The Most Human Human (20 page)

If we were talking about
x
, all the TA really had to know about was
x
. You could try to exceed those bounds out of curiosity, but doing so was rarely important or relevant to the matter at hand.

My philosophy seminars, though, were a completely different story. When you are trying to evaluate whether argument
y
is a good one or not, any line of attack is in play, and so is any line of defense. You’d almost never hear a seminar leader say something like “Well, that’s a good point, but that’s outside the scope of today’s discussion.”

“There is no shallow end,” a philosophy professor once told me. Because
any
objection whatsoever, from any angle, can fell a theory, you can’t carve out a space of philosophical territory, master it in isolation, and move on to the next.

My first day of class in the philosophy major, the professor opens the semester by saying that anyone who says that “philosophy is useless” is already philosophizing, building up an intellectual argument to make a point that is important to them, and therefore defeating their own statement in the very breath of uttering it. Poet Richard Kenney calls philosophy one of the “master disciplines” for this reason. You question the assumptions of physics and you end up in metaphysics—a branch of philosophy. You question the assumptions of history and you end up in epistemology—a branch of philosophy. You try to take any other discipline out at the foundations and you end up in philosophy; you try to take philosophy out at the foundations and you only end up in meta-philosophy: even deeper in than when you started.

For this reason the philosophy TAs tended to be Ph.D. students, and
even then, you’d frequently angle to get into the discussion section led by the professor him- or herself. Unlike the computer science professors and TAs, their whole training, their whole life experience—and the whole of the discipline—were in play at all times.

The other master discipline—concerning itself with linguistic Beauty rather than linguistic Truth—is poetry. As with philosophy, every attempt at escape lands you deeper than where you started.

“When I wrote ‘Howl’ I wasn’t intending to publish it. I didn’t write it as a poem,” Allen Ginsberg says, “just as a piece of writing for my own pleasure. I wanted to write something where I could say what I really was thinking,
rather than poetry
” (emphasis mine).

With poetry, as with philosophy, there is no exterior, only certain well-behaved interiors: in philosophy we call them sciences (physics originally began as the largely speculative field of “natural philosophy”), and in poetry we call them genres. If a play wanders too far from the traditions and conventions of playwriting, the script starts to be regarded as poetry. If a short story starts to wander out of safe short-story territory, it becomes a prose poem. But poetry that wanders far from the conventions of poetry is often simply—e.g., “Howl”—better poetry.

All this leads me to the thing I keep noticing about the relationship between human-made and human-mimicking bots and humans themselves.

The first few years that the Loebner Prize competition was run, the organizers decided they wanted to implement some kind of “handicap,” in order to give the computers more of a fighting chance, and to make the contest more interesting. What they chose to do, as we discussed, was to place
topic
restrictions on the conversations: at one terminal, you could only talk about ice hockey, at another terminal you could only talk about the interpretation of dreams, and so on.

The idea was that the programmers would be able to bite off some
kind of subset of conversation and attempt to simulate just that subdomain. This makes sense, in that most artificial intelligence research has been the construction of so-called “expert systems,” which hone just one particular task or skill (chess being a clear example).

Part of the problem with this, though, is that conversation is just so
leaky:
If we’re talking hockey, can I compare hockey to other sports? Or is that going outside the domain? Can I argue over whether top athletes are overpaid? Can I gossip about a hockey player who’s dating a movie actress? Can I remark on the Cold War context of the famous U.S.A.–U.S.S.R. Olympic gold medal hockey match in the 1980s? Or is that talking about “politics”? Conversational boundaries are just too porous and ill defined. This caused huge headaches for the prize committee.

This question of domain, of what’s in and what’s out, turns out to be central to the whole notion of the man-machine struggle in the Turing test—it may well embody the entire principle of the test.

