Failure reasons

In the “argument from failure”, a claim is made that “a machine can never perform an action X”. As examples of actions X Turing gives the following list.

To be kind, generous, beautiful, friendly, initiative, have a sense of humor, distinguish truth from lies, make mistakes, fall in love, enjoy strawberries and ice cream, make fall in love with yourself, learn from experience, use the words correctly, be an object of your own thoughts, exercise such same diversity in behavior, like a person, to create something really new.

Turing had to use his intuition to put forward suggestions about what would be possible in the future, and we have the happy opportunity to look back to find out what the computers have already been able to achieve. It is impossible to deny that computers today perform many actions that were previously the prerogative of people alone.

Programs play chess, checkers and other games, control parts on assembly lines, check spelling in documents entered with a text editor, drive cars and helicopters, diagnose diseases, and also perform hundreds of other jobs as well as people, or even it is better. Computers have made small but important discoveries in astronomy, mathematics, chemistry, mineralogy, biology, computer science, and other fields. In each of these cases, it was required to have abilities at the level of a human expert.

Given the fact that we know about computers, it is not surprising that they easily cope with such combinatorial tasks as playing chess. But the algorithms also operate at levels not lower than a person when solving such problems, which, at first glance, do not allow one to do without human judgment or, as Turing expressed it, make "learn from experience" and require the ability to "distinguish truth from falsehood."

As early as 1955, Paul Mehl studied decision-making processes by trained experts as subjective tasks, similar to predicting the student’s successful mastery of a training program or a repeat offense by a recidivist. In 19 of the 20 studies that he reviewed, Mel found that simple statistical learning algorithms (such as linear regression algorithms or naive Bayes algorithms) produce better predictions than human experts.

Since 1999, the US Educational Testing Service has used an automated program for grading review questions on GMAT exams. The results of the program were consistent with the results of the work of examiners who were assigned to give such marks in 97% of cases, and this level corresponds to the coincidence of marks of the two examiners.

Obviously, computers can perform many actions as successfully or even better than people, including such works, which, according to people, require tremendous human insight and understanding.

This, of course, does not mean that computers in performing these works show insight and understanding (these properties are not part of the behavior, as will be discussed below), but the point is that the first assumptions about the content of thought processes required to develop This particular behavior is often false. Of course, it is also true that in many areas computers have not yet achieved significant success (to put it mildly), including the task of conducting a conversation on a free topic set by Turing.