How can one refute John Searle's "syntax is not semantics" argument against strong AI?

Question

There are many refutations of John Searle's Chinese Room argument against Strong AI. But they seem to be addressing the structure of the thought experiment itself, as opposed to the underlying epistemic principle that it is trying to illustrate.

This principle is "syntax is not semantics" (See these lectures by John Searle): At the end of the day, computer software, even the most advanced AI conceivable, manipulates symbols according to a set of syntactic rules, regardless of their meaning.

Anybody who has studied formal logic knows that rules like De Morgans laws or the laws of idempotency ( e.g. A ^ A = A ) are independent of the meaning of the symbols being processed.

This idea, that syntax is independent of semantics, and therefore a computer can function perfectly without ever knowing the meaning of what it is computing seems like a much stronger argument against AI, and Mind-Body functionalism in general, than Searle's original Chinese Argument.

What are the main refutations advanced by proponents of functionalism and strong AI specifically of the "syntax is not semantics" argument?

Answer 1

The "syntax is not semantics" principle in the Chinese Room Argument (CRA) is based on the relationship between the Searle-computer and the Chinese symbols. Searle correctly characterizes this as a formal symbol processing relationship wherein the Searle-computer manipulates the symbols purely syntactically, according their shapes alone, without doing any subjective interpretation of them. This formal relationship is the linchpin of the CRA and Searle's rebuttal of computationalism (aka, computational functionalism, Strong AI).

Turing machine (TM) theory explains why this "linchpin" is merely a special case, and it exposes the huge gap in reasoning that comes from ignoring the most important part of the picture: the program. For example, the theory highlights this telling discrepancy:

If computers lack internal semantics, then why must the Searle-computer's programs be in English?

Searle never addresses this significant inconsistency in his position.

The Searle-computer is fully programmable, hence it is a universal TM (UTM). Every UTM has a two-part input: (1) a program and (2) a "nominal input" for the program to process. For example, if given program ADD, for addition, and nominal input "3, 4", the Searle-UTM would output "7". Because the digits "0-9" are just formal symbols to the Searle-UTM, they could be encoded as Chinese characters, and the Searle-UTM would still perform addition--just like the CRA. However, the same is not true for the Searle-UTM's other input, the ADD program. If it were written in Chinese, for example, then the Searle-UTM would fail.

Notice that the Searle-UTM can correctly process the Chinese symbols (#2) on a purely formal (syntactic) basis only because it also has a program input (#1) that is actually responsible for determining what to do with them. The program--not the Searle-UTM--determines how the Chinese symbols are actually processed, so the Searle-UTM need only manipulate them formally, acting as the program's vehicle or "middleman".

On the other hand, the Searle-UTM is the only thing responsible for correctly processing the program itself. The Searle-UTM must causally connect the program symbols with the physical entities and processes that they represent—a non-formal process that realizes symbolic representations as specific real events. Thus, the formal symbol processing that is Searle's linchpin, is just a consequence of the special relationship a UTM has to its nominal (formal) input, which is mediated by a program input that is processed non-formally by the UTM.

"It's the program, stupid!"

Q: What is a program?
A: It is the specification of how some TM works--a kind of blueprint for instantiating a TM that typically is non-universal.

Q: What happens when a UTM runs a program?
A: Two significant TM computations occur: (1) the universal computation instantiates (2) the computation of the program's TM. (The Searle-UTM can only introspect on the first one, his own universal computation, which entails reading the program instructions in English and executing them.)

Q: What, if anything, is happening semantically inside the Chinese Room?
A: We don't know because we don't know how the program works. It is useless to ask the Searle-UTM because he doesn't know either. He doesn't know if his program is doing a Chinese Turing Test or tic-tac-toe. He only knows about his own universal algorithm: "read the program and execute its steps on the nominal input". To know the nature of the computation responsible for the externally observed behavior, the only thing that matters is the program, and it is left unspecified.

