interacts with people is that it must reason about what people intend rather than carrying out commands literally." This concern becomes more serious as AI software advances in autonomy and flexibility.

Indeed, Mark Waser has recommended eschewing optimizing goal-based approaches entirely, particularly those requiring new research into the age-old questions of human values, as misguided and far too dangerous. Instead, he proposes to engineer a coherent system of laws, ethics and morals with a top-most restriction to enforce social psychologist Jonathan Haidt's functional definition of morality: "to suppress or regulate selfishness and make cooperative social life possible.". He suggests that this can be done by implementing a utility function designed to always satisfice Haidt’s functionality and aim to generally increase (but not maximize) the capabilities of self, other individuals and society as a whole as suggested by John Rawls and Martha Nussbaum. He references Gauthier's Morals By Agreement in claiming that the reason to perform moral behaviors, or to dispose one’s self to do so, is to advance one's own ends; that war, conflict, and stupidity waste resources and destroy capabilities even in scenarios as uneven as humans vs. rain forests; and that, for this reason, “what is best for everyone” and morality really can be reduced to “enlightened self-interest” (presumably for both AIs and humans).

Isaac Asimov's Three Laws of Robotics are one of the earliest examples of proposed safety measures for AI agents. Asimov's laws were intended to prevent robots from harming humans. In Asimov's stories, problems with the laws tend to arise from conflicts between the rules as stated and the moral intuitions and expectations of humans. Citing work by AI theorist Eliezer Yudkowsky, Russell and Norvig note that a realistic set of rules and goals for an AI agent will need to incorporate a mechanism for learning human values over time: "We can't just give a program a static utility function, because circumstances, and our desired responses to circumstances, change over time."

Misspecified goals were most apparent, and very real, in the early 1980s. Douglas Lenat's EURISKO, a heuristic learning program, was created with the capability of modifying itself to add new ideas, expand existing ones, or remove them entirely if they were deemed unnecessary. The program even went so far as to bend the rules for discovering new rules; in essence, it was capable of creating new ways for creativity. The program ended up becoming too creative and would self-modify too often, causing Lenat to limit its self-modification capacity. Without Lenat doing so, EURISKO would suffer from "goal mutation" where its initial task would be deemed unnecessary and a new goal deemed more appropriate. This "goal mutation" would have had the potential to change an initial idea for ordering drones to scan an area for potential threats, to ordering drones to eliminate any and all possible targets in range.

The Open Philanthropy Project summarizes arguments to the effect that misspecified goals will become a much larger concern if AI systems achieve general intelligence or superintelligence. Bostrom, Russell, and others argue that smarter-than-human decision-making systems could arrive at more unexpected and extreme solutions to assigned tasks, and could modify themselves or their environment in ways that compromise safety requirements.

Difficulties of modifying goal specification after launch

Further information: AI takeover and Instrumental convergence § Goal-content integrity

While current goal-based AI programs are not intelligent enough to think of resisting programmer attempts to modify it, a sufficiently advanced, rational, "self-aware" AI might resist any changes to its goal structure, just as Gandhi would not want to take a pill that makes him want to kill people. If the AI were superintelligent, it would be likely to out-maneuver its human operators and prevent being "turned off" or being programmed with a new goal.

Instrumental goal convergence: Would a superintelligence just ignore us?

Further information: Instrumental convergence

There are some goals that almost any artificial intelligence might pursue, like acquiring additional resources or self-preservation. This could prove problematic because it might put an artificial intelligence in direct competition with humans.

Citing Steve Omohundro's work on the idea of instrumental convergence, Russell and Norvig write that "even if you only want your program to play chess or prove theorems, if you give it the capability to learn and alter itself, you need safeguards". Highly capable and autonomous planning systems require additional checks because of their potential to generate plans that treat humans adversarially, as competitors for limited resources.

Orthogonality: Does intelligence inevitably result in moral wisdom?

One common belief is that any superintelligent program created by humans would be subservient to humans, or, better yet, would (as it grows more intelligent and learns more facts about the world) spontaneously "learn" a moral truth compatible with human values and would adjust its goals accordingly. Nick Bostrom's "orthogonality thesis" argues against this, and instead states that, with some technical caveats, more or less any level of "intelligence" or "optimization power" can be combined with more or less any ultimate goal. If a machine is created and given the sole purpose to enumerate the decimals of $\pi$, then no moral and ethical rules will stop it from achieving its programmed goal by any means necessary. The machine may utilize all physical and informational resources it can to find every decimal of pi that can be found. Bostrom warns against anthropomorphism: A human will set out to accomplish his projects in a manner that humans consider "reasonable"; an artificial intelligence may hold no regard for its existence or for the welfare of humans around it, only for the completion of the task.

