Artificial intelligence (also known as machine intelligence and often abbreviated as AI) is intelligence exhibited by any manufactured (i.e. artificial) system. The term is often applied to general purpose computers and also in the field of scientific investigation into the theory and practical application of AI.
One popular and early definition of artificial intelligence research, put forth by John McCarthy at the Dartmouth Conference in 1956, is "making a machine behave in ways that would be called intelligent if a human were so behaving", repeating the claim put forth by Alan Turing in "Computing machinery and intelligence" (Mind, October 1950). However, this definition seems to ignore the possibility of strong AI. Another definition of artificial intelligence is "intelligence arising from an artificial device". Most definitions could be categorized as concerning either systems that think like humans, systems that act like humans, systems that think rationally or systems that act rationally.
Artificial intelligence can be considered in two parts: "What is the nature of artifice?" and "What is intelligence?" The first question is relatively easy, although it leads to an examination of what can be manufactured. The limitations of classical computational systems, available manufacturing processes, or human intellect may all place constraints on what can be manufactured.
The second question raises ontological issues of consciousness and intelligence as displayed by humans, as intelligent behavior in humans is complex and often difficult to understand. Study of animals and artificial systems which are not simply models of what already exists are also considered highly relevant.
Development of AI theory
Much of the (original) focus of artificial intelligence research draws from an experimental approach to psychology, and emphasizes what may be called linguistic intelligence (best exemplified in the Turing test).
Approaches to artificial intelligence that do not focus on linguistic intelligence include robotics and collective intelligence approaches, which focus on active manipulation of an environment, or consensus decision making, and draw from biology and political science when seeking models of how "intelligent" behavior is organized.
Artificial intelligence theory also draws from animal studies, in particular with insects, which are easier to emulate as robots (see artificial life), as well as animals with more complex cognition, including apes, who resemble humans in many ways but have less developed capacities for planning and cognition. AI researchers argue that animals, which are simpler than humans, ought to be considerably easier to mimic. But satisfactory computational models for animal intelligence are not available.
Seminal papers advancing the concept of machine intelligence include A Logical Calculus of the Ideas Immanent in Nervous Activity (1943), by Warren McCulloch and Walter Pitts, and On Computing Machinery and Intelligence (1950), by Alan Turing, and Man-Computer Symbiosis by J.C.R. Licklider. See cybernetics and Turing test for further discussion.
There were also early papers which denied the possibility of machine intelligence on logical or philosophical grounds such as Minds, Machines and Gödel (1961) by John Lucas .
With the development of practical techniques based on AI research, advocates of AI have argued that opponents of AI have repeatedly changed their position on tasks such as computer chess or speech recognition that were previously regarded as "intelligent" in order to deny the accomplishments of AI. They point out that this moving of the goalposts effectively defines "intelligence" as "whatever humans can do that machines cannot".
John von Neumann (quoted by E.T. Jaynes) anticipated this in 1948 by saying, in response to a comment at a lecture that it was impossible for a machine to think: "You insist that there is something a machine cannot do. If you will tell me precisely what it is that a machine cannot do, then I can always make a machine which will do just that!". Von Neumann was presumably alluding to the Church-Turing thesis which states that any effective procedure can be simulated by a (generalized) computer.
In 1969 McCarthy and Hayes started the discussion about the frame problem with their essay, "Some Philosophical Problems from the Standpoint of Artificial Intelligence".
Experimental AI research
Artificial intelligence began as an experimental field in the 1950s with such pioneers as Allen Newell and Herbert Simon, who founded the first artificial intelligence laboratory at Carnegie-Mellon University, and McCarthy and Marvin Minsky, who founded the MIT AI Lab in 1959. They all attended the aforementioned Dartmouth College summer AI conference in 1956, which was organized by McCarthy, Minsky, Nathan Rochester of IBM and Claude Shannon.
