The Chinese Room Argument revisited

In the midst of 2015’s debates on the risks and possibilities of artificial intelligence which were fueled by the progress of self-driving cars, autonomous robotics, and the signature of an open letter in favor of AI security, Google invited philosopher John Searle to elaborate once more on his Chinese Room Argument of 1980.

The Chinese Room Argument was designed to demonstrate the impossibility of creating true intelligence out of computer code by comparing it to a hotel room that contains detailed instructions on how to respond to sequences of Mandarin characters. While a person who follows these instructions impeccably could lead an observer to believe that a dialogue between two Chinese speakers is taking place, the user of these instructions would in fact not need to understand a single character. Therefore, Searle argued, the Turing Test is inadequate, as it does not take true understanding of the input into account, but rather judges only the quality of the output. At its core, the Chinese Room Argument defends the notion of intelligence as a unique property of biological entities, something machines can at best simulate but never replace.

For an in-depth discussion of the thought experiment and its reception over the last three decades, see e.g. the Stanford Encyclopedia of Philosophy.

  • Date of recording: Mon, 2015-11-23
  • Language(s) spoken:

hypervideo Player

Download this file

00:01 Introduction

John Bracaglia: My name is John Bracaglia. And I’m a Googler working in YouTube operations. I also lead a group called the Singularity Network, an internal organization focused on discussions and rationality in artificial intelligence. I’m pleased to be here today with Mr. John Searle.

As a brief introduction, John Searle is the Slusser Professor of Philosophy at the University of California-Berkeley. He is widely noted for his contributions to the philosophy of language, philosophy of mind, and social philosophy. John has received the Jean Nicod Prize, the National Humanities Medal in the Mind and Brain prize for his work. Among his noble concepts is the Chinese room argument against strong artificial intelligence.

John Searle, everyone.

John Searle: Thank you Thank you. Many thanks. It’s great to be back at Google. It is a university outside of a university. And sometimes, I think, this is what a university ought really to look like. Anyway, it’s just terrific to be here.

And I’m going to talk about some— well, I’m going to talk about a whole lot of stuff. But, basically, I want to start with talking about the significance of technological advances. And America, especially, but everybody, really, is inclined to just celebrate the advances. If they got a self-driving car, who the hell cares about whether or not it’s conscious. But I’m going to say there are a lot of things that matter for certain purposes about the understanding of the technology. And that’s really what I’m going to talk about.

Now to begin with, I have to make a couple rather boring distinctions because you won’t really understand contemporary intellectual life if you don’t understand these distinctions.

01:49 Objectivity vs. subjectivity

In our culture, there’s a big deal about objectivity and subjectivity. We strive for an objective science. The problem is that these notions are systematically ambiguous in a way that produces intellectual catastrophes. They’re ambiguous between a sense, which is epistemic, where epistemic means having to do with knowledge— epistemic— and a sense, which is ontological, where ontological means having to do with existence. I hate using a lot of fancy polysyllabic words. And I’ll try to keep them to a minimum. But I need these two, epistemic and ontological.

Now the problem with objectivity and subjectivity is that they’re systematically ambiguous—I’ll just abbreviate subjectivity— between an epistemic sense and an ontological sense. Epistemically, the distinction is between types of knowledge claims.

If I say, Rembrandt died in 1606, well— no, he didn’t die then. He was born then. I’d say Rembrandt was born in 1606. That is to say, it’s a matter of objective fact. That’s epistemically objective.

But if I say Rembrandt is the greatest painter that ever lived, well, that’s a matter of opinion. That is epistemically subjective. So we have epistemic objectivity and subjectivity.
Underlying that is a distinction in modes of existence.

Lots of things exist regardless of what anybody thinks. Mountains, molecules, and tectonic plates have a mode of existence that is ontologically objective. But pains and pickles and itches, they only exist insofar as they are experienced by a subject. They are ontologically subjective.

So I want everybody to get that distinction because it’s very important because— well, for a lot of reasons, but one is lots of phenomena that are ontologically subjective admit of an account which is epistemically objective.

I first got interested in this kind of stuff. I thought, well, why don’t these brain guys solve the problem of consciousness. And I went over UCSF to their neurobiology gang and told them, why the hell don’t you guys figure out how the brain causes consciousness? What am I paying you to do?

And their reaction was, look, we’re doing science. Science is objective. And you, yourself, admit that consciousness is subjective. So there can’t be a science of consciousness. Now you’ll all recognize that’s a fallacy of ambiguity. Science is indeed epistemically objective because we strive for claims that can be established as true or false, independent of the attitudes of the makers and interpreters of the claim.

