AI from zero

Lesson 1 of 819506 min

Turing and the imitation game

In 1948 Alan Turing moved to Manchester University as deputy director of its Computing Laboratory. Two years later, in the October 1950 issue of the philosophy journal Mind, he published a paper called “Computing Machinery and Intelligence”.

It opened by proposing to consider the question “Can machines think?”, and then, almost at once, set it aside. The words “machine” and “think” were too slippery to argue about. Turing offered a game instead.

The game

The imitation game has three players. A man (A) and a woman (B) sit in one room; an interrogator (C) sits in another and can only exchange typed messages with them, ideally through a teleprinter so that voices give nothing away. The interrogator's task is to work out which is the man and which is the woman. A tries to mislead; B tries to help. Then Turing changes one thing: a machine takes A's place. If the interrogator is fooled as often as when a man was playing, Turing suggests, the original question has lost most of its force.

The move matters more than the details. Turing replaced a question about inner experience with one about observable behaviour. That is still how most of AI is judged: not by what a system is, but by what it does on a task, measured against a benchmark. It is also the test's best-known weakness. A program can produce convincing replies without anything we would call understanding, which is why passing for human in a short chat says little about whether an answer is true.

A prediction with numbers

Turing put figures on his expectations. He believed that in about fifty years it would be possible to program computers with a storage capacity of about 10⁹ (he meant binary digits) to play the game so well that an average interrogator would have no more than a 70 per cent chance of making the right identification after five minutes of questioning. It was a forecast about one game, not a definition of intelligence, and Turing never claimed it was a finishing line.

Nine objections and a child machine

Most of the paper answers critics in advance. Turing lists nine objections, from the theological to the mathematical, and gives each a reply. The best known is Lady Lovelace's objection: Ada Lovelace had written that Babbage's Analytical Engine could not originate anything, only do what it was ordered to do. Turing's answer pointed to the last section of the paper. Instead of programming an adult mind rule by rule, he proposed building a child machine with little built in and educating it, with rewards and punishments, so that it could learn. That is, in outline, the idea that would come to dominate the field: machine learning.

Key terms

Imitation game
Turing's test: can an interrogator tell a machine from a person through typed conversation alone?
Interrogator
The player who asks questions and has to make the identification.
Artificial intelligence
The broad field of building systems that perform tasks we associate with intelligence.
Machine learning
AI that improves at a task from data or experience rather than from explicit rules.
Benchmark
A fixed task and score used to compare systems by their observable performance.

Check what you learned

  1. 1.What did Turing put in place of the question “Can machines think?”

  2. 2.What did Turing predict for about fifty years after 1950?

  3. 3.Turing's “child machine” is an early statement of which idea?

Practise it for the exam

Sources