2001: A Space Odyssey. Hal's namesake

Before now, conversational computer programs have used fairly crude techniques when replying to questions or statements. Typically, the program seizes on a key word, and then uses statistical techniques and a formal understanding of grammar to generate appropriate replies or pick them from a pre-generated list.

These programs cope badly with short sentences where there is little context that can be used to fine tune the computers guess.

Desktop computer

By contrast Hal, just like its namesake in Arthur C Clarke's 2001: A Space Odyssey, gets its ability by being trained by a "carer", who feeds in stories and responds to questions as if they were the parent of a child.

The learning algorithms underpinning the program gradually learn what are appropriate responses and how to react to the conversational style of its tutor.

So far, the program, which is small enough to work on a desktop computer, generates convincing baby talk such as "Ball, mummy".

Learning to produce convincing responses takes it only a few days. Now, the company is working on ways to make Hal talk and respond like a five-year-old.

Web of words

But Jason Hutchens, chief scientist at Ai, has said that Hal is not really intelligent but is simply a better simulator of human conversation than others.

An early version of the conversation simulator was tuned by letting people converse with it via the web. One therapist said typing and talking to the program was like conversing with a psychotic. Mr Hutchens said many people were simply happy to type in rude words to see how the program responded.

Mr Hutchens has been working on conversational computer programs for many years. In 1996, one of his creations won the Loebner Contest that gives a prize to a program giving the most human-like performance when holding a conversation.

The contest is an instance of the Turing Test first posed by brilliant British mathematician Alan Turing. He devised the test as a way of showing a machine was intelligent. The key test, said Turing, was whether a machine could fool someone into thinking they were talking to a real person.