More than 60 years ago Alan Turing started the debate about
artificial intelligence when he published a paper titled, "Computing
Machinery and Intelligence." In it, he suggested that machines
could be tested for intelligence, in a process that is now known
as the Turing Test. The term "artificial intelligence,"
or "AI" was coined by John McCarthy a short time later,
and defined as "the science and engineering of making intelligent
machines," or "intelligence as exhibited by an artificial
(man-made, non-natural, manufactured) entity." By 1956 AI
became a distinct field of scientific study.
Predictions at that time that AI would become a reality by
the year 2000 were a bit optimistic. Computers can now beat the
best chess masters on the planet and do many other amazing things,
but they have little of what we call "comprehension."
The best computer still cannot handle a simple conversation with
us in a way that makes it appear as intelligent to the average
human. That, by the way, is more-or-less the basis of the Turing
test.
Turing started out with the question, "Can machines think?"
"Thinking" is difficult to define in a way that all
scientists could agree upon, so he changed the question to, "Are
there imaginable digital computers which would do well in the
[Turing test]"? One test he proposed was to have a human
judge carry on a natural conversation with a computer and another
human without being able to see them, while both try to "convince"
the judge that he or it is human. Since this is a test of intelligence
and not the ability to sound human, today we would do this by
keyboard and computer screen, perhaps in an online chat room.
The computer passes the test if the judge can't reliably and
consistently distinguish it from the real human.
So far, no luck. In the few tests that computers have "passed,"
they have done so only because the judges have been limited in
topics or the type of questions they could ask, or there were
other deviations from a true conversation test.
A program called ELIZA, and another called PARRY have passed
the Turing Test according to some, but this is disputed by most
scientists. In the tests with ELIZA, for example, 33 psychiatrists
were only able to identify the human from the program about half
of the time - the same as random chance would yield, but they
were judging from transcripts of conversations with the program,
and not from their own conversation in which they could ask what
they wanted.
As far as the evidence that I have seen goes, no machines
or programs have been able to pass a true natural conversation
test consistently. Now even the optimistic scientists have pushed
back their predictions for true AI to 2029 and beyond. Many skeptics
and those with a dualist view of the mind (believers in the idea
that the mind is at least partly non-physical and separate from
the brain), doubt it will happen at all.
Here's a short video I found on an anthropomimetic robot,
which it's creator thinks is a step toward AI:
I'm not sure what is meant by the statement that "without
a body artificial intelligence cannot exist." It seems that
any intelligence has to be exercised through some physical form,
but no particular reason it has to be a body like ours (although
I agree that the body itself shapes the way we think).
My own opinion is that we might see artificial intelligence
in this century. But I wonder about just how consistently any
computer will be able to pass that basic test of "humanness."
I can easily imagine a computer that can carry on a conversation
and fool people into believing it is another human. The problem
is, I can also easily imagine finding questions that will reveal
it as a computer.
Specifically, I would ask my conversation partners about matters
that involve the balancing of values. Classic moral dilemmas
come to mind. If you have time to do only one thing, do you save
your own child who is laying on the train tracks in front of
an oncoming train, or save the four stranger's children? A human
usually has some doubts about either which is the right course
of action or at least which he or she would really take, and
that would show in the conversation. Would the computer be convincing
enough in its moral struggling with an issue like this? Of course
it could be programmed for the classic dilemmas like this, but
there are many other ways to approach this questioning.
How would a computer "value" things? We consider
our pain, our future benefit, our stated moral beliefs, and all
sorts of other factors when thinking about what we value or what
we would like to see or do. These factors could not mean much
to a computer, and that lack of significance seems likely to
show. I suppose a computer that is designed for self-preservation
at least would have that as a basis for valuing this or that
idea or course of action, but that isn't quite the same. In fact,
one crucial difference is obvious here: a human sometimes values
some outcome more than his or her own life.
This is where we have to start questioning the validity of
the Turing Test as a test of artificial intelligence. A computer,
after all, even if it has what we would recognize as consciousness,
would not necessarily be indistinguishable in its thinking from
a human. Consider the fact that in a conversation in a chat room
we might quickly tell the difference between a conservative and
a liberal, or a shy and an extroverted person. Intelligence is
there in people of all categories even though we can create and
recognize categories. Certainly, then, a computer that has become
intelligent might still be very different from humans.
At this point in history, we are still facing the original
problem that Turing faced when designing his test, which is that
of defining intelligence. After all, we are not testing to see
if computers can do amazing things - they have passed that test
many times in many ways. We are asking whether they can truly
"think" about what they are doing - or about what they
"should" be doing, or "could" be doing.
These questions of machine intelligence are still seriously
debated in the scientific community. In 2010 the "Towards
a Comprehensive Intelligence Test" symposium was held at
De Montford University in the United Kingdom to address some
of them.
Here's another question to consider about AI: Is there a related
possibility we might refer to as "artificial wisdom?"
