Chapter 7
Artificial Intelligence
Intelligence is by essence unintelligible.
David Hume (1711-1776)
Although we often believe that the philosopher is right, we cannot deny that scientific progress, in particular in biology, has permitted us to cast a light on the following problems: studying intelligence, its partial simulation on machines, what can be formalized (with the current state of formal tools), appreciating its limits, better identify where the problems are important, and so on.
Furthermore, note that from a practical point of view, we need more and more intelligent tools (whatever the reasonable and perhaps informal characterization of “intelligence” we adopt).
7.1. Intelligent systems: AI
To begin to grasp the topic, we recall the etymology of (natural) intelligence1 :
in French:
Intelligence: 12th Century “understanding”, 15th Century “communication between people who understand each other”.
in English:
Intelligent: 16th Century, from the Latin words intellegere, ligere lit. choose among, formed on INTER + legere gather, choose.
In other words, choose among (which implicitly seems to admit the importance of handling non-determinism).
As far as its study and modeling is concerned, for a long time, mathematics has been used to study problems that arise in natural science, in particular on the behavior of the brain and the nervous system. Conversely, important developments in different topics of mathematics have been motivated by these problems. All this happened before there was even a research domain called AI.
DIGRESSION 7.1.– Ramón Llull (c. 1235-1315) and G. Leibniz (1646-1716) are often cited as pioneers of the “AI project”, i.e. of the belief that every thought (the essential human characteristic) can be formalized. This project became more concrete, at least partially, with the development of computers, and its huge influence on science and society in general.
Ramón Llull designed a reasoning machine (more precisely a deduction machine) with which he wanted to encode (in a combinatorial way) knowledge of creation in a universal language combining base symbols.
G. Leibniz believed that if we could produce a list of basic human thoughts (i.e. words denoting simple ideas), it would be possible to produce mechanically (by combination) all complex ideas. (This should be compared with the concept of a formal system.)
This reductionism seems to have a very long history. some religions (in their esoteric approach) postulated on the existence of absolute ideas and of a mathematic (algebra) of ideas.
The models that were used were constructed based on neural models. some chemical, electrical, and mechanical aspects were taken into account and differential equations were used as tools of modeling and analysis.
This tradition continues in a part of current AI, another larger part inherited from the discrete modeling approach (two states, 0 and 1), which originated with the work of McCulloch and Pitts, Turing, and others.
Of course, the study of other models, thanks to other tools, is possible (and likely) in the future.
Recently, some research has focussed on the cerebral bases of mathematical activity, for example, on the cerebral activity that corresponds to the understanding of numbers and calculus.
The importance of these studies is reflected in the creation at the Collège de France of a chair of “experimental cognitive psychology”.
How can AI be defined? For example, consider the following definition by Minsky, one of the pioneers of the domain.
Artificial Intelligence is the science of making machines do things that would require intelligence if done by men.
Is this definition really satisfactory? Consider motion for example, as well as the laws that characterize it. Motion can be defined independently from the objects that move…but intelligence has probably never been defined independently from the beings that we consider as intelligent.
If in Minsky’s definition we replace “intelligence” by “kinematics”:
Kinematics is the science of making machines do things that would require motion [if done by cars].
[]: redundant here but not in the previous definition.
If we say that intelligence (or that intelligent behavior) is an exclusive property of living beings (or of the animal kingdom), the problem is somehow solved: a computer system cannot exhibit an intelligent behavior, except if we believe a computer can be alive. We have then replaced the problem of defining intelligence by the problem of defining living beings.
At least since cybernetics (see section 7.5.2), connections have been made with life (in particular, with the way the brain works): self-reproducing automata (artificial), neural networks, etc.
REMARK 7.1.– The following definition is sometimes given to artificial life (a domain related to AI):
Artificial life denotes the study of artificial systems that exhibit a behavior that is characteristic to natural living systems.
This definition is similar to Minsky’s on (which would be a particular instance of artificial life), and in our opinion, is not satisfactory either.
To be an interesting definition, it presupposes that there is a decidable way of qualifying all the characteristics related to a living system, or at least all those that are inherent to the specific distinction between living and non-living systems.
