How will we know when we've built the first intelligent machine—a
computer capable of thought? It will pass the Turing Test. In 1950,
British mathematician Alan Turing—later revealed as one of his
country's chief code breakers during the Second World War—wrote an
article called "Computing Machinery and Intelligence" for Mind, a
small Oxford University philosophy journal. The article described
what would soon become the definitive method for identifying an
intelligent computer. Turing suggested that if a machine and a
person carry on a conversation—via e-mail, say—without the person
realizing that the other participant is a machine, then the machine
can be deemed to think.
Kevin Dowd, CEO of Brainhat, a Connecticut-based software
developer, likes to say that his company's natural language
operating system has passed the Turing Test. In development for the
past six years, the company's system is designed to analyze natural
language—ordinary English, for instance—and respond to it, often
using built-in speech recognition and speech synthesis software to
carry on verbal conversations.
Recently, in testing the system, Dowd's employees connected it to
a telephone line, and a woman accidentally dialed the number. The
Brainhat system, programmed to sound like the quintessential Valley
Girl, often inserting the word "like" between phrases, answered the
call and spoke to the woman for several minutes. "She got all the
way through the call without knowing she was talking to a computer,"
claims Dowd. "We passed the Turing Test."
Dowd's assertion is, of course, tongue-in-cheek. If you listen to
of the phone call you'll realize that the machine was by no
means carrying on a logical conversation. Yet, as some of Brainhat's
tests demonstrate, the system is more intelligent than your
average computer, and the company is well on its way to building a
machine that can carry on limited conversations, as are language
labs at IBM, Microsoft, and various universities around the
Currently available for download, the system operates in a
way familiar to anyone with a grade school education. "Remember
diagramming sentences? That's basically what we do," explains Dowd.
"We extract semantic value from language by parsing through it,
identifying different parts of speech, and organizing everything
within various data structures." Once a sentence is broken into
pieces that the system can recognize and, to a certain extent,
understand, it then manipulates these pieces—turns them into a
related question, say—in an effort to generate a feasible response
to the sentence.
As a conversation continues, the system can use what it's learned
about each sentence to better understand subsequent sentences and,
thus, provide better responses. An upcoming enterprise version of
the system—a way for businesses to develop their own natural
language applications—will also include a back-end SQL database of
conversational data that further facilitates the software's ability
to understand the context of a particular sentence.
In the near term, the system will likely be used to improve the
performance of today's speech recognition packages. If a package can
understand, on some level, the context of what a person is saying,
then it becomes much better at distinguishing between words. Several
organizations have also enlisted Brainhat to help them build more
proficient robots. "We're working with NASA to build a robot arm you
can talk to," says Dowd.
In the long term, Dowd and his staff hope to reach the point
where their system can handle customer support and other specialized
but practical tasks. "Wouldn't it be great to have a conversation
with your car to find where your wife took it last night?" he asks.
Ultimately, a machine has yet to legitimately pass the Turing Test,
but it may happen sooner than you think.