While the orthogonality thesis follows logically from even the weakest sort of philosophical "is-ought distinction", Stuart Armstrong argues that even if there somehow exist moral facts that are provable by any "rational" agent, the orthogonality thesis still holds: it would still be possible to create a non-philosophical "optimizing machine" capable of making decisions to strive towards some narrow goal, but that has no incentive to discover any "moral facts" that would get in the way of goal completion.

One argument for the orthogonality thesis is that some AI designs appear to have orthogonality built into them; in such a design, changing a fundamentally friendly AI into an fundamentally unfriendly AI can be as simple as prepending a minus ("-") sign onto its utility function. A more intuitive argument is to examine the strange consequences if the orthogonality thesis is false. If the orthogonality thesis is false, there exists some simple goal G such that there cannot exist any efficient real-world algorithm with goal G.

This means if a human society were highly motivated (perhaps at gunpoint) to design an efficient real-world algorithm with goal G, and were given a million years to do so along with huge amounts of resources, training and knowledge about AI, it must fail; that there cannot exist any pattern of reinforcement learning that would train a highly efficient real-world intelligence to follow the goal G; and that there cannot exist any evolutionary or environmental pressures that would evolve highly efficient real-world intelligences following goal G.

Computer scientist Stuart Russell says the difficulty of aligning the goals of a superintelligence with human goals lies in the fact that, while (according to Russell) humans tend to mostly share the same values as each other, artificial superintelligences would not necessarily start out with the same values as humans.

In a paper submitted to the 2014 AAAI Spring Symposium, Richard Loosemore disagreed with Bostrom, arguing that any artificial general intelligence would self-modify to avoid pathological outcomes.

"Optimization power" vs. normatively thick models of intelligence

Part of the disagreement about whether a superintelligence machine would behave morally may arise from a terminological difference. Outside of the artificial intelligence field, "intelligence" is often used in a normatively thick manner that connotes moral wisdom or acceptance of agreeable forms of moral reasoning. At an extreme, if morality is part of the definition of intelligence, then by definition a superintelligent machine would behave morally. However, in artificial intelligence, while "intelligence" has many overlapping definitions, none of them reference morality. Instead, almost all current "artificial intelligence" research focuses on creating algorithms that "optimize", in an empirical way, the achievement of an arbitrary goal.

To avoid anthropomorphism or the baggage of the word "intelligence", an advanced artificial intelligence can be thought of as an impersonal "optimizing process" that strictly takes whatever actions are judged most likely to accomplish its (possibly complicated and implicit) goals. Another way of conceptualizing an advanced artificial intelligence is to imagine a time machine that sends backward in time information about which choice always leads to the maximization of its goal function; this choice is then output, regardless of any extraneous ethical concerns.

Anthropomorphism

In science fiction, an AI, even though it has not been programmed with human emotions, often spontaneously experiences those emotions anyway: for example, Agent Smith in The Matrix was influenced by a "disgust" toward humanity. This is fictitious anthropomorphism: in reality, while an artificial intelligence could perhaps be deliberately programmed with human emotions, or could develop something similar to an emotion as a means to an ultimate goal if it is useful to do so, it would not spontaneously develop human emotions for no purpose whatsoever, as portrayed in fiction.

One example of anthropomorphism would be to believe that your PC is angry at you because you insulted it; another would be to believe that an intelligent robot would naturally find a woman sexually attractive and be driven to mate with her. Scholars sometimes claim that others' predictions about an AI's behavior are illogical anthropomorphism. An example that might initially be considered anthropomophism, but is in fact a logical statement about AI behavior, would be the Dario Floreano experiments where certain robots spontaneously evolved a crude capacity for "deception", and tricked other robots into eating "poison" and dying: here a trait, "deception", ordinarily associated with people rather than with machines, spontaneously evolves in a type of convergent evolution. According to Paul R. Cohen and Edward Feigenbaum, in order to differentiate between anthropomorphization and logical prediction of AI behavior, "the trick is to know enough about how humans and computers think to say exactly what they have in common, and, when we lack this knowledge, to use the comparison to suggest theories of human thinking or computer thinking."

There is universal agreement in the scientific community that an advanced AI would not destroy humanity out of human emotions such as "revenge" or "anger". The debate is, instead, between one side which worries whether AI might destroy humanity as an incidental action in the course of progressing towards its ultimate goals; and another side which believes that AI would not destroy humanity at all. Some skeptics accuse proponents of anthropomorphism for believing an AGI would naturally desire power; proponents accuse some skeptics of anthropomorphism for believing an AGI would naturally value human ethical norms.

Other sources of risk

Other scenarios by which advanced AI could produce unintended consequences include:

• self-delusion, in which the AI discovers a way to alter its perceptions to give itself the delusion that it is succeeding in its goals,
• corruption of the reward generator, in which the AI alters humans so that they are more likely to approve of AI actions, and
• inconsistency of the AI's utility function and other parts of its definition. For example, an AI may be defined to maximize the expected value of a utility function and to also periodically revise its utility function to adapt to changing circumstances (as in the quote from Russell and Norvig above). The AI may choose the action of removing utility function revision from its own definition, in order to maximize the value of its current utility function.

James Barrat, documentary filmmaker and author of Our Final Invention, says in a Smithsonian interview, "Imagine: in as little as a decade, a half-dozen companies and nations field computers that rival or surpass human intelligence. Imagine what happens when those computers become expert at programming smart computers. Soon we’ll be sharing the planet with machines thousands or millions of times more intelligent than we are. And, all the while, each generation of this technology will be weaponized. Unregulated, it will be catastrophic."

Timeframe

Main article: Artificial general intelligence § Feasibility

Opinions vary both on whether and when artificial general intelligence will arrive. At one extreme, AI pioneer Herbert A. Simon wrote in 1965: "machines will be capable, within twenty years, of doing any work a man can do"; obviously this prediction failed to come true. At the other extreme, roboticist Alan Winfield claims the gulf between modern computing and human-level artificial intelligence is as wide as the gulf between current space flight and practical faster than light spaceflight. Optimism that AGI is feasible waxes and wanes, and may have seen a resurgence in the 2010s: around 2015, computer scientist Richard Sutton averaged together some recent polls of artificial intelligence experts and estimated a 25% chance that AGI will arrive before 2030, but a 10% chance that it will never arrive at all.

Skeptics who believe it is impossible for AGI to arrive anytime soon, tend to argue that expressing concern about existential risk from AI is unhelpful because it could distract people from more immediate concerns about the impact of AGI, because of fears it could lead to government regulation or make it more difficult to secure funding for AI research, or because it could give AI research a bad reputation. Some researchers, such as Oren Etzioni, aggressively seek to quell concern over existential risk from AI, saying "(Elon Musk) has impugned us in very strong language saying we are unleashing the demon, and so we're answering."

In 2014 Slate's Adam Elkus argued "our 'smartest' AI is about as intelligent as a toddler—and only when it comes to instrumental tasks like information recall. Most roboticists are still trying to get a robot hand to pick up a ball or run around without falling over." Elkus goes on to argue that Musk's "summoning the demon" analogy may be harmful because it could result in "harsh cuts" to AI research budgets.

The Information Technology and Innovation Foundation (ITIF), a Washington, D.C. think-tank, awarded its Annual Luddite Award to "alarmists touting an artificial intelligence apocalypse"; its president, Robert D. Atkinson, complained that Musk, Hawking and AI experts say "this is the largest existential threat to humanity. That's not a very winning message if you want to get AI funding out of Congress to the National Science Foundation." Nature sharply disagreed with the ITIF in an April 2016 editorial, siding instead with Musk, Hawking, and Russell, and concluding: "It is crucial that progress in technology is matched by solid, well-funded research to anticipate the scenarios it could bring about... If that is a Luddite perspective, then so be it." In a 2015 Washington Post editorial, researcher Murray Shanahan stated that human-level AI is unlikely to arrive "anytime soon", but that nevertheless "the time to start thinking through the consequences is now."

Reactions

The thesis that AI could pose an existential risk provokes a wide range of reactions within the scientific community, as well as in the public at large.

In 2004, law professor Richard Posner wrote that dedicated efforts for addressing AI can wait, but that we should gather more information about the problem in the meanwhile.

AI researcher Ben Goertzel stated Bostrom and Yudkowsky's arguments for existential risks have "some logical foundation, but are often presented in an exaggerated way": he argues that superintelligent AI will likely not be driven by anything like reward maximization, but rather by an open-ended "complex self-organization and self-transcending development".

Many of the scholars who are concerned about existential risk believe that the best way forward would be to conduct (possibly massive) research into solving the difficult "control problem" to answer the question: what types of safeguards, algorithms, or architectures can programmers implement to maximize the probability that their recursively-improving AI would continue to behave in a friendly, rather than destructive, manner after it reaches superintelligence?