The two following definitions go in the same direction.
The definition of life:
Life is a self-maintained chemical system that is subject to Darwinian evolution.
and the definition of living organism
A living organism is a chemical system that is capable of regenerating its own constituents, and exchanges matter and energy with its environment; this system is capable of reproducing itself in an imperfect manner, generating slightly different replications of itself, possibly better adapted to the environment.
What do we do with suffering, pleasure, emotions, anguish, feelings, etc. that seem to interact with intelligence?
There seems to be a consensus between researchers on how difficult it is to characterize intelligence, one of the consequences of the evolution of life.
A key concept here is another concept that we shall study: the concept of explanation. Is a black box that to some inputs associates outputs that resemble what happens in an organism an explanation of the behavior of the organism? We should be very careful, because, for example, in a domain we know better, the way programs play chess, prove theorems, suggest clauses that explain data…is probably not the same as the way a human would proceed.
In any case, it is worth mentioning that it is generally not wise to identify the dreams of some researches (no matter how brilliant they may be) with reality. For example, two important names in AI predicted in 1958 that as soon as 1970, computers would be capable of composing classical music, writing masterpieces, discovering theorems, playing chess, understanding and translating languages, etc.
Of all these predictions, we can say that that until now, and although considerable progress has been made, the only one that was completely realized (in 2006) is the one on the game of chess (the program Deep Fritz defeated the world champion in six games, with two victories and four draws).
REMARK 7.2.– In the analysis of the different ways intelligence can be characterized, we frequently forget to mention a meaningful fact. Radio-astronomers who are searching for proofs of artificial life in the universe try to detect non-random signals coming from outer space. Producing signals that respect laws should therefore be a (sufficient) condition to characterize intelligence.
REMARK 7.3.– In another important activity of human beings: art, researchers (most of them in neurobiology) are trying to “extract” general laws on beauty (i.e. what leads us to qualifying objects, ideas, etc. as “beautiful”).
This domain of study could be named “artificial art” and in our opinion, its importance must not be underestimated. After all, three centuries after his death, Mozart is still considered a genius, whereas generations of scientists have disappeared without a trace.
7.2. What approaches to study AI?
We provide three possible approaches (others can of course be imagined).
1) Try to define intelligence independently from humans (animals).
2) Replace AI by an expression such as “design of assistant for intelligent tasks [or requiring manual skills…]”
3) Try to design systems that mimic capabilities that psychology, history, etc. consider as intelligent in a human being.
7.3. Toward an operational definition of intelligence
In physics, when a concept is defined by specifying what operations are necessary to measure the terms that occur in the definition, we say that the concept has an operational definition.
In 1950, Alan Turing proposed in his paper “Computing Machinery and Intelligence” an imitation game now known as the Turing test2. It is presented in many different ways in the literature, and these presentations do not always correspond to what is explicitly said in the aforementioned paper, and probably not to what Turing thought on the topic. The most popular version seems to be: A system (machine) that makes you believe you have communicated (or interacted) with a human being can be considered as intelligent.
Turing remained very cautious in his paper about the notions of thought, conscience, intelligence, etc.
Turing’s paper begins as follows:
I propose to consider the question “Can machines think?”
Turing replaces this question by another question that is closely related, and that he describes as an imitation game.
7.3.1. The imitation game proposed by Turing
– Three players: a man (A), a woman (B), and a questioner (C) who can be a man or a woman.
– Rules of the game
1) A and B are in the same room.
2) C is in another room, and cannot see or hear A or B. He can only communicate through (typed) written messages.
3) C gives names to the people in the other room. For example, he says: X is the one on the left-hand side and Y is the one on the right-hand side.
4) C can ask questions to A and B.
– Goal for the players
– A: force C to make a mistake.
– B: help C give the correct answer.
– C: determine who among A and B is the man and who is the woman (by saying, for example, X is A and Y is B).
– Example of questions and answers
– C: Will X please tell me the length of his or her hair?
– X: My hair is shingled, and the longest strands are about nine inches long.
Turing suggests to replace the question “Can machines think?” with “What will happen when a machine takes the part of A in this game?”. Will the questioner make a mistake with the same frequency as when the game is played by a man and a woman? (Turing test)
Given what Turing wrote, we can deduce that a program (system, etc.) that could replace A and fool the questioner with the same frequency as a human should be qualified as intelligent or capable of thought.
Some authors attribute a key position to the Turing test in the definition of AI:
Artificial Intelligence is the enterprise of constructing a physical symbol system that can reliably pass the Turing test.
Turing believed (at least, this is what he wrote) that such programs would exist by the end of the 20th Century.
How far are we from Turing’s prediction?
In January 2000, there was a congress on the Turing test and a competition was organized: six programs were enrolled.
Results:
The questioners gave 91% correct answers after five minutes, 93% after fifteen minutes. No program was able to fool a human.
REMARK 7.4.– Some authors have proposed extensions of the Turing test to take sensorimotor functionalities into account.
These extensions are far from arbitrary and correspond to human evolution. Speech, which is essential for intelligent activity, lies within Broca’s area. The development of these zones is related to the standing posture and to manual activity.
The difficulty in designing a test, such as the Turing test, is clearly illustrated by autism. Art seems to be one of the most high-level manifestations of the human spirit (the word the most frequently used to describe Mozart, for example, is genius). There are autistic children who have exceptional qualities for painting and music, or who have mnemonic, or shape recognition capacities that are well beyond those of “normal” people, even those considered to be intelligent. In the domain of mathematics, there also exist autistic people who are prodigies in calculus.
These exceptional capabilities restricted to a unique domain3 have led researchers to postulate on the existence of a multiplicity of intelligences, controlled by rules that are hard-wired in different neural areas.
Recently, a logician who was interested in the relationship of logic with other domains (S. Buss) wrote:
I wish to avoid philosophical issues about consciousness, self-awareness and what it means to have a soul, etc. and instead seek a purely operational approach to artificial intelligence. Thus I define artificial intelligence as being constructed systems which can reason and interact both syntactically and semantically. To stress the last word in the last sentence, I mean that a true artificial intelligence system should be able to take the meaning of statements into account, or at least act as if it takes the meaning into account.
7.4. Can we identify human intelligence with mechanical intelligence?
The aim of this argument is to show that trying to identify intelligence (and understanding) with a sequence of states leads to consequences that cannot reasonably be accepted.
The philosophical argument that is attacked is the argument that states that computers running programs can possess mental states and that when we possess the same states, similar programs are executed in our brains.
7.4.1. Chinese room argument
This argument (proposed by Searle in 1980) describes a situation showing that an entity can pass the Turing test without us being able to say that it actually thinks or understands in the traditional sense.
A person P1 who does not understand written (and spoken) Chinese at all is isolated in a room and can only communicate with the outside world through signs written on pieces of paper. The person has paper, a pencil, and an instruction manual (program) written in his mother tongue.
– He is given pieces of paper on which signs are “scrawled”.
– Using these signs and the instruction manual, P1 writes other signs on pieces of paper.
After some time, the experiment stops. A person P2 (the presence of which is ignored by P1) is outside and is the person giving the pieces of paper to P1 and receiving those passed out by P1. P2 understands Chinese perfectly. The papers that were given to P1 contained a story written in Chinese together with questions on this story. The papers that were received contained answers in Chinese.
For P2, entity P1 passed the Turing test and should be identified as intelligent (see section 7.3)…but of course, P1 does not understand Chinese. Which means that the Turing test can be passed for a spoken language without even understanding it!
Furthermore, everyone has laughed when reading automatic translations or has already been incapable of understanding (in their own mother tongue) the translations produced by translation software that is available on the Internet.
Other formulation
1) Algorithms are independent of the hardware on which they are programmed: in particular, the machine can be a human (here, of course, execution time is not taken into account).
2) We assume that there is a program P in a room that can produce speeches like someone whose mother tongue is Chinese. The program is supposed to produce speeches that a native Chinese speaker could not distinguish from those produced by a human being: the person listening believes someone Chinese is speaking.
