Concerns have recently been widely expressed that artificial intelligence presents a threat to humanity. For instance, Stephen Hawking is quoted in Cellan-Jones1 as saying: "The development of full artificial intelligence could spell the end of the human race." Similar concerns have also been expressed by Elon Musk, Steve Wozniak, and others.
Such concerns have a long history. John von Neumann is quoted by Stanislaw Ulam8 as the first to use the term the singularityathe point at which artificial intelligence exceeds human intelligence. Ray Kurzweil5 has predicted that the singularity will occur around 2045a prediction based on Moore's Law as the time when machine speed and memory capacity will rival human capacity. I.J. Good has predicted that such super-intelligent machines will then build even more intelligent machines in an accelerating 'intelligence explosion.'4 The fear is that these super-intelligent machines will pose an existential threat to humanity, for example, keep humans as pets or kill us all10or maybe humanity will just be a victim of evolution. (For additional information, see Dubhashi and Lappin's argument on page 39.)
I think the concept of the singularity is ill conceived. It is based on an oversimplified and false understanding of intelligence. Moore's Law will not inevitably lead to such a singularity. Progress in AI depends not just on speed and memory size, but also developing new algorithms and the new concepts that underpin them. More crucially, the singularity is predicated on a linear model of intelligence, rather like IQ, on which each animal species has its place, and along which AI is gradually advancing. Intelligence is not like this. As Aaron Sloman, for instance, has successfully argued, intelligence must be modeled using a multidimensional space, with many different kinds of intelligence and with AI progressing in many different directions.6
AI systems occupy points in this multidimensional space that are unlike any animal species. In particular, their expertise tends to be very high in very narrow areas, but nonexistent else-where. Consider, for instance, some of the most successful AI systems of the last few decades.
Is this situation likely to change in the foreseeable future? There is currently a revival of interest in AI general intelligence, the attempt to build a machine that could successfully perform any intellectual task that a human being can. Is there any reason to believe that progress now will be faster than it has been since John McCarthy advocated it more than 60 years ago at the 1956 inaugural AI conference at Dartmouth? It is generally agreed that one of the key enabling technologies will be commonsense reasoning. A recent Communications article2 argues that, while significant progress has been made in several areas of reasoning: temporal, geometric, multi-agent, and so forth, many intractable problems remain. Note also that, while successful systems, such as Watson and AlphaGo, have been applied to new areas, each of these applications is still narrow in scope. One could use a 'Big Switch' approach, to direct each task to the appropriate narrowly scoped system, but this approach is generally regarded as inadequate in not providing the integration of multiple cognitive processes routinely employed by humans.
Many humans tend to ascribe too much intelligence to narrowly focused AI systems.
I am not attempting to argue that AI general intelligence is, in principle, impossible. I do not believe there is anything in human cognition that is beyond scientific understanding. With such an understanding will surely come the ability to emulate it artificially. But I am not holding my breath. I have lived through too many AI hype cycles to expect the latest one to deliver something that previous cycles have failed to deliver. And I do not believe that now is the time to worry about a threat to humanity from smart machines, when there is a much more pressing problem to worry about.
That problem is that many humans tend to ascribe too much intelligence to narrowly focused AI systems. Any machine that can beat all humans at Go must surely be very intelligent, so by analogy with other world-class Go players, it must be pretty smart in other ways too, mustn't it? No! Such misconceptions lead to false expectations that such AI systems will work correctly in areas outside their narrow expertise. This can cause problems, for example, a medical diagnosis system might recommend the wrong treatment when faced with a disease beyond its diagnostic ability, a self-driving car has already crashed when confronted by an unanticipated situation. Such erroneous behavior by dumb machines certainly presents a threat to individual humans, but not to humanity. To counter it, AI systems need an internal model of their scope and limitations, so that they can recognize when they are straying outside their comfort zone and warn their human users that they need human assistance or just should not be used in such a situation. We must assign a duty to AI system designers to ensure their creations inform users of their limitations, and specifically warn users when they are asked to operate out of their scope. AI systems must have the ability to explain their reasoning in a way that users can understand and assent to. Because of their open-ended behavior, AI systems are also inherently hard to verify. We must develop software engineering techniques to address this. Since AI systems are increasingly self-improving, we must ensure these explanations, warnings, and verifications keep pace with each AI system's evolving capabilities.
The concerns of Hawkings and others were addressed in an earlier Communications Viewpoint by Dietterich and Horvitz.3 While downplaying these concerns, Dietterich and Horvitz also categorize the kinds of threats that AI technology does pose. This apparent paradox can be resolved by observing that the various threats they identify are caused by AI technology being too dumb, not too smart.
AI systems are, of course, by no means unique in having bugs or limited expertise. Any computer system deployed in a safety or security critical situation potentially poses a threat to health, privacy, finance, and other realms. That is why our field is so concerned about program correctness and the adoption of best software engineering practice. What is different about AI systems is that some people may have unrealistic expectations about the scope of their expertise, simply because they exhibit intelligencealbeit in a narrow domain.
