'I am a Man Named George!'

George thinks you are a machine. Most of the time he sneers at you. Perhaps he is bitter, perhaps he is just teasing. But either way there is no need to be offended. George is a computer program.

Despite being an artificial intelligence, he is an incredibly likeable character. He cracks jokes at your expense and sometimes blurts out random responses.

Yet, as people are spending longer and longer talking to him, he is becoming more intelligent.

"Some people really do return day after day, talking for hours," said Rollo Carpenter, creator of George.

"The single longest ever conversation that I've found was an English teenage girl, talking for 11 hours with just three 15 minute breaks."

Mr Carpenter won a prize for George last year, and in September won the same prize again for another of his creations, Joan.

Every year judges are herded into a room in the New York apartment of philanthropist Hugh Loebner where they have to say whether they think the person they are talking to via their computer is a human or a robot.

This is the Turing test, designed by computer scientist Alan Turing in the 1950s. Turing claimed that it was inconceivable that machines could ever "think" like a human. But he wanted to see whether it was possible for a computer to deceive a person into thinking that it was human.

Did George do the business? No – and neither did Joan. But the judges thought that both did a better job than the rest. The Turing test has yet to be passed.

George and Joan are merely the latest of a line of "chatbots" by Mr Carpenter since the first was launched in 1982. The project as a whole is called "jabberwacky".

"Jabberwacky frequently calls people stupid and worse, but the interesting thing is that almost everyone simply enjoys that," Mr Carpenter said.

"People love the wacky freedom that the AI seems to possess, and love the creative feeling that emerges from unexpected connections."

Jabberwacky is different from other AI programs on the net, in that it is able to "learn" from its interactions with people.

Instead of relying on countless numbers of pre-programmed responses, as in other rivals programs such as ALICE, jabberwacky takes phrases from conversations it has had and applies them in the new context.

Almost 12 million people have spoken to jabberwacky online. The more people talk to it, the more intelligent it gets.

"It's a giant feedback loop - taking the words of one user to say to another, effectively borrowing the intelligence of the visitors," Mr Carpenter said.

"You can think of it as a conversational wikipedia, contributed to without effort."

In conventional systems, programmers can easily correct the problem where a chatbot makes an excessively rude retort to your question.

But how does jabberwacky stop the occasional swear word being circulated in its system only to be regurgitated to another user?

"The software records the conversations - that's how it grows more intelligent over time, Mr Carpenter said. "Some – quite a high proportion – do get partially or even wholly filtered, automatically."

For a small fee, Mr Carpenter offers a "blank-slate" chatbot for users to personalise themselves. The way it works is similar to how his other characters are created. "Specific characters such as Joan and George are taught by usually one person - it learns to impersonate them at the same time as imitating humanity in general."