3) By assuming the philosophical thesis mentioned above is correct, any system on which program P is executed understands Chinese.
4) There remains to imagine that the program was executed by a human being who does not know Chinese at all: the conclusion is that the person still does not know Chinese in the usual sense, and neither does the computer.
This is an argument that shows the limits of the Turing test: someone can give the impression of “understanding” without “understanding” anything at all.
REMARK 7.5.– This argument may seem far-fetched and artificially created to defend a thesis.
Yet, it partially corresponds to what happens for the Etruscan language4. The following quote is from a book on the history of the Etruscan language (L.J. Calvet) published in 1996:
Etruscan writing does not pose any difficulty: it is an alphabet inspired from Greek,…But if we are fully capable of reading this alphabet, we do not really know what language it was transcribing: we can read aloud texts that we do not understand.
This is the same for the Iberian language (the first people of the Iberian peninsula).
Neuropathology also (more sadly) illustrates that some tasks that may seem intelligent to an observer are far from being intelligent.
Among some of the “autistic geniuses”, there are children who from the age of two are able to read books and newspapers very easily, but without understanding a single word (because they are only able to decode texts from a phonological point of view and this expertise does not encompass meaning).
It is worth recalling that our intellectual faculties can be affected by emotions and feelings (stress diminishes our capacities in general, a taste for certain topics can make them easier to understand, to discover, to solve related problems, etc.).
When the chess master Garry Kasparov lost against a computer in 1997, observers believed that he had been emotionally disturbed and thus, had played poorly5.
More recently (2006), the world champion Vladimir Kramnik made an unexplainable mistake and lost a game against the software Deep Fritz, a mistake that “normally” even a beginner should not make…but errare humanum est.
This relation seems to have been appreciated in the language of a great civilization (the Chinese civilization): the ideogram for think is obtained by merging the one for head with the one for heart.
7.5. Some history
Humans have tried for a long time to reproduce or mimic life, and in particular creatures that resemble them.
For example, the mythical Golem, man-robot created by magical or artificial means.
Furthermore, in mythical stories, labor has always been a punishment and a yoke that man has always tried to get rid of.
We classify the works in two large categories, that we can imagine as increasingly merging and that we will call Prehistory and History.
7.5.1. Prehistory
Emphasis is put on the reproduction of external properties of living beings: movements, gestures, sounds, etc.
The following facts are part of tradition (hard to verify).
– 5th-4th Century BC: carrier pigeon that could fly (Archytas).
– 4th-3rd Century BC: snail that could crawl (Demetrius).
– 4th Century BC: automatic signal to call Plato’s students to class.
– 3rd Century BC: android (Ptolemy II Philadelphus).
– 1st Century AD: theater shows with automata on the return home of the heroes of the Trojan war (Hero of Alexandria).
The name of this Greek engineer and mathematician is often cited as one of the pioneers of “programmable” automata, of cybernetics, and of robotics.
– Galen (2nd Century AD) showed the purpose of human organs by analogy to machines built by man. In some way, he is also a pioneer of cybernetics.
– 
– 12th Century AD: android that opened the door and greeted when a bell was rung.
–
– 16th Century AD: mechanical lion (L. da Vinci) whose stomach would open and free lily flowers.
–
7.5.2. History
Emphasis is put on the reproduction of intellectual faculties or abilities that require different forms of learning.
– During the 17th Century, mechanical philosophy (or mechanistic materialism) tried to explain the world without ever referring to vital forces or vital causes. It believed that the method used to study the stars could be applied to physiology and human psychology.
– The circulatory system was discovered around 1628 (Harvey6), and this dented the vitalist theory (according to which vital phenomena are irreducible to physicochemical phenomena).
– 1641-1642: automatic adder (Pascal).
– 1673-1674: multiplication machine (Leibniz).
– During the 17th Century, Descartes viewed man as a machine.
– Thomas Hobbes (16th-17th Century) suggested that it could be considered that automata (meaning machines that move by themselves) have an artificial life.