Searle completely ignores or dismisses the second TM computation, which arises from the program. Nevertheless, its existence is a mathematical fact, not just some philosophical assertion. It does not depend on anyone's subjective opinions or intuitions. It only requires an objective understanding of how UTMs work. This also explains why the Systems/Virtual-Mind reply has been the most popular kind of CRA rebuttal: http://www.scholarpedia.org/article/Chinese_room_argument#The_systems_reply

Searle's obsessive focus on a quirk in the nature of UTMs is somewhat understandable because UTM-computers and programs are so iconic in our culture. In philosophical discussions, however, it is crucial to focus instead on TM theory itself and on general TMs, not just UTMs. The failure to do so means that the CRA fails miserably as a refutation of computationalism while spawning decades of fruitless debate in the process.

UTMs vs. General TMs

Q: Aren't UTMs "universal" (i.e., representative of all TMs)?
A: While a UTM can instantiate any other TM via its program, a UTM's own internal algorithm and specialized input for this universal programmability is highly specific and not at all representative of TM computation in general. Focusing on this as the CRA does is an unhealthy distraction.

Q: What can a general TM do that a UTM can't?
A: A UTM must always act as a rote machine. It must faithfully ensure that the same given program will function the same way every time. In general, a non-universal TM could change it's own behavior over time based on its input-output history.

For a more complete explanation and discussion of this topic, see the articles provided here: http://www.chineseroom.info/

EDIT: A later version of this explanation is here, several reply-levels down: https://www.reddit.com/r/askphilosophy/comments/50igj8/if_you_could_chat_with_john_searle/

Answer 2

Wittgenstein in his intermediate period provided a response, before the age of AI research and Searle's objections. In a nutshell: semantics is another syntax. Words only mean as role players in a linguistic calculus, and their meaning reduces to the collection of rules governing their use in the calculus. Of course, he was thinking of mathematics and language at large rather than computers. Here is Wittgenstein on metamathematics as "semantics" of mathematics:

"What Hilbert does is mathematics and not metamathematics. It is another calculus just like any other. I can play with chessmen, according to certain rules. But I can also invent a game in which I play with the rules themselves. The pieces of my game are now the rules of chess, and the rules of the game are, say, the laws of logic. In that case I have yet another game and not a metagame... What is known as the ‘theory of chess’ isn’t a theory describing something, it’s a kind of geometry. It is of course in its turn a calculus and not a theory".

What Wittgenstein came to appreciate later, in Philosophical Investigations, is that realistic "language games" are not reducible to calculi, they are far too nuanced for that. But that did not mean reinstatement of "intentionality" and "meanings" as entities, it meant that even the "rules" are not entities that can be spelled out. "Meaning" is acquired in activity, linguistic practice. In a very different form and by a very different route the same conclusion was arrived at by others, who came to play an unexpectedly prominent role in AI research. Dreyfus, the perennial critic of what computers can do since 1960s, gives a very interesting account in Why Heideggerian AI Failed and how Fixing it would Require making it more Heideggerian:

"Using Heidegger as a guide, I began to look for signs that the whole AI research program was degenerating. I was particularly struck by the fact that, among other troubles, researchers were running up against the problem of representing significance and relevance – a problem that Heidegger saw was implicit in Descartes’ understanding of the world as a set of meaningless facts to which the mind assigned what Descartes called values and John Searle now calls function predicates. But, Heidegger warned, values are just more meaningless facts... One version of this relevance problem is called the frame problem. If the computer is running a representation of the current state of the world and something in the world changes, how does the program determine which of its represented facts can be assumed to have stayed the same, and which might have to be updated?

Merleau-Ponty’s work, on the contrary, offers a nonrepresentational account of the way the body and the world are coupled that suggests a way of avoiding the frame problem. According to Merleau-Ponty, as an agent acquires skills, those skills are “stored”, not as representations in the mind, but as a bodily readiness to respond to the solicitations of situations in the world. What the learner acquires through experience is not represented at all but is presented to the learner as more and more finely discriminated situations..."

