The Turing Test
In the 1950’s and 60’s, the development of the worlds first ‘thinking machine’ was thought to be just around the corner. In preparation for this, Alan Turing created a test which he claimed could determine whether or not a computer could ‘think’; he called his test the ‘imitation game’. The game, which came to be known as the ‘Turing test’, required three participants; a man, a woman and an ‘interrogator’ of either sex. The idea was for the man to try to convince the interrogator that he is the woman and for the woman to try and convey who she really is; the interrogator must then try to distinguish between the two. The conversation could involve any question that could be conveyed by teletype and would take place from different rooms so that the interrogator could not be influenced by visual cues. Turing proposed that without the interrogator’s knowledge, the man should be replaced by a computer. The computer must then try to convince the interrogator that it is the woman. If an interrogator was unable to distinguish between the computer and the woman as many times as between the man and the woman, then the computer would be said to be capable of thought. The ‘Turing test’ has undergone some modifications since it’s initial proposal; contemporary versions tend to eliminate the man vs. woman element, instead the interrogator must simply identify whether they are communicating with a human being or with a computer (French 2000). At the time of it’s conception, it was predicted that by the year 2000, computers would be able to pass the Turing test 70% of the time (Turing 1950), however this is not the case. Researchers into artificial intelligence are becoming increasingly aware of the difficulty of producing a machine able to pass even a restricted Turing test (Loebner 1994), not to mention a version that tests more complex cognitive and sub-cognitive abilities (French 2000).
Would passing the Turing test actually prove intelligence?
Is the Turing test even an adequate measure of artificial intelligence? John Searle (1980) used the example of the ‘Chinese Room’ to suggest that it is not. In Searle’s analogy, a non-Chinese speaking person is put in a room; a Chinese person is asked to write a note in Chinese and pass it into the room. The non-Chinese person looks up the symbols written on the paper and finds some corresponding symbols and passes these outside resulting in the Chinese person thinking that the person in the room can understand Chinese. Searle used this analogy to suggest that a computer that passed the Turing test would have no more understanding of the conversation that had taken place than the occupant of the ‘Chinese Room’. This argument does highlight a possible flaw in the Turing test but it could be argued that the non-Chinese person was simply a part of a system and that although he did not understand Chinese, on some level the system, as a whole, did (Harnard 2001); although a computer may only simulate the ability to think, it may take some intelligence to simulate intelligence. This leaves us with some important questions about our understanding of intelligence: does successful imitation of a human denote intelligence? Or just highly developed skills of imitation? Is it arrogant to assume that intelligence can only be measured on human terms? Perhaps a machine could be intelligent and yet be unable to imitate a human successfully? What about the ability to act creatively and autonomously and to give valid reasons for such behaviour?
Could the test be adapted to make it a better measure of intelligence?
Gödel (1931) proposed alterations to the Turing test on the basis that there is knowledge within any system (e.g. human society) that all members of said system believe to be true but that cannot be proven within that system. Thus there should be certain questions which are impossible for a machine to answer as a human would unless it has human-like intelligence. When asked questions to which there are no existing answers (e.g. “which word do you find prettier, blutch or farfalletta” or “rate banana splits as medicine”), humans would be expected to provide similar answers to one another. This is because of an underlying cognitive similarity, based on complex associations acquired over a lifetime of similar sensory, physical and emotional experiences. A machine would have to base it’s answers on pure chance and is unlikely to provide a similar pattern of answers as a human would. This ‘adapted’ Turing test would be reliant on ‘life experience’, something which a computer will not have had. Thus it is very unlikely that a machine could pass the test on these criteria unless it was able to learn, was given human-like sensory input and had been ‘brought up’ by humans so that it could learn to use the same systems that humans use to answer these questions.
Is imitation intelligence?
It seems unfair to assume that because an ‘intelligent’ machine has not had the chance to learn through experience that it cannot do so. Perhaps, a machine that was given a ‘human’ upbringing and a chance to learn as humans do would indeed be able to pass a test such as this. Perhaps this, then, would be considered to be a truly intelligent machine? However, is treating intelligence and ‘humanness’ as one and the same unfair? Someday the Turing test may be useful in identifying ‘artificially’ intelligent machines that also have ‘human-like’ thought capacity. It may allow us to decide whether or not it is ethical to destroy machines that are capable of independent thought and feeling. But in terms of judging intelligence, it seems to me that successful imitation of a human is not necessarily the bar by which we should judge! Perhaps, rather than expecting computers to prove themselves intelligent on human terms, we should be willing to accept machines as intelligent according to their own standards.