kw: book reviews, nonfiction, artificial intelligence, contests, turing test
How soon will it be that, when we talk on the telephone with a computer, we won't realize it? At present, the experience ranges from frustrating to unsettling, even when it isn't too hard to find what you called for. (If you hate to talk to computers, the source of this image, dialahuman.com, has lists of numbers that are answered by people, for many companies.)
If a computer-generated persona one day makes it through the Uncanny Valley and can converse without creeping us all out, it will have passed the Turing Test. For those who haven't heard, the Uncanny Valley is a phenomenon first reported by Masahiro Mori: when an object is very similar to human or animal, but not quite, it disturbs us more than something a lot less naturalistic. Thus, Pokemon figures are cute and funny, but the 3D-animated figures in the film Polar Express were just unhuman enough to bother a lot of people. For example, they didn't blink in a natural way, if at all.
How about if your only interaction with something is a character-based terminal? Words only, like text messaging. At what point will you find the interaction creepy, and at what point will you be unable to determine whether the something is a human? Actually, considering that some humans are decidedly creepy, a computer-driven teletype doesn't have to cross the valley all the way.
I've known about these issues for a long time. However, I never heard of the Loebner Prize before reading Brian Christian's book The Most Human Human: What Talking With Computers Teaches Us About What it Means to be Alive. The author participated in the competition in 2009 as one of the "confederates", humans who compete with the computer programs to convince the four judges that they are human. The title of the book is that of one of the awards.
The scoring, which I haven't figured out yet, assigns a "humanness" score to each of the contestants, known only by a number to the judges. The chatbot that achieves the highest score gets the award for "Most Human Computer". The confederate who achieves the highest score gets an award for "Most Human Human". Mr. Christian won the latter award in 2009.
Should any chatbot achieve a score of 0.5 or greater from any judge, thus beating out at least one human, it has gone a long way toward winning the prize. In 2009 none of the chatbots did so, but in prior years, scores approaching the threshold of "more human than the humans" have been achieved. Once a chatbot convinces a majority of the judges that it is more human than at least half the humans, it will win the prize, and the contest will be terminated. (Now, that's a pity. At that point, things just begin to get interesting!)
This would be a much more significant achievement than the defeat of Garry Kasparov by Deep Blue in 1997 or last year's victory at Jeopardy! by Watson. Computers excel at the heavy computation needed to play chess or to look up trivia. In each of these cases, the great majority of the work by their creators went into the real world knowledge needed to decipher and rank the input from the outside world. Faced with a more prosaic requirement, such as imitating the nematode Caenorhabditis elegans as it navigates in the soil, I don't think either Deep Blue or Watson would be up to the task. Not just because they weren't programmed to do so, but because no computer has yet been successfully programmed to model the nematode (See NemaSys at U of Oregon for an effort that seems to have foundered about ten years ago).
The book takes us on quite a journey, as the author prepares himself for the competition. Although he had been encouraged by many to "Just be yourself", he judged it unlikely that a person simply running on autopilot would be much different from a machine. Thus, he prepared, mainly by interviewing people who have been in the AI field, former "confederates", and psychologists, to learn how we are human and the ways we might be mistaken for nonhuman.
One achievement many people gain is some amount of familiarity with the piano. We struggle with scales and exercises and etudes, trying to commit them to "muscle memory", which is really brain memory, but of a different sort than remembering someone's name or face. A very well-trained pianist can play most pieces of music perfectly, at full speed, upon first sight. This is called sight-reading. An "ordinary" church organist has to be able to do so, particularly in churches that take requests from the congregation. Yet this is when the most highly-trained person is most machinelike, for a computer programmed to read music and play a piano can do so instantly, without error. It seems the ten to twenty years of practice have produced a piano machine in the pianist's nervous system. But a computer pianist cannot do what the human does easily: adjust tempo and volume to the congregation's singing.
This opens a window into the present lack in AI: senses that are as powerful as our own. The human visual system, for example, has about 140 million sensors in each eye (4% are color sensors) and 500 million neurons in the visual cortex to interpret what the eyes send in through the optic nerves. The best computer vision system so far devised has a few million pixels in each detector and a processor that is the equivalent of half a million neurons or so: roughly equal to the vision of a damsel fly. But a real damsel fly outperforms the computer as it navigates in three dimensions. We hear sounds that are interpreted by some 100 million neurons in the auditory cortex. A similarly-sized system interprets touch and other sensory receptors in our skin that number in the millions.
Just sitting in a chair, as your bottom tires its of contours, provides a sensory distraction that no chatbot feels. A whisper of moving air from a ventilation system may cause a confederate to relax subtly, and may affect his or her interaction with the judge. The chatbot can rely only on what its programmer put "in there" as it reacts to the judge's queries.
During months of preparation for the Loebner Prize event, the author learned of the history of the soul, as it migrated from the heart (whether considered literally or figuratively), to the abdomen, to the head most recently, and now to nowhere, because today's psychologists deny there is a soul. Are we a unitary being of some kind, with psychological "organs" attached, or are we a composite of those organs and nothing else? Are our memories just intellectual entities "somewhere in there"? It seems unlikely, particularly because when we dream, we sometimes smell or hear or feel things both tactilely and emotionally. Most "bad dreams" evoke fear, but there are also happy dreams, angry dreams, erotic dreams, and weepy dreams.
The author devotes one large chapter to "getting out of book", meaning breaking from convention ("Hi", "Hi. How are you?" "Fine, and you?" and so forth). In a five-minute typed "conversation" there is scant time for niceties. Better to jump right in with, "Hi, I'm Brian. What's your favorite hobby?". That'll get the judge into familiar territory and just maybe strike off a wide-ranging conversation that no computer could match. A chatbot that has been extensively programmed for political discourse might hit it off well with the judge who is a political junkie, but fall apart totally when asked, "How 'bout them Astros?"
Just in the past year, a supercomputer was constructed that has processing power and memory capacity both equivalent to a human brain. It won't be long before such a system will fit in a small room rather than a warehouse. How long will it be until the system can feel the wind on its cabinets, hear and see the whippoorwill, smell a rose, tell the taste of chablis from chardonnay (or from a banana!), and laugh at a groaner of a pun? Will a machine system ever be programmed to appreciate NOW NO SWIMS ON MON or "Madam, I'm Adam"? to design a mobile, given a box of small pieces of driftwood and a spool of wire? or at least, to discuss mobiles and Alexander Calder's other art with some felicity? This makes me hanker, just a bit, to try my hand at being a judge…
Mr. Christian concludes that humans are capable of creating "high surprisal" without seeming wholly wonky. Among my friends, this is called "chasing rabbits": one subject leads to another, and the conversation may range over a dozen subjects in a few minutes. This level of flexibility in a chatbot usually comes off seeming schizophrenic, particularly in one like Parry, which simulates the word salad produced by some paranoid schizophrenics. In others it just seems evasive. There is a technique to changing the subject, and most humans do so rather well. As always, too much of one thing or too much of another will give a machine away. Most of us are better balanced than that.
I found it highly entertaining as the author described some of his internal condition during the competition: his worries, his swings from panic to smugness and back. Could the judge detect this in him? Perhaps. And just perhaps, it was the sway of emotion oozing through the teletype-like interface that resulted in his being judged the "Most Human Human" for 2009.