I was talking to Dave Ackley about this kind of domain restriction. “If you make the discourse small enough, then the difference between faking it and making it starts to disappear,” he says. “And that’s what we’ve been seeing. So we’ve got, you know, voice recognition on corporate phone menus: you exploit the fact that you’re in a limited context and people either say digits or ‘operator.’ Or ‘fuck you,’ ” and we both chuckle. Somehow that “fuck you” touched off a kind of insight; it seems to perfectly embody the human desire to bust out of any cage, the human frustration of living life multiple-choice-style and not write-in-style.
⁸

If you ask the Army’s SGT STAR chatbot something outside the bounds of what he knows how to respond to, he’ll say something like “I have been trained to ask for help when I’m not sure about an answer. If you would like a recruiter to answer your question, please send the
Army an e-mail by clicking ‘Send Email’ and a live recruiter will get back to you shortly.” And most any telephone menu—infuriatingly, not all—will give you that “none of the above” option.
And that option takes you to a real person
.

Sadly, the person you’re talking to is frequently a kind of “expert system” in their own right, with extremely limited and delineated abilities. (“Customer service is often the epitome of empowerment failure,” writes Timothy Ferriss.) Often, in fact, the human you’re talking to is speaking from a script prepared by the company and not, in this sense, much more than a kind of human chatbot—this is part of what can make talking to them feel eerie. If what you want to communicate or do goes outside of this “menu” of things the employee is trained/allowed to do, then you must “exit the system”
again:
“Can I talk with a manager?”

In some sense, intimacy—and personhood—are functions of this kind of “getting out of the system,” “domain generality,” the move from “expertise” to “anti-expertise,” from strictly delimited roles and parameters to the unboundedness that human language makes possible. People frequently get to know their colleagues by way of interactions that are at best irrelevant to, and at worst temporarily impede the progress toward, the work-related goal that has brought them together: e.g., “Oh, is that a picture of your kids?” This is true even of the simple “How are you?” that opens virtually every phone call, no matter how agenda-driven. How two people’s
lives
are going is outside the agenda—but this initial “non sequitur,” however perfunctory, serves a profound purpose. These not-to-the-purpose comments remind us that we’re not just expert systems, not just goal-driven and role-defined. That we are, unlike most machines, broader than the context we’re operating in, capable of all kinds of things. Griping about the weather with the barista, instead of simply stating your order and waiting patiently, reinforces the fact that he or she is not simply a flesh-and-blood extension of the espresso machine, but in fact a
whole person
, with moods and attitudes and opinions about most everything under the sun, and a life outside of work.

One of the leading academics interested in the Turing test (and, as it turns out, an outspoken critic of the Loebner Prize) is Harvard’s Stuart Shieber, who actually served in the very first Loebner Prize contest as one of the “referees.” It’s a role that didn’t exist as I prepared for the 2009 test: the referees were there to keep the conversations “in bounds”—but what did that mean, exactly? The organizers and referees at the first Loebner Prize competition held an emergency meeting the night before the competition
⁹
to address it.

I called Shieber. “The night before the first competition there was a meeting with the referees,” he says. “How are we going to make sure that the confederates stay on topic and the judges don’t ask things outside of the—They’re not supposed to ask anything tricky—And what
is
a trick question? And it boiled down to, is it the kind of thing that would come up naturally in a conversation with a stranger on an airplane? You’re not going to ask someone out of the blue about sonnets or chess or something.” He pauses a split second. “If I were [in charge], that’s the first thing I’d get rid of.”

The Loebner Prize both has and hasn’t followed Shieber’s advice. After 1995, amid controversy on what a conversational domain
was
, let alone how to enforce it, the Loebner Prize committee decided to dissolve the referee position and move to an unrestricted test. Yet the “strangers on a plane” paradigm persists—enforced not so much by statute as by custom: it just comes off as kind of “square” to grill your interlocutors with weird, all-over-the-place questions. It just isn’t done. The results, I think, suffer for it.

