chatGPT R Us
first take:
I am a
materialist, by which I mean I believe that in the case of human beings in
particular, there is no substance other than the physical body to
account for our having a soul or mind or spirit – there is no “ghost in the
machine.” Somehow the functioning of this brain and body in the context of a
world and universe, which themselves contain no substance other than what
we are accustomed to characterizing as physical, produces in us the impression
of our having a soul or mind etc. It then becomes a matter of choice or utility
whether we continue to posit the existence of souls and minds etc., albeit understood
as not composed of any distinct substance from the body, or to banish
them from our ontology, except as illusions or hallucinations or delusions.
Now let us consider the remarkable
new AI phenom chatGPT. Apparently the fundamental operation that makes this app
tick is as follows: Compile a humongous body of digitized text – for example, Wikipedia
– and then design a program or algorithm that, when fed a digital verbal string
(the question), will generate text (the answer) based on the probability of
what the next word will be.[1] For example, if you typed
in “and so,” it might spill out “on” if that is the most likely word to follow
“and so.” If you prompted it with “and so on,” it might then produce “and so
forth” as the most likely string to follow “and so on.” And so on and so forth!
That chatGPT is, based on this
utterly simple operation, able to produce the texts we have seen, is remarkable
beyond ordinary belief. It would seem a hoax, like the mechanical chess player
(“Turk”) that really had a person inside it, if not for the authoritative
institutions that stand behind it. No, it is not yet as articulate or
idiosyncratic as an actual human being is capable of being; but it is
nevertheless an amazing achievement. Indeed, its very imperfections can mimic
the typical responses of human beings.
What prompts my writing this essay
(and, to give away the punch line, does so in essentially the same way
that a question prompts an answer from chatGPT) is the thought that what makes
chatGPT seem so remarkable is that it appears to understand any question
I pose to it. For example, here is an exchange I had with it:
“Do
you just quote sources like Wikipedia, or do you compose the wording of your
answers from scratch?”
To which chatGPT responded:
“As
an AI language model, my responses are generated by algorithms that draw from a
vast database of language patterns and knowledge. While I may use sources like
Wikipedia or other online resources to provide accurate and relevant
information, I do not simply quote them verbatim. Instead, I use my language
generation capabilities to compose my responses in my own words, tailored to
the specific question and context.”
As impressive as that answer is, to
me what is most noteworthy is that it feels irresistible to say and feel
that chatGPT must have understood my question in order to produce that
answer. It is beyond my intuitive mind to cognize chatGPT’s
answer as the result of only predicting what word or words follow some other
word or words. That it is the result of that “dumb” process – that
is what is astonishing.
However, my mind is
(presumably) capable of grasping how this trick is accomplished … provided I
were to become immersed in computer stuff. Or at a minimum I can say that I do
not seriously doubt that programming experts can pull off this trick, although
I really would have doubted it had I not “seen it with my own eyes.”
And this is not at all an unusual
phenomenon. In fact just about every modern-day convenience is just as
marvelous. Do I have the slightest idea how a performance of Beethoven’s Ninth
Symphony can be conveyed to me from an orchestra playing on the other side of
the Earth? Yes and no. It seems magical … when I am not simply taking it for
granted. Clearly today’s children will take chatbots far more extraordinary
than today’s chatGPT thoroughly for granted (although, if they have not been
genetically engineered otherwise, they, like me, in a moment of reflection,
will still be susceptible to wondering at it).
Furthermore even natural
phenomena partake of the same wonder. Hume pointed out that cause and effect
can obtain between the most oddly paired things, provided only that we become
used to it. But is it not wonderful that when you let go of something,
it drops to the ground? Is this not a miracle, looked at without (per
impossibile) preconceptions? But Darwin
has shown us that even the intuitively irresistible perception of all of
nature being fitted to every other part of nature, such that it must have
been intentionally made so (“designed”), can itself be deconstructed into a
simple fundamental principle that works without intention (survival of
the fittest).
