The Most Human Human (5 page)

What this also suggests, intriguingly, is that the task of translating (or writing) literary novels cannot be broken into parts and done by a succession of different
humans
either—not by wikis, nor crowdsourcing, nor ghostwriters. Stability of point of view and consistency of style are too important. What’s truly strange, then, is the fact that we
do
seem to make a lot of art this way.

To be human is to be
a
human, a specific person with a life history and idiosyncrasy and point of view; artificial intelligence suggests that the line between intelligent machines and people blurs most when a purée is made of that identity. It is profoundly odd, then—especially so in a country with a reputation for “individualism”—to contemplate how often we do just that.

The British television series
The Office
consists of fourteen episodes, all written and directed by the two series creators, Ricky Gervais and Stephen Merchant. The show was so successful that it was spun off into an American version: 130 episodes and counting, each written by a different person from the last and each directed by a different person from the last. The only thing stable from week to week seems to be the cast. The arts in America are strange that
way: we seem to care what our vision falls upon, but not whose vision it is.

I remember being enchanted as a kid with the early Hardy Boys books by Franklin W. Dixon, but after a certain point in the series, the magic seemed to disappear. It wasn’t until more than fifteen years later I discovered that Franklin W. Dixon never existed. The first sixteen books were written by a man named Leslie McFarlane. The next twenty were written by eleven different people. What I’d chalked up to the loss of something intangible in those later books was in fact the loss of something very tangible indeed: the author.

Aesthetic experiences like these for me are like an unending series of blind dates where you never follow up, conversations with a stranger on the bus (or the Internet) where you never catch the other person’s name. There’s nothing
wrong
with them—they’re pleasant, sometimes memorable, even illuminating—and all relationships start somewhere. But to live a whole
life
like that?

The
New York Times
reported in June 2010—in an article titled “The End of the Best Friend”—on the practice of deliberate intervention, on the part of well-meaning adults, to disrupt close nuclei of friends from forming in schools and summer camps.
⁴
One sleepaway camp in New York State, they wrote, has hired “friendship coaches” whose job is to notice whether “two children seem to be too focused on each other, [and] … put them on different sports teams [or] seat them at different ends of the dining table.” Affirms one school counselor in St. Louis, “I think it is kids’ preference to pair up and have that one best friend. As adults—teachers and counselors—we try to encourage them not to do that.” Chatroulette and Omegle users “next” each other when the conversation flags; these children are being nexted by force—when things are going too
well
.

The same thing happens sometimes in customer service, where the disruption of intimacy seems almost tactical. Recently a merchant made a charge to my credit card in error, which I attempted to clear up, resulting in my entering a bureaucratic Rube Goldberg machine the likes of which I had never before experienced. My record for the longest single call was forty-two minutes and
eight transfers
.

The ultimate conclusion reached at the end of this particular call was “call back tomorrow.”

Each call, each transfer, led me to a different service rep, each of whom was skeptical and testy about the validity of my refund request. If I managed to get a particular rep on my side, to earn their sympathy, to start to build a kind of relationship and come across as a distinct “nonanonymous” human being, it was only a few minutes before I’d be talking to someone else, anonymous again. Here’s my name, here’s my account number, here’s my PIN, here’s my Social, here’s my mother’s maiden name, here’s my address, here’s the reason for my call, yes, I’ve already tried that …

What a familiarity with the construction of Turing test bots had begun showing me was that we fail—again and again—to actually
be
human with other humans, so maddeningly much of the time. And it had begun showing me
how
we fail—and what to do about it.

Cobbled-together bits of human interaction do not a human relationship make. Not fifty one-night stands, not fifty speed dates, not fifty transfers through the bureaucratic pachinko. No more than sapling tied to sapling, oak though they may be, makes an oak. Fragmentary humanity isn’t humanity.

If the difference between a conversational purée and a conversation is continuity, then the solution, in this case, is extraordinarily simple:
assign a rep to a case. A particular person sees it through from start to finish. The
same
person.

For a brief period a tiny plastic tab that held the SIM card in my phone had gotten loose, and so my phone only worked when I was pressing on this plastic tab with my finger. As a result, I could only make calls, not receive them. And if I took my finger off the tab mid-call, the call dropped.

The tab is little more valuable than the plastic equivalent of a soda can’s pull tab, which it resembles in appearance, and is roughly as essential for the proper functioning of the device it’s attached to. I was out of warranty; protocol was that I was out of luck and needed a new, multi-hundred-dollar phone. “But this tab weighs one gram and costs a penny to manufacture,” I said. “I know,” said the customer service rep.

There was
no
way, no way at all, I couldn’t just purchase a tab from them?

“I don’t think it will work,” she said. “But let me talk to a manager.”

Then
the same woman
got back on the line. “I’m sorry,” she said. “But …” I said. And we kept talking. “Well, let me talk to a
senior
manager, hold on,” she says.

As I’m holding, I feel my hand, which has now been pushing down steadily on the plastic tab for about fifteen minutes, begin to cramp. If my finger slips off the tab, if she hits the wrong button on her console, if there is some glitch in my phone provider’s network, or hers—I am anonymous again. Anybody. A nobody. A number. This particular person and I will never reconnect.

I must call again, introduce myself again, explain my problem again, hear again that protocol is against me, plead my case again.

Service works by the gradual buildup of sympathy through failed attempted solutions. If person X has told you to try something and it doesn’t work, person X feels slightly sorry for you. X is slightly
responsible
for the problem now, having used up some of your time. Person Y, however, is considerably less moved that you tried following her colleague X’s advice to no avail—even if it is the same advice that
she herself would have given you had she been party to that earlier conversation. That’s beside the point. The point is that she wasn’t the one who gave you that advice. So she is not responsible for your wasted time.

