TITLE: A Reflection on Alan Turing, the Turing Test, and Machine Intelligence
AUTHOR: Eugene Wallingford
DATE: March 30, 2012 5:22 PM
DESC:
-----
BODY:
In 1950, Alan Turing published a paper that launched the
discipline of artificial intelligence,
Computing Machinery and Intelligence.
If you have not read this paper, go and do so. Now. 2012
is
the centennial of Turing's birth,
and you owe yourself a read of this seminal paper as part
of the celebration. It is a wonderful work from a wonderful
mind.
This paper gave us the Imitation Game, an attempt to replace
the question of whether a computer could be intelligent by
withn something more concrete: a probing dialogue. The
Imitation became
the Turing Test,
now a staple of modern culture and the inspiration for
contests and analogies and
speculation.
After reading the paper, you will understand something that
many people do not: Turing is not describing a way for us
to tell the difference between human intelligence and machine
intelligence. He is telling us that the distinction is not
as important as we seem to think. Indeed, I think he is
telling us that there is no distinction at all.
I mentioned in
an entry a few years ago
that I always have my undergrad AI students read Turing's
paper and discuss the implications of what we now call the
Turing Test. Students would often get hung up on religious
objections or, as noted in that entry, a deep and a-rational
belief in "gut instinct". A few ended up putting their
heads in the sand, as Turing knew they might, because they
simply didn't want to confront the implication of
intelligences other than our own. And yet they were in an
AI course, learning techniques that enable us to write
"intelligent" programs. Even students with the most diehard
objections wanted to write programs that could learn from
experience.
Douglas Hofstadter, who
visited campus this month,
has encountered another response to the Turing Test that
surprised him. On his second day here, in honor of the
Turing centenary, Hofstadter offered a seminar on some ideas
related to the Turing Test. He quoted two snippets of
hypothetical man-machoine dialogue from Turing's seminal
paper in his classic
Gödel, Escher, Bach.
Over the the years, he occasionally runs into philosophers
who think the Turing Test is shallow, trivial to
pass with trickery and "mere syntax". Some are concerned
that it explores "only behavior". Is behavior all there
is? they ask.
As a computer programmer, the idea that the Turing test
explores only behavior never bothered me. Certainly, a
computer program is a static construct and, however
complex it is, we can read and understand it. (Students
who take my programming languages course learn that even
another program can read and process programs
in a helpful way.) This was not a problem for Hofstadter
either, growing up as he did in a physicist's household.
Indeed, he found Turing's formulation of the Imitation
Game to be deep and brilliant. Many of us who are drawn
to AI feel the same. "If I could write a program capable
of playing the Imitation Game," we think, "I will have
done something remarkable."
One of Hofstadter's primary goals in writing GEB
was to make a compelling case form Turing's vision.
Those of us who attended the Turing seminar read a section
from Chapter 13 of
Le Ton beau de Marot,
a more recent book by Hofstadter in which he explores many
of the same ideas about words, concepts, meaning, and
machine intelligence as GEB, in the context of
translating text from one language to another. Hofstadter
said the focus in this book is on the subtlety of words
and the ideas they embody, and what that means for
translation. Of course, these are the some of the issues
that underlie Turing's use of dialogue as sufficient for
us to understand what it means to be intelligent.
In the seminar, he shared with us some of his efforts to
translate a modern French poem into faithful English. His
source poem had itself been translated from older French
into modern French by a French poet friend of his. I
enjoyed hearing him talk about "the forces" that pushed
him toward and away from particular words and phrases.
Le Ton beau de Marot uses creative dialogues of
the sort seen in GEB, this time between the Ace
Mechanical Translator (his fictional computer program) and
a Dull Rigid Human. Notice the initials of his raconteurs!
They are an homage to Turing. The human translator, Douglas
R. Hofstadter himself, is cast in the role of AMT, which
shares its initials with Alan M. Turing, the man who started
this conversation over sixty years ago.
Like Hofstadter, I have often encountered people who object
to the Turing test. Many of my AI colleagues are comfortable
with a behavioral test for intelligence but dislike that
Turing considers only linguistic behavior. I am
comfortable with linguistic behavior because it captures what
is for me the most important feature of intelligence: the
ability to express and discuss ideas.
Others object that it sets too low a bar for AI, because it
is agnostic on method. What if a program "passes the test",
and when we look inside the box we don't understand what we
see? Or worse, we do understand what we see and are
unimpressed? I think that this is beside the point.
Not to say that we shouldn't want to understand. If we
found such I program, I think that we would make it an
overriding goal to figure out how it works. But how an
entity manages to be "intelligent" is a different question
from whether it is intelligent. That is precisely
Turing's point!
I agree with Brian Christian, who won the prize for being
"The Most Human Human" in a competition based on Turing's
now-famous test. In
an interview with The Paris Review,
he said,
Some see the history of AI as a dehumanizing narrative;
I see it as much the reverse.
Turing does not diminish what it is to be human when he
suggests that a computer might be able to carry on a rich
conversation about something meaningful. Neither do AI
researchers or teenagers like me, who dreamed of figuring
just what it is that makes it possible for humans to do
what we do. We ask the question precisely because we
are amazed. Christian again:
We build these things in our own image, leveraging all the
understanding of ourselves we have, and then we get to see
where they fall short. That gap always has something new
to teach us about who we are.
As in science itself, every time we push back the curtain,
we find another layer of amazement -- and more questions.
I agree with Hofstadter. If a computer could do what it
does in Turing's dialogues, then no one could rightly say
that it wasn't "intelligent", whatever that might mean.
Turing was right.
~~~~
PHOTOGRAPH 1: the Alan Turing centenary celebration. Source:
2012 The Alan Turing Year.
PHOTOGRAPH 2: Douglas Hofstadter in Bologna, Italy, 2002. Source:
Wikimedia Commons.
-----