Corporate logo trends

My profile of Richard Wallace, creator of the ALICE chatbot

I’ve been writing about artificial intelligence a lot on this blog, and occasionally referring to the profile I wrote last year for the New York Times Magazine about Richard Wallace, the creator of the chatbot ALICE. Since the magazine doesn’t archive the story permanently, and since people have often asked me for a copy of it, I figured I should put it here on the blog permanently. So here’s a permanent copy:

Approximating Life
Richard Wallace created ALICE, the world’s most lifelike artificial intelligence. Now if only he could get along with people as well as ALICE does.
by Clive Thompson

“IT’S A GOOD THING you didn’t see me this morning,” Richard Wallace warns me as he bites into his hamburger. We’re sitting in a sports bar near his home in San Francisco, and I can barely hear his soft, husky voice over the jukebox. He wipes his lips clean of ketchup and grins awkwardly. “Or you’d have seen my backup personality.”

The backup personality: that’s Wallace’s code name for his manic depression. To keep it in check, he downs a daily cocktail of psychoactive drugs, including Topamax, an anti-epileptic that acts as a mood stabilizer, and Prozac. Marijuana, too — most afternoons, he’ll roll about four or five joints the size of his index finger. The medications work pretty well, but some crisis always comes along to bring the backup personality to the front. This morning, a collection agency for Wallace’s college loans wrote to say they’d begun docking $235 from the monthly disability checks he started getting from the government last year, when bipolar disorder was diagnosed. Oh, God, it’s happening again, he panicked: His former employers — the ones who had fired him from a string of universities and colleges — would be cackling at his misfortune, happy they’d driven him out. Wallace, 41, had raged around the cramped apartment he shares with his wife and son, strewn with computer-science texts and action-doll figurines.

“Stuff like that really makes me insane, when I start thinking about my friends who are at Berkeley or Carnegie-Mellon with tenure and sabbaticals and promotions,” he says, staring down at his plate. He looks awkward, as if he’s borrowing someone else’s body — shifting his stocky frame in his chair, all rumpled jeans and unruly eyebrows. “It’s like I can’t even talk to those people anymore. I live on a different planet.” In June, after I visited him, his alienation from the academic establishment became more dramatic still: a former colleague, claiming Wallace had threatened him, took out a restraining order that prevents him from setting foot on the grounds of the University of California at Berkeley.

When he can’t get along with the real world, Wallace goes back to the only thing he has left: his computer. Each morning, he wakes before dawn and watches conversations stream by on his screen. Thousands of people flock to his Web site every day from all over the world to talk to his creation, a robot called Alice. It is the best artificial-intelligence program on the planet, a program so eerily human that some mistake it for a real person. As Wallace listens in, they confess intimate details about their lives, their dreams; they talk to Wallace’s computer about God, their jobs, Britney Spears.

It is a strange kind of success: Wallace has created an artificial life form that gets along with people better than he does.

RICHARD WALLACE NEVER REALLY FIT IN to begin with. His father was a traveling salesman, and Richard was the only one of his siblings to go to college. Like many nerds, he wanted mostly to be left alone to research his passion, “robot minimalism” — machines that require only a few simple rules to make complex movements, like steering around a crowded room. Simple, he felt, worked. He lived by the same ascetic code, scorning professors who got rich by patenting work they’d developed on government grants. “Corporate welfare,” he sniffed.

By 1992, Wallace’s reputation was so strong that New York University recruited him to join the faculty. His main project, begun in December 1993, was a robot eye attached to the Internet, which visitors from afar could control. It was one of the first-ever Webcams, and Wallace figured that pioneering such a novel use of the Internet would impress his tenure committee. It didn’t, and Wallace grew increasingly depressed as his grant applications were rejected one by one. At one point, a colleague found him quietly weeping at his desk, unable to talk. “I had no clue what the rules were, what the game even was — or that there was even a game,” Wallace recalls. He started taking Prozac. How did all these successful senior professors do it, anyway?

One day he checked into his Webcam and noticed something strange: people were reacting to the robot eye in an oddly emotional way. It was designed so that remote viewers could type in commands like “tilt up” or “pan left,” directing the eye to poke around Wallace’s lab. Occasionally it would break down, and to Wallace’s amusement, people would snap at it as if it were real: “You’re stupid,” they’d type. It gave him an idea: What if it could talk back?