Agre, Brooks, Wheeler, Winograd and other big names in AI eventually came to assimilate what Dreyfus was selling on behalf of Heidegger and Merleau-Ponty. This is now called "embodied-embedded cognition", and "Heideggerian AI" is a term of art too. Even Dennett's Cog incorporated some of these insights, although that did not save it. So Searle is right, semantics is not syntax, but he is unlikely to like the conclusions AI researchers drew from it. Namely, that meanings are not representational entities mysteriously connected to the real world, as Descartes would have it, and intentionality is not a special goo exuded by organic systems, as Searle would have it, but dynamic effects emerging in the process of interacting with environment, including other actors. Semantics is not syntax because meaning and intentionality are prerogatives of active players. Computer can not be such a player, it has to be an AI robot of some sort. In a way, this attitude shows in more recent versions of the Systems and Robot replies to the Chinese room.

Whether this "embodied-embedded intentionality" works out remains to be seen. As Dreyfus's title suggests, it is a work in progress, and he charges that all existing implementations are still too representational.

Answer 3

You ask the obvious question. "If semantics is not syntax, then what is it?"

If they two are truly separate, you have a terrible difficulty explaining how semantics is teachable. You either end up in some kind of mandatory idealism where the basics of meaning needed to bootstrap semantics are already 'there', or with a functionalist model like Wittgenstein, Desassure, or Lacan.

In the latter case there is not syntax and semantics, there is just a continuum of semiotics with two unreachable ends. (A 'cuter' way to put the case @Conifold makes. So I am not going to bother repeating the reasoning here.) Semantics is just the syntax of behavior in general, rather than the syntax of a specific narrow range of behaviors you do with your vocal apparatus, with text, or with gestures. If your distinction is a spectrum, and pure forms of neither extreme are real, the argument can no longer be made.

In the former case, the one more consonant with Searle, the argument becomes much larger, and takes too many forms to kill them all off at once. But in most forms of idealism that allow for a basic internal structure to the mind independent of its function in reality, semantics is not real, either.

The meaning in the form it can occupy the mind is real, and it is meaning, not semantics. And the connection between minds that transfers meaning through behavior is real. And it is, being made of behavior, syntax, not semantics.

Intelligence, then, as it is functionally displayed through behavior by humans, is just this generalized syntactic wrapper around an essentially different process. Maybe one cannot artificially reproduce that process, but that is a different statement. There is no reason why the wrapper itself, intelligence, cannot exist without the customary stuffing of mind and will.

Answer 4

I believe the exact sentence by Searle in the Chinese Room paper was 'syntax is not enough for semantics'. He made the meaning of the sentence more precise by proposing a case in which the syntax of a language is perfectly operated without any semantical comprehension arising from the process. Now, semantics in Searle's sense is some mental comprehension and it is rather obvious that it is possible to operate the syntax of a language without a mental glimpse of its meaning emerging from the operation. Therefore, in Searle's sense, syntax is indeed not enough for semantics.

Answer 5

This idea, that syntax is independent of semantics, and therefore a computer can function perfectly without ever knowing the meaning of what it is computing seems like a much stronger argument against AI, and Mind-Body functionalism in general, than Searle's original Chinese Argument.

there's mysticism here. a simulation of a human is a human. it's the same thing. human minds are software running in a classical computer. that's what you already are – a computer running software. the hardware details don't matter to the computations. no soul or organic molecules required.

an intelligence software program has to do certain things. included on the list is create knowledge. the only known knowledge-creating process is evolution. knowledge can be created through replication with variation and selection. (in the case of ideas this would more normally be called brainstorming and critical thinking to eliminate errors). software that does this within various parameters, and does a few other things, would be a thinking person. that's all there is to it. stuff like emotions are emergent properties of software, they aren't tied to souls, hardware made of organic molecules instead of silicon, etc