Turing, And The Bogus Rivalry With Machine-Based Intelligence

IBM's Watson Wikipedia.org
IBM’s Watson           Wikipedia.org

In reality, humans have nothing to fear. Most measures of artificial intelligence use the wrong yardsticks.

We are awash in articles, books and films about the coming age of “singularity:” the point at which machines will supposedly duplicate and surpass human intelligence.  For decades it’s been the stuff of science fiction, reaching perhaps its most eloquent expression in Stanley Kubrick’s 1968 motion picture, 2001: A Space Odyssey.  The film is still a visual marvel.  Who would have thought that Strauss waltzes and images of deep space could be so compatible?  Functionally, the waltzes have the effect of juxtaposing the familiar with a hostile void, making the film a surprising celebration of all things earth bound.  But that’s another story.

The central agent in the film is HAL-9000, the controlling computer that begins to turn off the life support of the crew during a long voyage, mostly because it “thinks” the humans aren’t up to the enormous task facing them.

Kubrick’s vision of very smart computers is also evident in the more recent A.I., Artificial Intelligence (2001), a project started just before his death and eventually brought to the screen by Steven Spielberg.  It’s a dystopian nightmare. In the film intelligent “mechas” (mechanical robots) are generally nicer than the humans who created them.  In pleasant Haddonfield New Jersey, of all places, they are shot on sight for sport.

Fantasies of machine intelligence have lately given way to IBM’s “Big Blue” and “Watson,” mega-computers with amazing memories and—with Watson—a stunning speech recognition capability. And then there’s the Turing Test, the challenge of the World War II code-breaker Alan Turing which has been turned into a recurring world competitions.  The goal is to construct a “chatterbot” that can pass for a human in blind side-by-side “conversations” that include real people.  Can a machine pass for a person?  For artificial intelligence engineers, the trick is to fool a panel of questioners at least 30 percent of the time over 25 minutes. According to the BBC, a recent winner was a computer from the University of Reading in the U.K. It passed itself off as a Ukrainian teen (“Eugene Goostman”) speaking English as a second language.  O.K.: maybe we’re not at the point of singularity yet.

In actual fact, humans have nothing to fear.  Most measures of artificial intelligence use the wrong yardsticks.  These computers are never embodied.  The rich information of non-verbal communication is not present, nor can it be.  Proximate human features are not enough.  Watson’s “face” in its famous Jeopardy challenge, for example, was a set of cheesy electric lights.  Moreover, these smart machines tend to be asked questions that we would ask of Siri or other informational databases.  What they “know” is often defined as a set of facts.  And, of course, these machines lack what we so readily reveal in our conversations with others: that we have a sense of self, that we have an accumulated biography of life experiences that shape our reactions and feelings, and that we want to be understood.

Just the issue of selfhood should remind us of the special status that comes from living through dialogue with others.  A sense of self is complicated, but it includes the critical ability to be aware of another’s awareness of who we are.  If this sounds confusing, it isn’t.  This process is central to all but the most perfunctory communication transactions.  As we address others we are usually “reading” their responses in light of what we believe they already discern about us.  We triangulate between our  perceptions of who we are, who they are, and what they are thinking about our behavior. This all happens in a flash, forming what is sometimes called “emotional intelligence.”  It’s an ongoing form of self-monitoring that functions to oil the complex relationships. Put this sequence together, and you get a transaction that is full of feedback loops that involve estimates if intention and interest, and—frequently—a general desire to protect the feelings of the other.

It’s an understatement to say these transactions are not the stuff of machine-based intelligence, and probably never can be.  We are not computers.  As Walter Isaacson reminds us in his recent book, The Innovators, we are carbon based creatures with chemical and electrical impulses that mix to create unique and idiosyncratic individuals.  This is when the organ of the brain becomes so much more: the biographical homeland of an experience-saturated mind.  With us there is no central processor.  We are not silicon-based. There are the nearly infinite forms of consciousness in a brain with 100-billion neurons with 100-trillion connections.  And because we often “think” in ordinary language, we are so much more—and sometimes less—than an encyclopedia on two legs.

Comments: Woodward@tcnj.edu