The
same
woman, as if miraculously, again returns. “I can make an exception for you,” she says.

It occurs to me that an “exception” is what programmers call it when software breaks.

Sometimes even a single, stable point of view, a unifying vision and style and taste, isn’t enough. You also need a
memory
. In the 2004 comedy
50 First Dates
, Adam Sandler courts Drew Barrymore, but in the process discovers that due to an accident she can’t form new long-term memories.

Philosophers interested in friendship, romance, and intimacy more generally have, in recent times, endeavored to distinguish between the
types
of people we like (or, the things we like
about
people) and the
specific
people we feel connections with in our lives. University of Toronto philosopher Jennifer Whiting has dubbed the former “impersonal friends.” The difference between the numerous “impersonal friends” out there, who are more or less fungible, and the few individuals we care about
specifically
, who aren’t fungible with anyone on the planet, lies, she says, in so-called “historical properties.” Namely, your actual friends and your innumerable “impersonal friends”
are
fungible—but only at the moment the relationship begins. From there, the relationship puts down roots, builds up a shared history, shared understanding, shared experiences, sacrifices and compromises and triumphs …

Barrymore and Sandler really
are
good together—life-partner good—but she becomes “someone special” to him, whereas he is doomed to remain merely “her type.” Fungible. And therefore—being
no different from the
next
charming and stimulating and endearing guy who shows up at her restaurant
—vulnerable
to losing her.

His solution: give her a historical-properties crash course every morning, in the form of a video primer that recaps their love. He must fight his way out of fungibility every morning.

A look at the “home turf” of many chatbots shows a conscious effort on the part of the programmers to make Drew Barrymores of us: worse, actually, because it was her
long
-term memory that kept wiping clean. At 2008 Loebner Prize winner Elbot’s website, the screen refreshes each time a new remark is entered, so the conversational history evaporates with each sentence; ditto at the page of 2007 winner Ultra Hal. At the Cleverbot site, the conversation fades to white above the box where text is entered, preserving only the last three exchanges on the screen, with the history beyond that gone: out of sight, and hopefully—it would seem—out of the user’s mind as well. The elimination of the long-term influence of conversational history makes the bots’ jobs easier—in terms of both the psychology and the mathematics.

In many cases, though, physically eliminating the conversation log is unnecessary. As three-time Loebner Prize winner (’00, ’01, and ’04), programmer Richard Wallace explains, “Experience with [Wallace’s chatbot] A.L.I.C.E. indicates that most casual conversation is ‘state-less,’ that is, each reply depends only on the current query, without any knowledge of the history of the conversation required to formulate the reply.”

Not all types of human conversations function in this way, but many do, and it behooves AI researchers to determine which types of conversations are “stateless”—that is, with each remark depending only on the last—and to attempt to create these very sorts of interactions. It’s our job as confederates, as humans, to resist it.

One of the classic stateless conversation types, it turns out, is verbal abuse.

In 1989, twenty-old University College Dublin undergraduate Mark Humphrys connects a chatbot program he’d written called MGonz to his university’s computer network and leaves the building for the day. A user (screen name “SOMEONE”) from Drake University in Iowa tentatively sends the message “finger” to Humphrys’s account—an early-Internet command that acts as a request for basic information about a user. To SOMEONE’s surprise, a response comes back immediately: “cut this cryptic shit speak in full sentences.” This begins an argument between SOMEONE and MGonz that will last almost an hour and a half.

(The best part is undoubtedly when SOMEONE says, a mere twenty minutes in, “you sound like a goddamn robot that repeats everything.”)

Returning to the lab the next morning, Humphrys is stunned to find the logs, and feels a strange, ambivalent emotion. His program might have just passed the Turing test, he thinks—but the evidence is so profane that he’s afraid to publish it.

Humphrys’s twist on the age-old chatbot paradigm of the “non-directive” conversationalist who lets the user do all the talking was to model his program, rather than on an attentive listener, on an abusive jerk. When it lacks any clear cue for what to say, MGonz falls back not on therapy clichés like “How does that make you feel?” or “Tell me more about that” but on things like “you are obviously an asshole,” “ok thats it im not talking to you any more,” or “ah type something interesting or shut up.” It’s a stroke of genius, because, as becomes painfully clear from reading the MGonz transcripts,
argument is stateless
.

I’ve seen it happen between friends: “Once again, you’ve neglected to do what you’ve promised.” “Oh, there you go right in with that tone of yours!” “Great, let’s just dodge the issue and talk about my tone instead! You’re so defensive!” “
You’re
the one being defensive! This is just like the time you
x
!” “For the millionth time, I did not
even remotely
x
!
You’re
the one who …” And on and on. A close reading of this dialogue, with MGonz in mind, turns up something interesting, and very telling: each remark after the first is
only about the previous remark
. The friends’ conversation has become stateless, unanchored from all context, a kind of “Markov chain” of riposte, meta-riposte, meta-meta-riposte. If we can be induced to sink to this level, of course the Turing test can be passed.

Once again, the scientific perspective on what types of human behavior are imitable shines incredible light on how we conduct our own, human lives. There’s a sense in which verbal abuse is simply
less complex
than other forms of conversation. Seeing how much MGonz’s arguments resemble our own might shame us into shape.

Retorts, no matter how sharp or stinging, play into chatbots’ hands. In contrast, requests for elaboration, like “In what sense?” and “How so?” turn out to be crushingly difficult for many bots to handle: because elaboration is hard to do when one is working from a prepared script, because such questions rely
entirely
on context for their meaning and, because they extend the relevant conversational history, rather than resetting it.