The advantage of specific prescribed topics was, at least, that conversations tended to hit the ground running. Looking back at those years’ transcripts, you see some hilariously specific opening volleys, like:

In comparison, with the topic-less conversations, you often see the judge and the interlocutor groping around for something to talk about—the commute? the weather?—

For some reason I begin enjoying books much less when I’m almost done with them, because some inner drive starts yearning for “completion.” The beginning of the book is about pleasure and exploration, the end is about follow-through and completeness, which interest me much less.
¹⁰

Somehow I’m particularly susceptible to this notion of purpose or
project completion. Some weeks ago a few friends of mine all met up at one of our houses, and we’d decided to walk to a bar from there. As we’re putting on our coats, Led Zeppelin’s “Ramble On” comes on the stereo and someone spontaneously begins jumping around the room, flailing at an air guitar; one by one, we all join in. Yet the whole time I’m anxious to go, thinking, C’mon guys, we’re wasting time, we were supposed to be hanging out by now! Obviously, we already were.

“In our everyday life we are usually trying to do something, trying to change something into something else, or trying to attain something,” I read recently in the book
Zen Mind, Beginner’s Mind
. “When you practice zazen you should not try to attain anything.” But there’s the paradox waiting around the corner, which is to treat the mind state of non-attainment as itself the goal to be attained … It’s a bit like trying to look at the floaters in your own eyes, those little spots you see in your peripheral vision when you look at a uniform blue sky, which of course always slide away when you try to center them in your view. You go directly after non-attainment and always miss, of course.

The mind trips itself up, because as soon as you start saying, “Good job, self! I succeeded at doing something non-teleological,” you fail on that very score.

There was a commercial during the 1990s where a man puts on a pair of headphones and lies down on a couch. He’s got a relaxation tape, and he presses the play button—all of a sudden a harsh Germanic voice comes on: “Commence relaxation, NOW!” The man stiffens into a “relaxed” posture on the couch. Regarding non-goal-directed behavior as, itself, a goal will get you into these sorts of problems.

As twentieth-century philosopher Bertrand Russell argues: “Men as well as children have need of play, that is to say, of periods of activity having
no purpose.
” And Aristotle, who stressed the “teleology” of everything from men to microbes, went out of his way to describe the best form of friendship as one with no particular purpose or goal. Dolphins, allegedly, and bonobos are the only animals besides humans that have sex “for fun.” We also tend to regard them as the
smartest animals besides ourselves. Indeed, it seems that the list of “smartest” animals and the list of animals with some form of “play” or recreation in their daily lives are more or less the same list.

One of the odd things about domain-general chatbots at the Loebner Prize competitions—programs that, owing to the setup of the Turing test, must be jacks of all trades and masters of none—is this “What’s the
point
?” question. And it’s this question that contributes to what seems, at times, uncanny about them; it’s also what makes them so underfunded. In contrast, their cousins, the “expert systems,” the conversational equivalents of the hammer or the saw—you buy airline tickets, file a customer service complaint, etc.—are becoming increasingly richly funded, and are increasingly being rolled out into commercial applications.

Philip Jackson, the 2009 contest’s organizer, explains that one of the reasons the Turing test has been such a resilient one is that programs that do well often get co-opted by larger corporations, which then put the technology to some particular
use
. Some critics of the Loebner Prize describe its programmers as “hobbyists” rather than professionals; this isn’t true on the whole. Cleverbot’s author, Rollo Carpenter, who won the Most Human Computer award in 2005 and 2006, contributed the AI for the “interrogation” stages in
221b
, the 2009 computer game whose release accompanied the most recent Sherlock Holmes film. The Most Human Computer award winner from 2008, Elbot’s programmer, Fred Roberts, is part of the company behind the customer service chatbot at the IKEA website, among a number of others. These are professionals indeed: it’s just that the bots that make money are “domain specific” (divulge clues to move the game narrative ahead, point the user to the curtains department), and the bots that win Turing tests are “domain general,” conversing, as humans do, about whatever comes up. Jackson explains that companies and research-granting agencies appear to be having a hard time thinking of a reason—yet, anyway—to direct money into developing domain-general bots, conversational “universal machines.”