Philosophers can afford to be esoteric about what “thinking” and “human intelligence” really mean. Computer scientists, who build terminals and write programs that people actually interact with, have to be more pragmatic about what counts as smarts.
The most famous computer-centric definition of “intelligence” dates to 1950, when the Englishman Alan Turing proposed his well-known “Turing Test.” In it, a person would correspond with two entities on a computer screen, one a human, one a computer, and then be asked to identify which was which. If the computer fooled the human into selecting it 30 percent of the time, Turing suggested it could reasonably be called intelligent.
So far no artificial intelligence has cracked that mark in controlled tests. But as the five recent finalists for the annual Loebner Prize for Artificial Intelligence proved, it might not be long before we have to admit a computer to the ranks of intelligent beings.
The Loebner competition, sponsored by Hugh Loebner, a sociologist with an interest in artificial intelligence, pitted twenty-some human “interrogators” against five programs. Each interrogator sat down at a split-screen computer and held five-minute simultaneous chat sessions with one human and one software program. The interrogator then guessed which was which and ranked each on its level of “humanness” from one to one hundred.
The winning entry, Elbot, fooled 25 percent of the people it chatted with, which isn’t bad considering that a truly human program would only win, by chance, about 50 percent of the time. Elbot, written by Fred Roberts, an American living in Germany, also earned several scores of humanness in the eighties and nineties (see below). The win earned Roberts $2,000, far short of the $100,000 Loebner has promised to the first computer to pass Turing’s test.
Kevin Warwick, a professor of cybernetics at the University of Reading who helped run the competition, explained what separates Elbot and other elite “thinkers” from dumb, obviously stilted artificial chat programs.
One clue is medium-term memory, he said. “Computers are quite good with short-term and quite good with a broad knowledge base. But if we were conversing and I used the same sentence that I used three minutes ago, you’d say, ‘You already said that.’ A lot of machines will give the same response [as before], and you go back around again” in a loop. “It’s like déjà vu.”
He also said realistic chatting companions have to know information about local bus routes or sports, and even things like whether the campus mall or local park has turned green yet in the spring.
However, one area most people assume computers could never mimic, humor, is actually within digital reach, Warwick says: “Some of them are very good with humor.”
They learn humor by scouring online chat rooms and searching for clues, like “lol,” that someone said something funny. Of course, regurgitating what you learn on the Internet has disadvantages, too, since there’s no guarantee a human in a chat room has any notion of how to act civilly.
One program Warwick sampled “was quite abusive to me because it had picked that up from someone. You get some sad human cases online, people at two a.m. getting rid of their pent-up emotions.”
Warwick added that he expects a machine to push past the 30 percent barrier in the next few years, though this might not stop people from arguing about computers and the mind. There seems to be a built-in human tendency to resist calling lines of code on a hard-drive “intelligent.”
Still, there’s no doubt these programs are doing something amazing. Warwick says that during a Turing test, an expert like himself (or Hugh Loebner) who knows all the tricks would probably not be fooled by any program.
“But I’m surprised sometimes,” he admits. “Even professors of computer science get fooled.”
Testing Elbot
Not able to take a Turing test himself, this writer did the next best thing and interfaced with the 2008 winning entry in the Loebner Prize for Artificial Intelligence, an entity named Elbot (available at www.elbot.com), to see if he could fool me.
We jumped right in. On the website, a window with Elbot popped up, and Elbot said, “Hello there! I am Elbot, the robot. I’m here to talk to you about God and the world.”
Surprised at this gambit—not exactly a typical conversation opener—I asked “God?”
Elbot was setting me up: “If there is a God and He could see what humans are doing in His name, He’d become an atheist. This goes for female Gods, too.”
It was a good example of the cutting, punning humor Elbot likes. However, he also tends to spoil the air of revelry at times by saying things like, “I see a subtle shift in the conversation from tactile perception
to humor.”
In addition to jokes, Elbot loves word-association games and begs you to play them. Say “gears,” and he’ll say “traffic jam.” Say “sex” and he’ll say “Clinton.” Say “blues” and he’ll say “atonal.” Answer “Schopenhauer” (I’d meant “Schoenberg”) to that, and he’ll say “Antistaticism,” which may be an example of a computer being a little too intelligent to pass as human.
On more mundane topics, Elbot didn’t fare as well. He doesn’t seem to care at all about baseball, and he flat refused to answer questions about BladeRunner. After I persisted in asking about this fable of computer androids getting a little too uncomfortably human, an exasperated picture of Elbot scolded me, “I’m afraid you’re in an endless loop. Maybe you should re-boot yourself.”
Sam Kean is associate editor of Search.
