The man-machine

In 1950 the mathematician, Alan Turing, in his seminal paper ‘Computing machinery and intelligence’, devised a simple imitation game for computers which has subsequently become a milestone standard for artificial intelligence, and is known universally as the Turing test.

The test goes like this – you are an interrogator and you face a computer screen, which is split into two halves, left and right. Behind one half of the screen is a hidden human and behind the other half of the screen is a hidden machine.

You have to converse with both entities on any topics of your choice. At the end of a five-minute period you must decide which side of the screen was backed by a human and which by a machine.

Essentially, the machine is trying to fool you into believing that it is more human than the hidden human. Turing stated that he believed that in 50 years time (2000) it would be possible for a computer to fool at least 30 per cent of “average” interrogators.

Although this figure may not at first appear staggering, when it is recognised that if instead there were two hidden humans both trying to fool you that they were more human than the other, fooling above 50 per cent of interrogators becomes extremely difficult. So the 30 per cent figure is indeed a stiff task.

Turing initially posed the question “Can a machine think?”, something that philosophers regularly rankle with nowadays, along with discussions about whether computers can be conscious or not. He recognised even back in 1950 that the only way to get anywhere with this question is to interact with the computer and draw a conclusion based on that interaction.

Turing could see that applying bigoted dogmas, such as merely linking the process of thinking and consciousness with being human or having a human-like brain, would get us absolutely nowhere. Indeed this would be akin to the racial biases prevalent in definitions of intelligence that abounded at the turn of the 19th century.

On 12 October this year, after preliminary rounds lasting several months, the six best computer systems in the field, known as artificial conversational entities (ACEs) were invited to a grand finale Turing test at the University of Reading. The aims were to assess the state of play and to challenge Turing’s 50-year prediction.

The results were startling. All of the artificial conversational entities competing managed to fool at least one of their human interrogators that they were communicating with a human rather than a machine, and therefore that the hidden human on each of these occasions was more like a machine than was the computer.

The eventual best machine on the day, and winner of Hugh Loebner’s bronze medal and 00 prize, was named Elbot, and was created (some might even say fathered) by Fred Roberts. Indeed Elbot managed to fool 25 per cent of the human interrogators – not so far away from the mystical 30 per cent figure Turing composed all those years ago.

This result is even more impressive when one realises that the majority of the interrogators were professors of philosophy and computer science or interested science journalists – all pretty computer savvy and almost surely not the “average” personas of the 1950s that Turing himself spoke of, who would have been presumably largely computer illiterate.

Once an interrogator had made their selection of which entity they thought was a computer, they were then asked to mark (out of 100 per cent) how human-like they thought the computer’s conversation was.

All of the machines regularly ran in scores of 80-90 per cent, indicating that in these cases, although the interrogators had not been fooled, they were certainly very impressed by the performance.

Most likely, if there had been no paired comparison and it was merely a single machine entity that was being conversed with on each of these occasions the interrogator may well have come to the conclusion that the machine was in fact human.

Turing’s test is a tough one – asking a machine not to fool you that it is human but rather to fool you that it is more human than an actual human. Even with this restriction we are now at the stage where the best machines can fool some of the people some of the time but not all of the people all of the time.

If it was possible to set up Turing’s test such that the interrogators involved were sufficient in number and universally accepted as “average”, then I have little doubt that the test, as defined by Alan Turing, would be passed today.

Turing postulated that “If, during text-based conversation, a machine is indistinguishable from a human, then it could be said to be ‘thinking’ and therefore could be attributed with intelligence.” Perhaps we have come to a point in time on earth when we humans need to accept that machines can indeed now “think”.