The current focus on the very remote threat of super-human intelligence is obscuring this very real threat from subhuman intelligence.
But could such dumb machines be sufficiently dangerous to pose a threat to humanity? Yes, if, for instance, we were stupid enough to allow a dumb machine the autonomy to unleash weapons of mass destruction. We came close to such stupidity with Ronald Reagan and Edward Teller's 1983 proposal of a Strategic Defense Initiative (SDI, aka 'Star Wars').b Satellite-based sensors would detect a Soviet ballistic missile launch and super-powered x-ray lasers would zap these missiles from space before they got into orbit. Since this would need to be accomplished within seconds, no human could be in the loop. I was among many computer scientists who successfully argued that the most likely outcome was a false positive that would trigger the nuclear war it was designed to prevent. There were precedents from missile early-warning systems that had been triggered by, among other things, a moonrise and a flock of geese. Fortunately, in these systems a human was in the loop to abort any unwarranted retaliation to the falsely suspected attack. A group of us from Edinburgh met U.K. Ministry of Defence scientists, engaged with SDI, who admitted they shared our analysis. The SDI was subsequently quietly dropped by morphing it into a saner program. This is an excellent example of non-computer scientists overestimating the abilities of dumb machines. One can only hope that, like the U.K.'s MOD scientists, the developers of such weapon systems have learned the institutional lesson from this fiasco. We all also need to publicize these lessons to ensure they are widely understood. Similar problems arise in other areas too, for example, the 2010 flash crash demonstrated how vulnerable society was to the collapse of a financial system run by secret, competing and super-fast autonomous agents.
Another potential existential threat is that AI systems may automate most forms of human employment.7,9 If my analysis is correct then, for the foreseeable future, this automation will develop as a coalition of systems, each of which will automate only a narrowly defined task. It will be necessary for these systems to work collaboratively, with humans: orchestrating the coalition, recognizing when a system is out of its depth and dealing with these 'edge cases' interactively. The productively of human workers will be, thereby, dramatically increased and the cost of the service provided by this multiagent approach will be dramatically reduced, perhaps leading to an increase in the services provided. Whether this will provide both job satisfaction and a living income to all humans can currently only be an open question. It is up to us to invent the future in which it will do, and to ensure this future is maintained as the capability and scope of AI systems increases. I do not underestimate the difficulty of achieving this. The challenges are more political and social than technical, so this is a job for the whole of society.
As AI progresses, we will see even more applications that are super-intelligent in a narrow area and incredibly dumb everywhere else. The areas of successful application will get gradually wider and the areas of dumbness narrower, but not disappear. I believe this will remain true even when we do have a deep understanding of human cognition. Maggie Boden has a nice analogy with flight. We do now understand how birds fly. In principle, we could build ever more accurate simulations of a bird, but this would incur an increasingly exorbitant cost and we already achieve satisfactory human flight by alternative means: airplanes, helicopters, paragliders, and so forth. Similarly, we will develop a zoo of highly diverse AI machines, each with a level of intelligence appropriate to its tasknot a new uniform race of general-purpose, super-intelligent, humanity supplanters.
1. Cellan-Jones, R. Stephen Hawking warns artificial intelligence could end mankind. BBC Interview, (Dec. 2014); http://www.bbc.co.uk/news/technology-30290540.
Thanks to Stephan Schulz, Lucas Dixon, the St. Andrews University Student Debating Society, and two anonymous reviewers for feedback on earlier versions of this Viewpoint.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2017 ACM, Inc.
The following letter was published in the Letters to the Editor in the March 2017 CACM (http://cacm.acm.org/magazines/2017/3/213824).
The Viewpoints by Alan Bundy "Smart Machines Are Not a Threat to Humanity" and Devdatt Dubhashi and Shalom Lappin "AI Dangers: Imagined and Real" (both Feb. 2017) argued against the possibility of a near-term singularity wherein super-intelligent AIs exceed human capabilities and control. Both relied heavily on the lack of direct relevance of Moore's Law, noting raw computing power does not by itself lead to human-like intelligence. Bundy also emphasized the difference between a computer's efficiency in working an algorithm to solve a narrow, well-defined problem and human-like generalized problem-solving ability. Dubhashi and Lappin noted incremental progress in machine learning or better knowledge of a biological brain's wiring do not automatically lead to the "unanticipated spurts" of progress that characterize scientific breakthroughs.
These points are valid, but a more accurate characterization of the situation is that computer science may well be just one conceptual breakthrough away from being able to build an artificial general intelligence. The considerable progress already made in computing power, sensors, robotics, algorithms, and knowledge about biological systems will be brought to bear quickly once the architecture of "human-like" general intelligence is articulated. Will that be tomorrow or in 10 years? No one knows. But unless there is something about the architecture of human intelligence that is ultimately inaccessible to science, that architecture will be discovered. Study of the consequences is not premature.
Displaying 1 comment