As the latest group of contestants fails the infamous Turing Test, John Pavlus explores what it will take to create a computer that can think like a human.
How can you tell if a computer can think? In 1950, Alan Turing, the father of computer science, suggested a simple test. First, design a computer program that can mimic human conversation. (This was quite a challenge, considering how basic mid-twentieth-century computers were.) Next, hide the computer behind a screen or otherwise keep it out of sight. Then, invite a person to chat with the computer using text messages. Finally, ask the person if they believe their unseen conversation partner is human or a machine.
If the person mistakes the computer-generated conversation for a human, then voilà: according to Turing, the computer can be said "to think."
It sounds more like a parlor game than a serious experiment, and many in the field see it that way. Still, this "Turing test" has driven decades of research in artificial intelligence. It even led to an annual contest since 1991 called the Loebner Prize, where judges have short conversations with hidden AI programs and humans, then decide which is which. If no program passes the test, a smaller prize is awarded to the most "humanlike" one.
Turing predicted that a computer would pass his test before the year 2000, but so far, no program has succeeded—not even the winner of the 2012 Loebner Prize, Chip Vivant. This is partly because the test's parameters are unclear—for example, how long should the conversation last before the human decides if it's a person or a machine? Five minutes? Three hours? Turing never specified. Additionally, perfectly imitating human conversation is more complex than anyone anticipated. So, what would it take to build a machine that can pass Turing's famous test?
Mind your language
One thing is certain: brute-force logic won't work. In the early days of artificial intelligence research, "thinking" was thought to be simply a matter of connecting symbols using specific rules. "In the 1960s, there was this idea of dividing the world into objects and actions you can name: book, table, talking, running," says Robert French, a cognitive scientist at The French National Centre for Scientific Research. "All the words in the dictionary are symbols that refer to the world. So if you put them together carefully, intelligence should emerge, roughly."
But it doesn't. This approach, known as "symbolic AI," falls apart with even a bit of ambiguity. No dictionary rule can tell a computer how to respond appropriately to a casual question like, "What's up?" (If you answer "Up is the opposite of down," you've just failed the Turing test.) A densely interconnected database might hold "intelligent" information, but it isn't intelligent itself.
A much better way to simulate human conversation is to sidestep logic and aim for a quality called "statelessness". In a stateless conversation, each response only has to vaguely reference the one that came immediately before it. This behaviour is much easier to program into a computer, which is why so-called “chatbots” have become so prevalent online.
In the mid-1960s, ELIZA, one of the world's first chatbots, effectively impersonated a psychotherapist by parroting users' language back at them. In the 2000s, a more sophisticated chatbot called ALICE won the Loebner Prize three times by using essentially the same technique. Still, these stateless interactions are hardly what anyone would call "intelligent" – ironically, it is their almost total vacuousness that makes them seem so human. But not human enough, it seems: ALICE may have outperformed other chatbots, but it still could not fool human judges consistently enough to pass the Turing test.
So if knowing plenty of facts and making hollow small talk are not skills that allow a computer to pass the Turing test, what aspect of "human-like intelligence" is being left out? The one thing we have that computer programs do not: bodies.