– Leibniz had an eclectic mind and made connections between domains that seemed disconnected. He came up with the idea of a language for thought (lingua philosophica or characteristica universalis) in which everything that can be thought of could be transcribed, thus permitting reasoning to be (quasi)automated: the calculus raciocinator7.
– 17th Century: probability calculus (probabilities are essential in many intelligent tasks: decision making on rational databases, etc.).
– 18th Century: the flute player, Vaucanson’s duck (an engineer from Grenoble, France).
– 19th Century: Jacquard’s weaving loom, punch cards.
– Ampère (1834) included cybernetics (science of direction) in “politics”.
– 19th Century: development of probability calculus (Laplace, Poisson, etc.).
– End of the 19th Century: Ramon Cajal discovered the nature of neurons and their interconnections.
– 1924: the word robot, from the Czech word robota meaning “forced labour” appears in a play by Karel Capek to denote artificial workers.
– 1927: Hilbert wrote: “The key idea of my proof theory is nothing more than the depiction of the activity of our intelligence, of listing and analyzing the rules that guide the way our thoughts really function”.
– Some physiologists (Belle, Young, Helmholtz, and others) convert to physics (H. Helmholtz believed that “no other forces than those of physics and chemistry are active within an organism”).
– ≈ 1936: Turing.
– ≈ 1936: Wiener et al. cybernetics (science of direction and communication in living organisms and machines).
– ≈ 1940: neural networks (MacCulloch, Pitts).
– Shortly after World War II, interdisciplinary exchanges (mathematicians, neurophysiologists, psychologists and sociologists) took place among other things on the possibility of imitating human intelligence.
– ≈ 1950: electronic turtles capable of recharging themselves.
– ≈ 1950: electronic mouse that could learn a path.
– … AI (the “official birth” of AI is supposed to have taken place in 1956).
– What is considered as the first AI program Logic Theory Machine (A. Newell, H. Simon), a program for theorem proving, is presented in 1956.
REMARK 7.6.– In the near future, recent advances in biology and in particular the study of the brain via medical imaging will probably influence the models that are used in AI, as well as its foundations and techniques. If that were to happen, it would be an item to add in the list above.
7.6. Some undisputed themes in AI
– design of expert systems
– different kinds reasoning (inferences): deductive, inductive, abductive, probabilistic, non-monotonic, under uncertainty, etc.
– games (chess, go, etc.)
– knowledge representation
– learning
– robotics, vision, image analysis
– speech, writing recognition
– human-computer interaction
– natural language processing
– multi-agent systems
– planning
– constraint satisfaction
– computational linguistics
– neural networks
We have seen (section 7.2) that one of the approaches to studying AI was to try to mimic the way humans do things.
When we analyze (in particular by introspection) the way a human solves a problem, the most striking characteristics are the diversity of approaches that are used and the capability of humans to distinguish the “right” context and to use the “right” properties and relations for the problem under consideration.
This is particularly striking in so-called clever or elegant solutions.
In contrast, mechanical solutions would be qualified as “uniform”.
Furthermore, the etymology of the word “find” is worth mentioning here:
Find 
twist, in Greek evolution, change, related to change,
which seems to suggest that to find a solution, it is necessary to consider a problem by “twisting” it, i.e. by analyzing it from different angles.
1 The term “intelligent” is currently used, often in a glamorous way, on anything that has more or less surprising properties (for example, in nanotechnology, there are “intelligent materials”).
2 A lot of information, discussions, etc. on this test can be found on the Internet.
3 There are some rare cases in which several of these capabilities have been known to coexist within the same individual.
4 Not much is known about the Etruscan civilization: it originated in Italy around 700 BC and disappeared around 350 BC.
5 As a neurobiologist wrote: “Thus, the functioning of the limbic system, which supports emotions, memory, and therefore the cognitive system, depends on a perfect and delicate tuning of different neuromodulators. Too much or not enough chemical activity prevents these systems from functioning normally.”
6 And maybe even before M. Servet.
7 Leibniz prophesied that once the calculus had been perfected, men of good faith who wanted to settle a question would take a pencil and paper and say: Calculemus!