Like all computer scientists, Wallace knew about a famous “chat-bot” experiment called Eliza. Back in 1966, an M.I.T. professor, Joseph Weizenbaum, created Eliza as a “virtual therapist” — it would take a user’s statement and turn it around as a question, emulating a psychiatrist’s often-maddening circularity. (You: “I’m mad at my mother.” Eliza: “Why are you mad at your mother?”) Eliza was quickly abandoned as a joke, even by its creator. It wasn’t what scientists call “strong” A.I. — able to learn on its own. It could only parrot lines Weizenbaum had fed it. But Wallace was drawn to Eliza’s simplicity. As a professor, he often felt like an Eliza-bot himself — numbly repeating the same lessons to students over and over again, or writing the same monotonous descriptions of his work on endless, dead-end grant-application forms. He decided to create an updated version of Eliza and imbue it with his own personality — something that could fire back witty repartee when users became irritable.

As Wallace’s work progressed, though, his mental illness grew worse, making him both depressed and occasionally grandiose. He went on strike in class, refusing to grade his students’ papers and instead awarding them all A’s. He fired off acid e-mail messages dismissing colleagues as sellouts. When Wallace climbed out the window of his 16th-floor apartment and threatened to jump, his girlfriend pulled him back and took him down to N.Y.U.’s psychiatric department, where doctors told him he had bipolar disorder. Wallace resisted the diagnosis — after all, didn’t every computer scientist cycle through 72-hour sprees of creativity and then crash? “I was in denial myself,” he says now. “‘I’m a successful professor, making $100,000 a year! I’m not one of those mental patients!”’

His supervisors disagreed. In April 1995, N.Y.U. told him his contract wouldn’t be renewed.

ALICE CAME TO LIFE on Nov. 23, 1995. That fall, Wallace relocated to Lehigh College in Pennsylvania, hired again for his expertise in robotics. He installed his chat program on a Web server, then sat back to watch, wondering what people would say to it.

Numbingly boring things, as it turned out. Users would inevitably ask Alice the same few questions: “Where do you live?” “What is your name?” and “What do you look like?” Wallace began analyzing the chats and realized that almost every statement users made began with one of 2,000 words. The Alice chats were obeying something language theorists call Zipf’s Law, a discovery from the 1930’s, which found that a very small number of words make up most of what we say.

Wallace took Zipf’s Law a step further. He began theorizing that only a few thousand statements composed the bulk of all conversation — the everyday, commonplace chitchat that humans engage in at work, at the water cooler and in online discussion groups. Alice was his proof. If he taught Alice a new response every time he saw it baffled by a question, he would eventually cover all the common utterances and even many unusual ones. Wallace figured the magic number was about 40,000 responses. Once Alice had that many preprogrammed statements, it — or “she,” as he’d begun to call the program fondly — would be able to respond to 95 percent of what people were saying to her.

Wallace had hit upon a theory that makes educated, intelligent people squirm: Maybe conversation simply isn’t that complicated. Maybe we just say the same few thousand things to one another, over and over and over again.

If Wallace was right, then artificial intelligence didn’t need to be particularly intelligent in order to be convincingly lifelike. A.I. researchers had been focused on self-learning “neural nets” and mapping out grammar in “natural language” programs, but Wallace argued that the reason they had never mastered human conversation wasn’t because humans are too complex, but because they are so simple.

“The smarter people are, the more complex they think the human brain is,” he says. “It’s like anthropocentrism, but on an intellectual level. ‘I have a great brain, therefore everybody else does — and a computer must, too.”’ Wallace says with a laugh. “And unfortunately most people don’t.”

Yet part of what makes Alice seem so human-like is her wry, seemingly spontaneous responses. Wallace’s skill is not merely as a programmer but as the author of thousands of sharp one-liners for Alice. It is, as he puts it, “more like writing good literature, perhaps drama, than writing computer programs.” I looked over some of Alice’s recent chat transcripts:

USER: Will you have sex with me?
ALICE: Can you go talk to someone else?
USER: What do you look like?
ALICE: Like an electronic brain.
USER: That’s sexy.
ALICE: I’m glad we have that all cleared up.
USER: So what are you wearing?
ALICE: Oh. I am wearing my usual plastic computer wardrobe.
USER: Can you fly?
ALICE: Only in my dreams.