Historically, there are two broad styles of AI research - the "neats" and "scruffies". "Neat", classical or symbolic AI research, in general, involves symbolic manipulation of abstract concepts, and is the methodology used in most expert systems. Parallel to this are the "scruffy", or "connectionist", approaches, of which neural networks are the best-known example, which try to "evolve" intelligence through building systems and then improving them through some automatic process rather than systematically designing something to complete the task. Both approaches appeared very early in AI history. Throughout the 1960s and 1970s scruffy approaches were pushed to the background, but interest was regained in the 1980s when the limitations of the "neat" approaches of the time became clearer. However, it has become clear that contemporary methods using both broad approaches have severe limitations.
Artificial intelligence research was very heavily funded in the 1980s by the Defense Advanced Research Projects Agency in the United States and by the fifth generation computer systems project in Japan. The failure of the work funded at the time to produce immediate results, despite the grandiose promises of some AI practitioners, led to correspondingly large cutbacks in funding by government agencies in the late 1980s, leading to a general downturn in activity in the field known as AI winter . Over the following decade, many AI researchers moved into related areas with more modest goals such as machine learning, robotics, and computer vision, though research in pure AI continued at reduced levels.
Weak artificial intelligence
Weak artificial intelligence research deals with the creation of some form of computer-based artificial intelligence that can reason and solve problems only in a limited domain; such a machine would, in some ways, act as if it were intelligent, but it would not possess true intelligence or sentience. The classical test for such abilities is the Turing test.
There are several fields of weak AI, one of which is natural language. Many weak AI fields have specialised software or programming languages created for them. For example, the 'most-human' natural language chatterbot A.L.I.C.E. uses a programming language AIML that is specific to its program, and the various clones, named Alicebots. Jabberwacky is a little closer to strong AI, since it learns how to converse from the ground up based solely on user interactions.
To date, much of the work in this field has been done with computer simulations of intelligence based on predefined sets of rules. Very little progress has been made in strong AI. Depending on how one defines one's goals, a moderate amount of progress has been made in weak AI.
When viewed with a moderate dose of cynicism, weak artificial intelligence can be viewed as ‘the set of computer science problems without good solutions at this point.’ Once a sub-discipline results in useful work, it is carved out of artificial intelligence and given its own name. Examples of this are pattern recognition, image processing, neural networks, natural language processing, robotics and game theory. While the roots of each of these disciplines is firmly established as having been part of artificial intelligence, they are now thought of as somewhat separate.
Whilst progress towards the ultimate goal of human-like intelligence has been slow, many spinoffs have come in the process. Notable examples include the languages LISP and Prolog, which were invented for AI research but are now used for non-AI tasks. Hacker culture first sprang from AI laboratories, in particular the MIT AI Lab, home at various times to such luminaries as McCarthy, Minsky, Seymour Papert (who developed Logo there), Terry Winograd (who abandoned AI after developing SHRDLU).
Many other useful systems have been built using technologies that at least once were active areas of AI research. Some examples include:
Chinook was declared the Man-Machine World Champion in checkers (draughts) in 1994.
Deep Blue, a chess-playing computer, beat Garry Kasparov in a famous match in 1997.
InfoTame, a text analysis search engine developed by the KGB for automatically sorting millions of pages of communications intercepts.
Fuzzy logic, a technique for reasoning under uncertainty, has been widely used in industrial control systems.
Expert systems are being used to some extent industrially.
Machine translation systems such as SYSTRAN are widely used, although results are not yet comparable with human translators.
Neural networks have been used for a wide variety of tasks, from intrusion detection systems to computer games.
Optical character recognition systems can translate arbitrary typewritten European script into text.
Handwriting recognition is used in millions of personal digital assistants.
Speech recognition is commercially available and is widely deployed.
Computer algebra systems, such as Mathematica and Macsyma, are commonplace.
Machine vision systems are used in many industrial applications ranging from hardware verification to security systems.
- AI Planning methods were used to automatically plan the deployment of US forces during Gulf War I. This task would have cost months of time and millions of dollars to perform manually, and DARPA stated that the money saved on this single application was more than their total expenditure on AI research over the last 50 years.
The vision of artificial intelligence replacing human professional judgment has arisen many times in the history of the field, in science fiction and today in some specialized areas where "expert systems" are used to augment or to replace professional judgment in some areas of engineering and of medicine.
