The Turing test

Let us now return to the behaviouristic attitude and comment on the Turing test. This test was introduced in 1950 by the computer scientist Alan Turing (1912-1954). Its purpose was to find out if a computer really could have the capability to think. The idea was as simple as it was brilliant. A person poses questions to an invisible respondent, either a computer or a human being. The impossibility of recognizing whether the answers come from a computer or a human being is said to be proof that the computer actually can think.

To ensure that the test is reasonably realistic, the communication must be exchanged in a technically neutral way, let us say with the help of a keyboard and a screen. All information about the situation is conveyed solely by way of the keyboard during the session. Also the human respondent must always tell the truth and the computer must lie if necessary to give the impression that it is a human. This point is especially important when the computer answers smart questions concerning, for example, number-crunching where it is known that the computer far exceeds the human being. (What is the square of 30497.034 times 2004.3 divided by 0.39794?)

The Turing test has inspired many efforts, among them the famous $100000 Loebner Prize in Artificial Intelligence. The requirement for entrance into this contest is to submit a conversation computer program smart enough to mislead a jury of eight interrogators, to convince them that the conversation is taking place with a human instead of a computer. Hitherto, only a single-topic program has succeeded. Whimsical Conversation, written in 1991 by Joseph Weintraub, was able to outwit four persons in the jury. In 1992 he won the bronze medal and US $2000 with the program Men vs Women. But according to the famous linguist Marvin Minsky, the Turing Test is nothing but ‘a test to see how easily a person can be fooled’.

While many natural reasons for the lack of success can be found, the main cause is what is called ‘the common sense knowledge problem’. The problem is how to store in a computer and then access the total number of identified facts that human beings use in their everyday life. On closer examination, this task very soon takes on astronomical dimensions. To cope with a new environment, human beings base their actions on a recognition of certain similarities between the present situation and well- established past experience. Appropriate responses are gradually developed through trial-and- error, through training and imitation. Mentally, human beings advance from the past into the future, with the memory of the past going before them organizing the way new events are interpreted.

To store in a computer the myriad experiences of a lifetime — with small details such as how to give correct tip or to compliment a beautiful woman — is not technically possible. The fact that Eskimos have 40 different words for snow, Japanese have fifteen ways to say no and Arabs use 60 separate words for camel must be a nightmare for the AI programmer. To create a program that uses such details with flexibility, judgement and intuition then seems even more hopeless.

The use of language in the Turing test may put the computer at an unfair disadvantage. A special test, the Chess test, was therefore designed by the psychologist, William Hartston, and a group of chess masters, to examine if it is possible to differentiate between (the intelligence of) man and computer in a game of chess. A set of chess positions is presented and an examiner compares the computer and its human adversary. The positions used were designed to show both the computer’s and the human’s playing strengths. Computers are very capable of playing complicated positions, instantly calculating the immediate consequences of possible moves. Human beings have their strength in recognizing the long- term strategic implications of a chess position.

In the test, one minute was given to each player, human and computer, to find the best move from eight presented positions. The results were measured according to a system which gave a negative score for responses belonging to the typical computer and a positive score for responses typical of a human player. The examiner had no problem in differentiating between man and computer because the latter always strove for a material gain and often captured pieces if possible. Even Deep Blue the world’s most capable contemporary chess computer, had difficulties to resist the material gain. The sacrificing of short-term interest for future benefit seems to be exceedingly difficult to implement in a chess-playing computer program.

In the spring of 1997, the world chess champion Garri Kasparov was defeated in a chess contest with a staff of programmers using Deep Blue. This attracted great attention but must be considered quite natural taken in consideration the continuously growing brute calculation force implemented in the computer. The computational complexity of the different branches of the chess game is estimated to exceed the number of atoms in the universe. The number of alternatives to choose increases exponentially with the depth of moves one looks ahead. Therefore, grandmasters seem to rely on positional, pattern-recognition-like heuristics rather than extensive chains of forward move analysis (the very strength of Deep Blue). On the other hand one can easily imagine what would happen if a slight change was introduced in the rules of the game in the middle of a contest. The human side would quickly adapt to  the new circumstances, whereas the computer opponent would be left helpless.

In spite of several decades of hard work devoted to the development of AI, no evidence of what we normally define as mental qualities can be traced in computers in operation today. No computer algorithm, however complicated, has demonstrated some kind of genuine understanding and no computer has yet passed a serious Turing test. And of course no computer has shown the slightest hint of what is considered to define the human mind: emotion, free-will, creativity and ethical awareness. Nevertheless, from several other points of view, the work done by AI researchers must be considered to be fruitful.

Source: Skyttner Lars (2006), General Systems Theory: Problems, Perspectives, Practice, Wspc, 2nd Edition.

Leave a Reply

Your email address will not be published. Required fields are marked *