Turing test: what is it and why is it so difficult to pass? Alan Turing
The phrase "Turing test" is more appropriate to use to refer to a sentence that concerns the question of whether machines can think. According to the author, such a statement is “too pointless” to deserve discussion. However, if we consider the more specific question of whether a digital computer is able to cope with some kind of imitation game, then there is the possibility of an exact discussion. Moreover, the author himself believed that not too much time would pass, and computing devices would appear that would be very “good” in this.
The expression "Turing test" is sometimes used more generally to refer to certain behavioral studies of the presence of mind, thought, or intelligence in supposedly intelligent subjects. For example, it is sometimes suggested that the prototype of the test is described in Descartes’s Discourse on the Method.
Who came up with the Turing test?
In 1950, the work “Computing Machines and Intellect” was published, in which the idea of playing imitation was first proposed. The one who invented the Turing test is an English computer scientist, mathematician, logician, cryptanalyst, and theoretical biologist Alan Matheson Turing. His models allowed him to formalize the concepts of algorithm and computation, and also contributed to the theory of artificial intelligence.
Turing describes the next kind of game. Suppose there is a person, a car and a person asking questions. The interviewer is in a room separated from the other participants who pass the Turing test. The purpose of the test is that the questioner determines who the person is and who the machine is. The interviewer both subjects are known under the X and Y labels, but at least at the beginning he does not know who is hiding behind the X mark. At the end of the game, he must say that X is a person and Y is a machine, or vice versa. The interviewer is allowed to ask test subjects the following Turing test questions: “Well, will X be kind enough to tell me if X is playing chess?” Anyone who is X must answer questions addressed to X.The purpose of the machine is to deceive the questioner, and he mistakenly concluded that she is a human being. Man must help establish the truth. In 1950, Alan Turing said about this game: “I think that in 50 years it will be possible to program computers with a memory capacity of about 109in such a way that they can successfully play the imitation, and the average interviewer with a probability greater than 70% will not be able to guess who the machine is in five minutes. ”
Empirical and conceptual aspects
There are at least two kinds of questions that arise regarding Turing predictions. First, empirical - is it true that computers already exist or will soon appear, capable of playing imitation so successfully that the average interviewer with a probability not exceeding 70% makes the right choice within five minutes? Secondly, conceptual - is it true that if the average interviewer after five minutes of interrogation with a probability of less than 70% correctly identified a person and a car, then we must conclude that the latter shows a certain level of thinking, intelligence or reason?
Few doubt that Alan Turing would be disappointed with the state of the imitation game by the end of the twentieth century. Participants in the Lebner Competition (an annual event during which computer programs are subjected to a Turing test) are far from the standard presented by the founder of computer science. A quick look at the protocols of the participants over the past decades shows that the car can be easily detected with the help of not very sophisticated questions. Moreover, the most successful players constantly declare the difficulty of the Lebner competition due to the lack of a computer program that could have a decent conversation for five minutes. It is generally recognized that competitive applications are developed solely for the purpose of receiving a small prize, awarded to the best participant of the year, and they are not designed for more.
Turing test: passage is delayed?
By the middle of the second decade of the 21st century, the situation remained almost unchanged. True, in 2014, claims arose that the Eugene Goostman computer program passed the Turing test when it deceived 33% of the judges in the 2014 competition. But there were other one-time competitions in which similar results were achieved. Back in 1991, PC Therapist misled 50% of judges. And in the 2011 demonstrationCleverbot had an even higher success rate. In all these three cases, the duration of the process was very small, and the result was not reliable. None of them gave a strong reason to believe that the average interviewer with a probability of more than 70% correctly identifies the respondent during the 5-minute session.
Method and forecast
In addition, and much more important, it is necessary to distinguish between the Turing test and the prediction that he made about his passing by the end of the twentieth century. The probability of correct identification, the time interval during which the test takes place, and the number of questions needed are adjustable parameters, despite their limitation to a specific prediction. Even if the founder of computer science was very far from the truth in the prediction he made about the situation with artificial intelligence by the end of the twentieth century, it is likely that the proposed method is fair. But before you approve the Turing test, you should consider various objections that need to be considered.
Do I have to be able to speak?
Some people consider the Turing test as chauvinistic in the sense that it recognizes the mind only in objects that are able to keep up the conversation with us.Why can not there be reasonable objects that are unable to lead a conversation, or, in any case, a conversation with people? Perhaps the thought behind this question is true. On the other hand, it can be assumed that there are qualified translators for any two intelligent agents who speak different languages and can conduct any conversation. But in any case, the charge of chauvinism is completely irrelevant. Turing asserts only that if something can lead a conversation with us, then we have good reason to believe that he has a consciousness like ours. He does not say that only the ability to talk with us indicates a potential possession of a mind similar to ours.
Why so easy?
Others consider the Turing test insufficiently demanding. There are anecdotal evidence that completely stupid programs (for example, ELIZA) may seem to the ordinary observer to be owners of the intellect for quite some time. In addition, in such a short time as five minutes, it is likely that almost all interviewers can be deceived by cunning, but completely unreasonable applications.However, it is important to remember that the Turing test program cannot pass by deceiving "ordinary observers" in different conditions than those in which the test should take place. The application must be able to withstand the interrogation of one who knows that one of the two other participants in the conversation is a machine. In addition, the program must withstand such interrogation with a high degree of success after multiple trials. Turing does not mention what specific number of tests will be required. However, we can safely assume that their number should be large enough so that we can talk about the average value.
If the program is capable of this, then it seems plausible to say that we, at least previously, will have reason to assume the presence of intelligence. It may be worth emphasizing once again that there may be a smart subject, including a smart computer, that fails to pass the Turing test. You can allow, for example, the existence of machines that refuse to lie for moral reasons. Since it is assumed that the human participant must do everything possible to help the interviewer, the question “Are you a machine?” Will quickly distinguish such pathologically true subjects from people.
Why so hard?
There are those who doubt that the machine will ever be able to pass the Turing test. Among the arguments put forward by them are the difference in the time of recognition of words in native and foreign languages among people, the ability to rank neologisms and categories, and the presence of other features of human perception that are difficult to simulate, but which are irrelevant to the presence of reason.
Why a discrete machine?
Another controversial aspect of the work of the Turing test is that its discussion is limited to "digital computers". On the one hand, it is obvious that this is important only for the forecast, and does not concern the details of the method itself. Indeed, if the test is reliable, then it is suitable for any entity, including animals, aliens and analog computing devices. On the other hand, it’s a very controversial statement that “thinking machines” should be digital computers. It is also doubtful that Turing himself thought so. In particular, it is worth noting that the seventh objection, considered by him, concerns the possibility of the existence of machines of continuous states, which the author recognizes as being distinct from discrete ones.Turing argued that even if we are automatic machines of continuous states, then a discrete machine will be able to imitate us well in the imitation game. However, it seems doubtful that his considerations are sufficient to establish that in the presence of machines of continuous states that have passed the test, you can make a discrete finite state machine, which will also successfully cope with this test.
In general, the important point is that although Turing recognized the presence of a much more extensive class of machines, in addition to discrete finite automata, he was confident that a properly designed discrete automaton could succeed in an imitation game.