But
the amazingness of chatGPT’s seeming understanding is still not my punch line. The
truly astounding revelation is that the “illusion” of chatGPT’s
understanding is equivalent to our attributing understanding to human beings
… even to oneself.[2]
For – at least on the materialist view – we ourselves are physical beings
functioning entirely according to natural laws, and hence are, in this sense,
mechanisms exactly like chatGPT. And as a kind of evidence of this, ask
yourself: Do you have any more idea what will come out of your mouth before you
say it, or even into your mind before you think it, than chatGPT does of what
it will write in answer to your question?
Granted,
chatGPT was programmed by intelligent beings, and so is “artificial,” whereas
nobody programmed us, so we are “natural.” So the laws that govern our behavior
are the laws of nature, whereas the laws that govern chatGPT’s behavior (or,
more precisely, the “behavior” of any material object that is running it) are
both natural and intentionally imposed. But, it seems to me, in the most
fundamental way we are the same, in that we both are entirely physical in
composition and (ultimately) governed entirely by laws (if only statistical at
the quantum level).
So here again we have a choice: We
can (to suit our purposes) maintain either that we do not understand
things -- no more than does chatGPT (nor even a more mature chatGPT that truly
will pass any Turing test) – or that we do understand things, but so
does chatGPT. I see this as exactly analogous to choosing between asserting
that we have emotions and so do other animals, or that nothing has emotions.
Some animal experimenters are notorious for claiming or assuming (unless “proven”
otherwise) that we have emotions but other animals do not … and no doubt some
cognitive scientists and philosophers would claim that we have
intelligence and understanding, whereas robots do not and could not. But my
main point in this essay is that chatGPT demonstrates in a particularly
striking way that this dichotomy is purely tendentious, and that it’s really
all or nothing: Either everything that appears to understand does understand,
or nothing does. In other words, I am arguing that if it looks like a duck and
sounds like a duck, then it is a duck.
By now you have probably succumbed to your curiosity and
given chatGPT a try. And you have been as amazed as anyone. In no time flat it
begins to type out an answer to whatever query you have posed, and that answer
appears wholly intelligent. You can even carry on a conversation by following
up your initial question with a second one in response to chatGPT’s answer,
which it will then answer in turn, and so on indefinitely. It seems as if
chatGPT or a successor chatbot stands a good chance of passing the Turing Test,
which is to say, it may soon be indistinguishable from a human interlocutor.
After all, even now would you be able to tell if this very article
is being written by chatGPT? Indeed, and ironically, my own opinion (assuming
you believe I am the human being who appears to be authoring this article) is
that the most likely reason a chatbot would fail the Turing Test is that
it will have a better command of “proper” English (or whichever
language) and sound logic and the knowledge base than the average human speaker
or writer or thinker.
Nevertheless,
most of the experts who have written about this remarkable new product of AI
have assured us that chatGPT is not, and could never be, intelligent. The
reason is that this chatbot functions differently from a human brain, and in
fact by the “mindless” following of an algorithm or set of rules. It is nothing
but a program running inside a piece of hardware somewhere, and hence its
appearance of intelligence is just that – an appearance. No matter how closely
a simulacrum matches the real thing, it ain’t -- no more than a perfect replica
of the Declaration of Independence is the actual Declaration of Independence
that John Hancock signed.
That
argument is persuasive. However, I do not consider it definitive. Let me
present a very simple counter-argument. Return to your original moment of
amazement, as chatGPT begins to type out its answer to your first question. Now
pause to reflect. When you typed your question, were you doing anything less
amazing? Most likely you did not mentally rehearse your question before typing
it on your keyboard. Your fingers began to tap, and right before your very eyes
your question appeared on the screen in front of you in exactly the same way
chatGPT’s answer did. Both you and chatGPT drew upon some unknown source to
produce your respective question and answer. Even if you had mentally
rehearsed your question beforehand, the same is true: There was no
pre-rehearsal rehearsal, for that would lead ultimately to an infinite regress.