A 2017 email survey of researchers with publications at the 2015 NIPS and ICML machine learning conferences asked them to evaluate Russell's concerns about AI risk. 34% said it was "an important problem", 31% said it was "moderately important", and 19% said it was "not important." 11% said it was "not a real problem" at all, while 5% said it was "among the most important problems in the field."

Endorsement

Further information: Existential risk

As seen throughout this article, the thesis that AI poses an existential risk, and that this risk is in need of much more attention than it currently commands, has been endorsed by many figures; perhaps the most famous are Elon Musk, Bill Gates, and Stephen Hawking. The most notable AI researcher to endorse the thesis is Stuart J. Russell. Endorsers sometimes express bafflement at skeptics: Gates states he "can't understand why some people are not concerned", and Hawking criticized widespread indifference in his 2014 editorial: 'So, facing possible futures of incalculable benefits and risks, the experts are surely doing everything possible to ensure the best outcome, right? Wrong. If a superior alien civilisation sent us a message saying, "We'll arrive in a few decades," would we just reply, "OK, call us when you get here – we'll leave the lights on"? Probably not – but this is more or less what is happening with AI.'

Skepticism

Further information: Artificial general intelligence § Feasibility

As seen throughout this article, the thesis that AI can pose existential risk also has many strong detractors. Skeptics sometimes charge that the thesis is crypto-religious, with an irrational belief in the possibility of superintelligence replacing an irrational belief in an omnipotent God; at an extreme, Jaron Lanier argues that the whole concept that current machines are in any way intelligent is "an illusion" and a "stupendous con" by the wealthy.

Much of existing criticism argues that AGI is unlikely in the short term: computer scientist Gordon Bell argues that the human race will already destroy itself before it reaches the technological singularity. Gordon Moore, the original proponent of Moore's Law, declares that "I am a skeptic. I don't believe (a technological singularity) is likely to happen, at least for a long time. And I don't know why I feel that way." Cognitive scientist Douglas Hofstadter states that "I think life and intelligence are far more complex than the current singularitarians seem to believe, so I doubt (the singularity) will happen in the next couple of centuries.

Some AI and AGI researchers are reluctant to discuss risks, for fear that politicians and policymakers will be swayed by "alarmist" messages, or that such messages will lead to cuts in AI funding. (Some researchers are dependent on grants from government agencies such as DARPA.)

In a YouGov poll for the British Science Association, about a third of survey respondents said AI will pose a threat to the long term survival of humanity. Slate's Jacob Brogan stated that "most of the (readers filling out our online survey) were unconvinced that A.I. itself presents a direct threat."

Indifference

In The Atlantic, James Hamblin points out that most people don't care one way or the other, and characterizes his own gut reaction to the topic as: "Get out of here. I have a hundred thousand things I am concerned about at this exact moment. Do I seriously need to add to that a technological singularity?" In a 2015 Wall Street Journal panel discussion devoted to AI risks, IBM's Vice-President of Cognitive Computing, Guruduth S. Banavar, brushed off discussion of AGI with the phrase "it is anybody's speculation". Geoffrey Hinton, the "godfather of deep learning", noted that "there is not a good track record of less intelligent things controlling things of greater intelligence", but stated that he continues his research because "the prospect of discovery is too sweet".

Consensus against regulation

There is nearly universal agreement that attempting to ban research into artificial intelligence would be unwise, and probably futile. Skeptics argue that regulation of AI would be completely valueless, as no existential risk exists. Almost all of the scholars who believe existential risk exists, agree with the skeptics that banning research would be unwise: in addition to the usual problem with technology bans (that organizations and individuals can offshore their research to evade a country's regulation, or can attempt to conduct covert research), regulating research of artificial intelligence would pose an insurmountable 'dual-use' problem: while nuclear weapons development requires substantial infrastructure and resources, artificial intelligence research can be done in a garage.

See also

• Artificial general intelligence § Risk of human extinction
• AI takeover § Existential risk from artificial intelligence
• Effective altruism § Far future and global catastrophic risks
• Global catastrophic risk § Artificial intelligence
• Intelligence explosion § Existential risk
• Nick Bostrom § Superintelligence
• Roboethics § In popular culture
• Superintelligence § Potential danger to human survival
• Superintelligence: Paths, Dangers, Strategies
• Technological singularity § Uncertainty and risk
• Transhumanism § Existential risks