UK university holds AI test

Published Oct 14, 2008

Share

Computers argued, cracked jokes and parried trick questions, all part of an annual test of artificial intelligence carried out at the University of Reading.

Typing away at split-screen terminals on Sunday, a dozen volunteers carried out two conversations at once: one with a chat program, the other with a human. After five minutes, they were asked to say which was which. Some were not sure who - or what - they were talking to.

"There was one time when I was speaking to the two, and there was an element of humour in both conversations. That's the one that stumped me more than others," said Ian Andrews, one of the judges in Reading, just west of London.

Transcripts of the conversations showed some savvy judges ruthlessly trying to trip programs up with questions about the day's weather, the global financial turmoil and the colour of their eyes.

"Blue, of course!" answered Eugene Goostman, a "chatbot" designed by Pennsylvania-based programmer Vladimir Vesselov. Eugene was one of five programs competing to pass themselves off as flesh and blood. A sixth program, Alice, dropped out when it could not be set up in time.

Fred Roberts' Elbot scooped the day's top award: the Loebner Artificial Intelligence Prize's bronze medal, for duping three out of 12 judges assigned to evaluate it.

"I wish I was as good at conversation as Elbot," the Hamburg, Germany-based consultant joked after receiving the prize.

The contest draws on the ideas of British mathematician Alan Turing, who came up with a subjective but simple rule for determining whether machines were capable of thought. Writing in 1950, Turing argued that conversation was proof of intelligence. If a computer talked like a human, then for all practical purposes it thought like a human too.

But judging a computer's eloquence was tricky: Humans might be prejudiced against a machine. So Turing devised the test in which a human judge would swap messages simultaneously with a computer and another human, without being told which was which. If the judge had trouble telling his correspondents apart, Turing said, then the computer met the human standard of intelligence.

Each program took a slightly different tack on Sunday.

Eugene often made pointed references to his native Odessa and "Aunt Sonya in America."

Cleverbot, designed by Rollo Carpenter, used humour to try to fool the judges.

Roberts said Elbot worked by catching some of the judges off-guard with provocative answers or impishly hinting that it was, in fact, a machine.

"Hi. How's it going?" one judge began.

"I feel terrible today," Elbot replied. "This morning I made a mistake and poured milk over my breakfast instead of oil, and it rusted before I could eat it."

Roberts also said Elbot tried to dominate the conversation to keep it from wandering into areas it was not properly programmed to handle.

Although Turing died in 1954 without laying down precise rules for the competition, American scientist and philanthropist Hugh Loebner has been overseeing an annual series of tests based on Turing's principles since 1991.

The bronze prize handed out by Loebner goes to the piece of software that best mimics human conversation in text form.

No program has won the gold or silver prizes. The silver would go to a machine that could pass a longer version of the Turing Test and fool at least half the judges. The gold would go to a machine that could process audio and visual information rather than just text.

Cybernetics professor Kevin Warwick, one of this year's contest organisers, said passing the test would be something akin to the moment IBM supercomputer Deep Blue beat Russian chess master Garry Kasparov in 1997.

Despite - or maybe because of - his win, Roberts said he did not buy Turing's argument.

"I don't think it's anything like thought," he said of Elbot's conversational prowess. "If you know a magic trick, you know how it's done, it's not magic anymore. Sorry to be so pessimistic." - AP

Related Topics: