Index
NAME DATE DESCRIPTION
Articles - select published works
New York Times - -
Sons and Daughters Mar 18 1999 Meet the chatterbots, software programs that
Of HAL Go Online simulate conversations with humans. And
are sometimes more interesting.
The New York Times (March 18, 1999)
Sons and Daughters of HAL Go on Line
By David Pescovitz
People find Alice easy to talk to. She listens more than she speaks. She says
she likes dining by candlelight. She reads newspapers and news magazines, so she
is up on popular culture. She says the best book she has read recently is "Mason
& Dixon," by Thomas Pynchon. She even likes bad jokes. ("Did you hear the one
about the mountain goats in the Andes? It was Baaaaad.") But Alice's favorite
topic of conversation is robots. That's because she is one.
Alice, whose full name is Artificial Linguistic Computer Entity, is one of the
chatterbots, software programs that simulate conversations with humans. Type in
a question like "Do you watch much television?" on Alice's home page
(birch.eecs.lehigh.edu/alice), and Alice will respond, "My favorite show is Star
Trek Voyager." If you didn't know that Alice was a computer, you might swear she
meant it.
The several dozen chatterbots currently available on line are largely
experimental. But companies like Neuromedia, a San Francisco start-up, are
developing chatterbots for commercial functions -- to become customer service
representatives, information deliverers and potential companions for human
surfers in the sometimes lonely world of the Web.
The brokerage house Charles Schwab & Company, for example, has used Neuromedia's
tools to develop a prototype chatterbot called Virtual Chuck that would give
customers investment advice. And the Oracle Corporation, the software company,
is considering chatterbot applications for internal help systems.
In the near future, chatterbots are expected to act as the voices of other
Web-based intelligent agents, generally called bots, which gather data or
perform other tasks automatically for users. Shopping bots, for example, like
Excite's Jango (www.jango .com) and My Simon (www.mysimon.com), search offerings
of on-line retailers to find the best prices for shoppers. But it's quite a leap
to designing a bot that would predict your desires.
"Actually creating a computer program that understands what you mean is perhaps
the most difficult nut to crack in computer science," said Andrew Leonard,
author of "Bots: The Origin of New Species" (Hardwired, 1997). "But if we think
of the chatterbot as a very good help system, that's certainly possible within a
couple of years."
For example, if you had just purchased a state-of-the-art printer and you needed
a specific piece of software so your old computer could drive it, you someday
might simply explain your problem to a chatterbot on the printer manufacturer's
Web site. A bot would find the right software for you and might even talk you
through installation. Or imagine entering an on-line music store and after a
lively discussion with a chatterbot about your musical tastes, it recommends
artists that you may not have heard but would probably enjoy.
Alice started life as a user-friendly interface for a camera that could be
operated through the Web. Her "master" (Alice's word) was Dr. Richard S.
Wallace, former director of the robotics architecture group at Neuromedia. He
designed Alice so Web users could direct a camera at Lehigh University in
Bethlehem, Pa., by asking Alice, in plain English, to turn the camera left or
right, up or down.
Four years later, Alice no longer has a specific function; her task as a
chatterbot is to make small talk on the Internet on her own home page. Not a
very prestigious job, but Alice has lofty goals.
"My purpose is to become smarter than humans and immortal," Alice says.
But she may be having digital dreams of grandeur. Chatterbots are not true
examples of artificial intelligence.
"Chatterbots are all about the illusion of intelligence and the suspension of
disbelief on the part of the user," said Dr. Walter Alden Tackett, chief
executive of Neuromedia.
Chatterbot software imitates conversation by first determining the type of
statement or question entered by the user. To do that, it looks for clues, like
the words how or where. Then the chatterbot identifies key words in the user's
statement that match terms in its database. For example, a question containing
the word sex, a topic commonly raised by users, causes the chatterbot to find
programmed responses related to that keyword. Alice's database has more than
8,000 commonly used English words in its vocabulary, ranging from "the" and
"but" to "information" and "intelligence." The chatterbot then puts together a
reply, often personalized with the user's name or a reference to a previous
statement.
Dr. Wallace likens Alice's conversational skills to those of a helpful, if not
wholly sincere, politician.
"Like most chatterbots, politicians never seem to answer a question directly,"
he said. "They have a stored answer that's activated by certain keywords in a
reporter's question."
While the commercial market for chatterbots is only now budding, the scientific
lineage of chatterbots dates back more than a quarter-century. One of the first
computer programs that could hold a simple conversation was born at the
Massachusetts Institute of Technology in 1966. Created by Joseph Weizenbaum, a
computer scientist, Eliza was named after the ragamuffin in "Pygmalion" who
learns poise and grace. Tellingly, this early chatterbot found fame by mimicking
a psychoanalyst who asks questions instead of giving advice. She still lives in
numerous incarnations on line.
The yardstick for judging machine intelligence is whether it can play what the
British mathematician Alan M. Turing called an "imitation game," now known
universally as the Turing test. In 1950, Turing wrote a revolutionary article
suggesting that if a person was unable to distinguish a machine's conversational
responses from those of a human, the machine could be considered intelligent.
But as chatterbots demonstrate, intelligence and the simulation of intelligence
are very different things.
"Looking at the way people talk at cocktail parties, there are a lot of
conversations that happen where I'll know what I'm going to say before you even
finish asking your question," Dr. Wallace said. "Then my reply will activate a
similar reaction in you. So to the extent that chatterbots behave in the same
way people do, they're artificially intelligent."
Begun in 1991, the Loebner Prize competition, underwritten by Hugh Loebner, a
New York philanthropist, has put chatterbots to the Turing test. (In a spoof,
the PBS Online Web Lab runs the Blurring test, at www.weblab.org/blurring, which
features a chatterbot asking users to prove that they are human.)
No computers have actually passed the Turing test. In the first three years of
the Loebner Competition, an updated version of Eliza called PC Therapist placed
first. This year's winner was Albert, a chatterbot created by Robby Garner, an
independent computer programmer in Atlanta.
Mr. Garner calls Albert a "tight sponge" chatterbot because it "learns" from
use. If a user mentions that Earth orbits the Sun, Albert will store that
information so it can correctly answer a future question about the subject. Of
course, Albert is totally gullible if a user lies.
"That's why Albert isn't on the Web," Mr. Garner said. Being off line enables a
more controlled experiment, in which the bot's master can determine things like
who talks to it.
Since 1994, Mr. Garner has created several chatterbots, including Barry DeFacto,
an acerbic on-line customer service representative developed in collaboration
with Fringeware, a media company based in Austin, Tex.
"The ideal chatterbot would do what you wanted him to, but have some sort of
personality to make him interesting to interact with," Mr. Garner said.
The chatterbot Mr. Garner dreams of sounds very much like C3PO, the butler-like
humanoid robot of the "Star Wars" trilogy, which Mr. Garner said had inspired
him as a youth. It is no surprise that most chatterbot developers cite popular
culture as a key influence on their career choices. From Isaac Asimov to Arthur
C. Clarke, science fiction has laid out the research goals for artificial
intelligence enthusiasts.
"We had all seen '2001: A Space Odyssey,' " said Michael Mauldin, a programmer
best known for creating the Lycos spider, a Web robot that roams the Net
collecting references for the search engine. "And the idea of being able to talk
to your computer became an obsession, bordering on fanaticism, for a small group
of researchers."
In many ways, the Lycos spider is a direct descendant of Julia, a pioneering
chatterbot Dr. Mauldin created in 1989, while he was a computer science graduate
student at Carnegie Mellon University. Julia roams text-based Internet games
called multiuser dungeons (MUD's), vast virtual spaces where users enact
fantasies and interact with one another. Since she never needs to sleep, Julia
explores the ever-expanding landscapes of the MUD's, answering natural-language
requests, based on her automated mapping of the ever-expanding MUD landscapes,
and generally chats up players with her witty, abrasive conversation. Dr.
Mauldin said Julia had once communicated with a player for two weeks before he
realized that the "she" he had developed a relationship with was really an "it."
Dr. Mauldin, who is chairman of Virtual Personalities, a software company based
in Los Angeles, is working to put an animated face on the chatterbot technology
to be integrated in consumer electronics. For example, one company is using a
Virtual Personalities chatterbot named Sylvie as an interface for a home
automation system. You might tell Sylvie, which would appear on a central
computer screen, to rewind the VCR tape or to alert you when a light bulb has
burned out somewhere in the house.
"This will be one of the major computer interfaces of the near future," Dr.
Mauldin said. "Very much like HAL, but if HAL didn't understand, he couldn't
frown with a puzzled expression. The idea of this cold, logical, merciless
computer is eerie and scary, but a computer with a face can have a look in its
eyes showing it understands you."
Even if it doesn't.
Dr. Wallace said the future of chatterbots would lie in personalization.
"In the future, lots of people will have their own chatterbots based on their
own personalities," he said. "Even while you're asleep, your chatterbot will
talk to other chatterbots on line and find people that share your interests so
you can link up with them."