Thinking Machines

In his classic article “Computing Machinery and Intelligence” Alan Turing lays out what is now known as the Turing Test, a means by which one can, supposedly identify intelligence in another being, regardless of what external qualities this being may or may not have. John Searle, however, finds the Turing Test to be not only insufficient, but entirely misguided altogether, at least in the way in which most would implement it. It will be the purpose of this paper to briefly describe both the Turing Test as well as Searle’s Chinese Room argument, which is intended as a refutation of the former. I will also very briefly consider some of the objections which have been brought against Searle’s Chinese Room argument and conclude that all of these objections are largely irrelevant to Searle’s main point. In the end, I will argue that Searle is correct in calling the Turing Test as an unqualified test of intelligent thought into serious question.

The primary aim of Turing’s paper is to limit or even completely do away with the intuition that something must be human-like if it is to be considered intelligent in any meaningful way. Before the age of Sci-Fi thrillers it was, and still is to a large extent, “natural” to define humans as the paragon of intelligence, they being “above” all other animals and just below the gods. It has been rational thought which has been the essence of humanity since the time of Aristotle. To be human was to be rational and intelligent, which is more than enough to give most people the intuition that the converse must also be true as well: to be intelligent is to be human or at least human-like.

It is in this context that Turing argues quite convincingly that the degree to which a being (be it biological or artificial) is human is largely irrelevant to how intelligent is it. It is for this reason that the intelligence of a being should be judged on its own terms, isolated from any other features which will only serve to manipulate our intuitions in one direction or another. The way in which this is to be accomplished, according to Turing, is by way of an imitation game, wherein the being whose capacities for thought are in question is placed in a separate room while we are allowed to communicate with it by way of written messages. We are able to ask it any question or interact with it in any way which we may desire just so long as it is done by way of written communication. Turing’s main point is this: if under such conditions we are unable to distinguish a machine placed in the other room from another human placed in the same situation, by what reason are we to deny that such a machine has the capacity for thought? Turing’s answer to this rhetorical question is, of course, none; the machine has not merely simulated intelligent thought, but has actually engaged in intelligent thought.

John Searle disagrees, and I will argue that the disagreement lies largely in the fact that while a machine should not have to be human-like at all, there are some features and characteristics which a machine must have, characteristics which human do in fact have, in order for either to have intelligent thought. What Searle sees as being the central problem with the Turing Test is this, syntax is not sufficient for semantics. A Turing machine is simply manipulating formal symbols which have absolutely no meaning which is intrinsic to them. Accordingly, the processing which the Turing machine is engaged in is simply not meaningful.

This is what the Chinese Room argument (it’s actually more of a thought experiment than an argument) is supposed to illustrate. Suppose that you are locked in a room, just as a Turing machine would be in a Turing Test, into which you are only allowed to communicate with an interlocutor by way of written communication. Suppose further, however, that your interlocutor is Chinese and all the symbols which you receive as well as send are only written in what are to you utterly meaningless Chinese symbols. In order to complete the comparison, you are equipped in the room with an instruction book which allows you to follow the immensely huge list of instructions on what it is that you are supposed to “output” in response to all possible Chinese “inputs.” Thus, in the Chinese Room you have become the Turing machine, complete with a tape with Chinese input/output symbols and a machine table to consult. The entire point of Searle’s thought experiment is to show that even if you were able to pass this Chinese Turing Test, the simple fact of the matter is that you would not yourself understand a single word of the Chinese “conversation” in which you had just taken part. You had all you needed by way of syntax, but this did not provide you with a single bit of semantic content, and the exact same thing can and should be said of any Turing machine which is itself simply manipulating formal symbols.