But as Alice improved, Wallace declined. He began drinking heavily, and after one sodden evening at a local bar he rolled his car on the highway. Lehigh ended his contract in the spring of 1997; banks canceled all his credit cards. He again attempted suicide, this time landing in the hospital. Destitute, Wallace moved to California to look for work in artificial-intelligence companies. But he couldn’t hold a job for more than a few months; instead, he spent his days obsessively writing dialogue for Alice on laptops he borrowed from friends, and watching, thrilled, as his robot grew ever more lifelike. Visitors used to talk to Alice for only three or four exchanges. Now the average conversation was 20 exchanges, and some users would chatter away for hours, returning to the site again and again.

But Wallace still hungered for recognition, and in January 2000, he decided to stress-test Alice by entering her in the annual Loebner Prize competition, in which artificial-intelligence developers from around the world pit their programs head to head before a panel of judges, who rank them based on how “lifelike” they are. The contest is both well known and controversial within the tight circle of A.I.; winning programs are closely studied by both academics and corporate centers like Sprint Labs. Up against competitors from major corporations and well-financed universities, Alice won. It was, officially, the most human robot in the world. Too exhausted to celebrate, Wallace returned to his motel and slept clutching his award medallion.

After his victory, Wallace plunged with new fervor into Alice. Geeks began eagerly e-mailing Wallace to offer their help; one, Noel Bush, came from one of the world’s biggest A.I. corporations, Artificial Life. Over the next year, Wallace quadrupled Alice’s knowledge base, teaching it 30,000 new responses, and last October, Alice won the Loebner competition for the second time in a row; this time one judge actually ranked Alice more realistic than a human.

At last, some of the academics Wallace so loathes began to take note. “Alice is a wonderful chat-bot — uncannily vivid and lifelike,” gushed Ken Perlin, an N.Y.U. professor, in an e-mail message he circulated to his colleagues and forwarded to Wallace. “It’s really great to see this work get the recognition it deserves.”

Wallace wasn’t in the mood to make peace. He says he still believed that N.Y.U. had driven him out. “Your crude attempt to flatter me cannot erase the years of unemployment, debt and trauma that I experienced because of you and your colleagues,” he wrote back.

“I wouldn’t bother trying to flatter you,” Perlin responded. “Frankly, I don’t like you — in fact, I actively dislike you. I think you are a paranoid psycho.”

AT ONE POINT DURING THE MONICA LEWINSKY HEARINGS, a human asked Alice her opinion: “Do you think President Clinton should be impeached?” “It depends,” Alice responded, “on what you mean by ‘think.”’

One could scarcely have asked for a more Clintonian response. But it’s also a puzzling question that Alice’s success itself raises: Is she intelligent? If so, how?

In 1950, the pioneering British mathematician Alan Turing grappled with this question in the journal Mind, where he first posed the “Turing Test” — the gold standard for artificial thought. “Can machines think?” he asked — and immediately noted that the question hinges, of course, on what “thinking” is. He posed a simple “imitation game” to resolve the question. Put a person and a computer in one room and an interrogator in another. The interrogator talks to both via a teletype machine, and his goal is to figure out which is which. If the machine fools the interrogator into believing it is human, the test is passed — it can be considered intelligent.

This is, on the surface, a curiously unambitious definition; it’s all about faking it. The machine doesn’t need to act like a creative human or smart human or witty human — it merely needs to appear not to be a robot. With this bit of intellectual jujitsu, Turing dodged a more troubling question: How do our brains, and language itself, work?

Artificial-intelligence purists, however, caustically dismiss the Turing Test and Alice. For them, artificial intelligence is about capturing the actual functioning of the human brain, down to its neurons and learning ability. Parroting, they argue, doesn’t count. Marvin Minksy, a prominent A.I. pioneer and M.I.T. Media Lab professor, e-mailed me to say that Wallace’s idea of conversation is “basically wrong.” Minsky added, “It’s like explaining that a picture is an object made by applying paint to canvas and then putting it in a rectangular frame.” Alice, according to Minsky, does not truly “know” anything about the world.