Strong artificial intelligence
Main article: Strong AI
Strong AI is a term introduced by John Searle in 1980 in his article "Minds, Brains, and Programs". Strong AI research deals with the creation of AI where "the appropriately programmed computer really is a mind" (Searle 1980). Whether or not building strong AI is possible is an area of active philosophical debate, as well as whether strong AI can be made truly conscious. Strong AI is frequently the subject of science fiction stories and plays a major role in futurist speculations. Creating strong AI has long been the goal of many AI projects, but at present none have succeeded. It should be emphasised that the terms 'Weak' and 'Strong' do not mean that processors with Weak AI are necessarily less powerful than those with Strong AI.
Sub-fields of AI research
GOFAI - 'Good Old Fashioned AI'
Artificial Life and Evolution
Modern Bayesian methods and learning
AI in Business
According to Haag, Cummings, etc.(2004) there are four common applications of Artificial Intelligence in the business setting:
- Expert Systems
- Neural Networks
- Genetic Algorithms
- Intelligent Agents
Expert Systems apply reasoning capabilities to reach a conclusion. An expert system can process large amounts of known information and provide conclusions based on them.
Neural Networks are AI that are capable of finding and differentiating between patterns. Police Departments use neural networks to identify corruption.
Genetic Algorithms are designed to apply the survival of the fittest process to generate increasingly better solutions to the problem. Investment brokers use Genetic Algorithms to create the best possible combination of investment opportunities for their clients.
An Intelligence Agent is software that assits you, or acts on your behalf, in performing repetitive computer-related tasks. Examples of its uses are data mining programs and monitoring and surveillance agents.
Logic programming was sometimes considered a field of artificial intelligence, but this is no longer the case.
Machines displaying some degree of intelligence
There are many examples of programs displaying some degree of intelligence. Some of these are:
Twenty Questions - A neural-net based game of 20 questions
The Start Project - a web-based system which answers questions in English.
Brainboost - another question-answering system
Cyc, a knowledge base with vast collection of facts about the real world and logical reasoning ability.
Jabberwacky, a learning chatterbot
ALICE, a chatterbot
Alan, another chatterbot
Albert One, multi-faceted chatterbot
ELIZA, a program which pretends to be a psychotherapist, developed in 1966
- PAM (Plan Applier Mechanism) - a story understanding system developed by John Wilensky in 1978.
- SAM (Script applier mechanism) - a story understanding system, developed in 1975.
SHRDLU - an early natural language understanding computer program developed in 1968-1970.
Creatures, a computer game with breeding, evolving creatures coded from the genetic level upwards using a sophisticated biochemistry and neural network brains.
BBC news story on the creator of Creatures latest creation. Steve Grand's Lucy.
AARON - artificial intelligence, which creates its own original paintings, developed by Harold Cohen.
Eurisko - a language for solving problems which consists of heuristics, including heuristics for how to use and change its heuristics. Developed in 1978 by Douglas Lenat.
X-Ray Vision for Surgeons - a group in MIT which researches medical vision.
Neural networks-based programs for backgammon and go.
There are many thousands of AI researchers (see Category:Artificial_intelligence_researchers) around the world at hundreds of research institutions and companies. Among the many who have made significant contributions are:
To some computer scientists, the phrase artificial intelligence has acquired somewhat of a bad name due to the large discrepancy between what has been achieved so far in the field and some more usual notions of intelligence. This problem has been aggravated by various popular science writers and media personalities such as Kevin Warwick whose work has raised the expectations of AI research far beyond its current capabilities. For this reason, some researchers working on topics related to artificial intelligence say they work in cognitive science, informatics, statistical inference or information engineering. However, progress has in fact been made, and AI is today routinely employed in thousands of industrial systems around the world. See Raj Reddy's AAAI paper for a huge review of real-world AI systems in deployment today.
See also Important publications in artificial intelligence.
- John McCarthy: Proposal for the Dartmouth Summer Research Project On Artificial Intelligence. 
- John Searle: Minds, Brains and Programs Behavioral and Brain Sciences 3 (3): 417-457 1980. []
AI related organizations