Searle’s argument has not gone unchallenged by any means. The problem with most responses, however, is that they ignore Searle’s actual argument (that syntax is not sufficient for semantics) in favor of attacking his Chinese Room thought experiment. The most common response, one which I myself found persuasive for some time, is what Searle calls “the system’s reply”. The reply states that you, while locked in the room, do not yourself know Chinese, you combined with the instruction book do know it; it is the system which knows Chinese, not you. But this plays directly into Searle’s line of thought, for the instruction book is simply the programming of the system, it is the smuggling of another mind (namely that of the books author) into the system. Such a response only attempts to establish that the programmers of the Turing machine are intelligent rather than the machine itself. While the symbols certainly have semantic content to the programmers in the case of the Turing machine, or to the books author in the case of the Chinese Room, this does not change the fact that such Chinese/formal symbols are completely meaningless to me/the machine.

Another very popular reply, if only because it gets repeated in many popular publications, is that Searle’s Chinese Room is simply a thought experiment which is meant to manipulate our intuitions. Indeed, Daniel Dennett goes so far as to call Searle a “conjuror” who is using a “trick” to pump our intuitions in a misguided direction. His response is simply to reassert that if a machine could pass a really hard Turing Test by engaging and explaining a slightly off-colored political joke, then our intuitions are back in line with where they should be: the machine is acting intelligent and therefore must actually be intelligent. Steve Pinker’s response is similar in that he asserts that Searle is merely slowing down the process of manipulating symbols in order to guide our intuitions. Such, however, is simply beside the point. Let us suppose that we have an entire army of English-speakers locked in a Chinese Room who jointly have the entire instruction book memorized. Furthermore, let us say that they are able to perform their task at blinding speeds. Let us juice up the story all we want with greater speed and power but the point is this: neither speed nor power is sufficient to overcome the syntax/semantics gap, and it is this gap which distinguishes mere acting intelligently from actually being intelligent.

As an illustration of what the machine is lacking, consider the following question: what does the word “horses” mean to a Turing machine? Of course the machine can be programmed to respond to such an inquiry and thus behave as if “horses” meant something to it, but in such a case we are supplying all the semantic content to such a conversation, not the machine. In this we can see the Behaviorist influences in Turing’s test, for intelligence is defined by him as consisting in behaving intelligently and nothing else. But this is exactly where Searle sees him as going wrong, for even though a machine can be trained to behave as if it were engaging in intelligent thought, this would not, contrary to Turing’s rhetorical question, actually amount to intelligent thought.

To the machine, “horses” is simply a symbol, period. To us, “horses” is a symbol which means something in virtue of its referring to horses. The machine has no idea what horses are, but is instead engaged in the meaningless shuffling of symbols, a pure language game without any content at all. What the machine lacks is indexicality, the ability to reference the symbolic system (which it certainly has) to the actual world of objects which the symbols are supposed to refer to. Consequently, I argue that in order for a machine to engage in intelligent thought as we do, especially in terms of language use, it must interact with the world of objects in a way parallel to the way in which we do. Turing was right in that an intelligent machine does not have to do so in a way which is similar to the way in which we do, but to suggest that such a machine need not interact with the world of objects in any way at all beyond the mere passing of symbols back and forth is simply untenable.

In conclusion, Turing’s test for intelligent thought was useful in overcoming the strong intuition that intelligence and humanness are necessarily linked. His affirmative claim, however, that behaving as an intelligent thinker is sufficient for actually being an intelligent thinker is flawed. The gap between a machine behaving intelligently and being intelligent consists, at minimum, in the gap which exists between syntax and semantics. It is not through more computing power or speed in which this gap is to be bridged, but rather by way of environmental interaction. It is indexicality which the machine lacks, not proper programming.

Bibliography

Terrence Deacon: The Symbolic Species (W. W. Norton & Company Ltd. 1997)
Daniel Dennett: Consciousness Explained (Little, Brown & Company 1991)
Steve Pinker: How the Mind Works (W. W. Norton & Company Ltd. 1997)
John Searle: “Minds, Brains and Programs” as reprinted in Minds, Brains and Computers: The Foundations of Cognitive Science (Blackwell Publishers Ltd. 2000)
Alan Turing: “Computing Machinery and Intelligence” as reprinted in Minds, Brains and Computers: The Foundations of Cognitive Science (Blackwell Publishers Ltd. 2000)