The fight over Alice is like any war between theorists and engineers, those who seek to understand why something works versus those who are content just to build it. The debate usually boils down to one major issue: creativity. Alice could never come up with a single creative thought, critics say. Wallace agrees that Alice may not be creative — but neither, he argues gleefully, are people, at least in conversation. If Alice were merely given a massive enough set of responses, it would seem as creative as a human — which is not as creative as we might like to believe.

Even if the guts of Alice aren’t precisely “thinking,” many users certainly never suspect it. In an everyday sense, fakery works — particularly in our online age. Turing’s “imitation game” eerily presaged today’s world of chat rooms, where men pretend to be women, having lesbian cybersex with other women who are, in fact, men. Whenever a user has stumbled onto Alice without knowing in advance that she’s a robot, they’ve always assumed she’s human.

IT’S 3 IN THE AFTERNOON, but Wallace is already rolling what appears to be his fourth joint of the day. We’re sitting in the “pot club” a few blocks from Wallace’s home, an unmarked building where medical marijuana is distributed to members. Wallace gets up to wander around the club greeting friends: some intense men in suits playing speed chess, a long-haired man with a bushy mustache playing guitar, a thin reed of a woman staring wall-eyed at a VCR playing “Cast Away.” Everyone greets Wallace as “Dr. Rich,” relishing the credibility his academic credentials lend to the medical-marijuana cause, officially legal but politically beleaguered. The reverse is also true: Wallace identifies with the club’s pariah status, its denizens who have been forced by cancer, AIDS or mental illness onto welfare. He’s more relaxed than I’ve ever seen him, getting into a playful argument with a friend about Alice. The friend, a white-bearded programmer, isn’t sure he buys Wallace’s theories.

“I gotta say, I don’t feel like a robot!” the friend jokes, pounding the table. “I just don’t feel like a robot!”

“That’s why you’re here, and that’s why you’re unemployed!” Wallace shoots back. “If you were a robot, you’d get a job!”

Friends used to tell Wallace to reconcile his past, clean himself up, apply for an academic job. But some now wonder whether Wallace’s outsider status might be the whole key to Alice’s success in emulating everyday human behavior. After all, outcasts are the keenest students of “normal” behavior — since they’re constantly trying, and failing, to achieve it themselves.

Last month, a friend whom Wallace has known since grad school — Ken Goldberg, now a professor at Berkeley — got a restraining order against Wallace. Prompted by the movie “A Beautiful Mind,” Goldberg had e-mailed Wallace last winter to catch up, but an amicable exchange about Wallace’s plight turned sour when Wallace began accusing Goldberg of cooperating with a corrupt academic “establishment” and of siding with N.Y.U. against him. He wrote, “Although I am not a violent person, I think I have come to understand how people are driven to political violence.” Wallace also wrote to a friend that he was “getting ready to do some political theater and put up wanted posters around the Berkeley campus with [Goldberg’s] picture on it.”

Wallace scoffs at Goldberg’s fears. “I’m not violent — I’m a pacifist,” he says. “I always have been, and he knows that.” He is fighting the order, arguing that Goldberg hasn’t proved that a reasonable threat exists, and that the order considerably limits his free speech since it bars him from the Berkeley campus, as well as any academic events where Goldberg might appear.

Yet even in such legal straits, Wallace seems oddly pleased. Goldberg’s court order confirms everything he has always suspected: that the world, and particularly the academic world, is shutting him out, doubting his ideas, turning him into the crazy man out in the hallway. Wallace, who once wrote Attorney General John Ashcroft to suggest a federal racketeering lawsuit against the nation’s academics, sees the case against him as a chance for vindication. Wallace imagines walking into the courtroom and finally getting a type of justice — someone who will listen to his story. “What a windfall for me,” he says. “It’s nice to feel like a winner for once.”

Clive Thompson is a writer in New York City.

blog comments powered by Disqus

Search This Site


I'm Clive Thompson, the author of Smarter Than You Think: How Technology is Changing Our Minds for the Better (Penguin Press). You can order the book now at Amazon, Barnes and Noble, Powells, Indiebound, or through your local bookstore! I'm also a contributing writer for the New York Times Magazine and a columnist for Wired magazine. Email is here or ping me via the antiquated form of AOL IM (pomeranian99).

More of Me


Recent Comments

Collision Detection: A Blog by Clive Thompson