We could also say that both of you
relied on the functioning of a physical object: in the case of chatGPT, a
computer somewhere in the Cloud, and in your case, your brain. I still call
these sources “unknown” because you yourself (unless you are a computer
engineer or a brain physiologist) have very little conception of what produced
the words on the screen. And in fact, the illusion of an intelligence typing
the answer is matched by an illusion of you typing the question -- that
is, if you conceive of yourself as an intelligent mind or soul or self, and not
just a(n admittedly remarkable) mass of brain tissue.
We
attribute all sorts of magical qualities to our “self,” such as intelligence
and free will. But, if you are a materialist anyway, you believe that it’s
really the brain that is running the show, and the brain works according to
deterministic laws like all the rest of the physical universe. (Note: At the
quantum level the laws may be statistical, but lawful they remain.) Granted, the particular way in which the
brain functions is different from the way chatGPT functions. But does a
difference at the micro level always signify a difference at the macro level? A
fly’s eyes function differently from a human’s eyes, but do we hesitate to call
both eyes, and say that both flies and humans see? No, we do not. So are
the experts really justified in denying intelligence to chatGPT (or its
successor) just because its behavior is produced by a different mechanism from
our own? My argument is that it is a matter of convenience or utility whether
we designate chatGPT or human beings as intelligent. This
position is sometimes called instrumentalism: We attribute qualities to thing
based on the usefulness of doing so.
Here is another way to think about
this. The philosopher Ludwig Wittgenstein famously argued that the meaning of a
word derives from its use. This hit home for me the other day when, once again
(like now), I was typing away on my keyboard – this time replying to an email from
a colleague – and found myself typing the word “nominal” in a sentence. For
years I have been hazy on the meaning of that word, but have come across it now
and again. My colleague, who is an engineer, had used it some time previously
in a technical discussion we were having, and I had taken the opportunity to
ask him what it means. He gave me an explicit definition. But I must admit it
still left me uncomprehending.
Lo and behold, there I was
spontaneously using the word myself – its coming, as if “ex nihilo,” just like
all the other words I wrote. And I was also confident in my usage. How
had this happened? I surmise that I (or my brain), after years of exposure to
the word being used by others, must finally have crossed some threshold and grokked
the meaning of the word. I had picked it up by a kind of osmosis. And is this
not precisely analogous to how chatGPT comes up with its sentences, through
exposure to a mass of data input? And now I can push the point further and
argue that it suggests not only that intelligence and the self are (if we find
it useful to think of them this way) illusory, but so is meaning. Or,
like Wittgenstein, we can choose to “retain” the notion of meaning as something
real, but analyze it as nothing but how we use a word. There is
nothing “deeper” here. So why insist there must be for chatGPT? Once again: It
is a matter of choice or convenience or utility how we decide what trait to
assign to something. Reality itself is mind-dependent in this way.
Still, the experts will object that
surely chatGPT is not conscious (and surely we are). And I agree. But my
reason is the same for nevertheless ascribing intelligence to
chatGPT. I, and I suspect you, find it intuitively irresistible to
imagine we are conversing with an intelligent being when we use chatGPT. But we
balk at attributing consciousness to a computer. Now, that may be
a mere prejudice; and if chatGPT were hooked up to a metal or plastic body with
a fully human face, my intuition about that could change too. Alternatively
(again, it is a matter of choice) I am very sympathetic to philosopher Daniel
Dennett’s notion of consciousness as itself an instrumental illusion. But even
if we deem it something real, now that I understand verbal meaning to be a
matter of how we use words, what is so important about being conscious
of that meaning? Does my consciousness add anything to my ability to type the
very words I am typing now? I doubt it. So how could the denial of
consciousness to chatGPT prove detrimental to the attribution of intelligence
to it?
I conclude that chatGPT, despite its obvious differences from us, might nevertheless be intelligent.
[1] “ChatGPT generates realistic responses by making guesses
about which fragments of text should follow other sequences, based on a
statistical model that has ingested billions of examples of text pulled from
all over the internet.” (Benjamin Weiser, “Here’s What Happens When Your Lawyer Uses ChatGPT,” The New York Times,
May 28, 2023)
[2] I am, in effect, offering a counterargument to Searle’s famous Chinese Room.