But epistemic objectivity of the theory does not preclude an epistemically objective account of a domain that’s ontologically subjective.

I promised you I wouldn’t use too many big words, but anyway there are a few. The point is this. You can have an epistemically objective science of consciousness, even though consciousness is ontologically subjective. Now that’s going to be important.

And there’s another distinction. Since not everybody can see this, I’m going to erase as I go along. There’s another distinction which is crucial. And that’s between phenomena that are observer-independent. And there I’m thinking of mountains and molecules and tectonic plates, how they exist regardless of what anybody thinks.

But the world is full of stuff that matters to us that is observer-relative. It only exists relative to observers and users. So, for example, the piece of paper in my wallet is money. But the fact that makes it money is not a fact of its chemistry. It’s a fact about the attitudes that we have toward it. So money is observer-relative. Money, property, government, marriage, universities, Google, cocktail parties, and summer vacations are all observer-relative. And that has to be distinguished from observer-independent.

And notice now, all observer-relative phenomenon are created by human consciousness. Hence, they contain an element of ontological subjectivity. But you already know that you can have, in some cases, an epistemically objective science of a domain that is observer-relative. That’s why you can have an objective science of economics even though the phenomena studied by economics is, in general, observer-relative, and hence contains an element of ontological subjectivity.

Economists tend to forget that. They tend to think that economics is kind of like physics, only it’s harder. When I studied economics, I was appalled. We learned that marginal cost equals marginal revenue in the same tone of voice that in physics we learned that force equals mass times acceleration. They’re totally different because the stuff in economics is all observer-relative and contains an element of ontological subjectivity. And when the subjectivity changes— ffft— the whole thing collapses. That was discovered in 2008. This is not a lecture about economics. I want you to keep all that in mind.

Now that’s important because a lot of the phenomena that are studied in cognitive science, particularly phenomena of intelligence, cognition, memory, thought, perception, and all the rest of it have two different senses. They have one sense, which is observer-independent, and another sense, which is observer-relative. And, consequently, we have to be very careful that we don’t confuse those senses because many of the crucial concepts in cognitive science have as their reference phenomena that are observer-relative and not observer-independent. I’m going to get to that.

OK, everybody up with us so far? I want everything to sound so obvious you think, why does this guy bore us with these platitudes? Why doesn’t he say something controversial?

08:50 The emergence of cognitive science

Now I’m going to go and talk about some intellectual history.

Many years ago, before any of you were born, a new discipline was born. It was called cognitive science. And it was founded by a whole bunch of us who got sick of behaviorism in psychology, effectively. That was the reason for it. And the Sloan Foundation used to fly us around to lecture, mostly to each other. But anyway, that’s all right.

We were called Sloan Rangers. And I was invited to lecture to the Artificial Intelligence Lab at Yale. And I thought, well, Christ, I don’t know anything about artificial intelligence. So I went out and bought a book written by the guys at Yale. And I remember thinking, $16.95 plus tax— money wasted. But it turned out I was wrong. They had in there a theory about how computers could understand.

And the idea was that you give the computer a story. And then you ask the computer questions about the story. And the computer would give the correct answer to the questions even though the answer was not contained in the story.

A typical story: A guy goes into a restaurant and orders a hamburger. When they brought him the hamburger, it was burned to a crisp. The guy stormed out of the restaurant and didn’t even pay his bill. Question, did the guy eat the hamburger?

Well, all of you computers know the answer to that. No, the guy didn’t eat the hamburger. And I won’t tell you the story where the answer is yes. It’s equally boring. Now, the point was this proves that the computer really understands the story.

So there I was on my way to New Haven on United Airlines at 30,000 feet. And I thought, well, hell, they could give me these stories in Chinese. And I could follow the computer program for answering stories. And I don’t understand a word of the story And I thought, well, that’s an objection they must have thought of. And besides that won’t keep me going for a whole week in New Haven.

Well, it turned out they hadn’t thought of it. And everybody was convinced I was wrong. But interestingly they all had different reasons for thinking I was wrong. And the argument has gone on longer than a week. It’s gone on for 35 years. I mean, how often do I have to refute these guys?

11:17 The Chinese Room Argument

But anyway, let’s go through it. The way the argument goes in its simplest version is I am locked in a room full of Chinese— well, they’re boxes full of Chinese symbols and a rule book in English for manipulating the symbols. Unknown to me, the boxes are called a database, and the rule book is called a program.

In coming in the room, I get Chinese symbols. Unknown to me, those are questions. I look up what I’m supposed to do. And after I shuffle a lot of symbols, I give back other symbols. And those are answers to the questions.

Now we will suppose— I hope your bored with this, because I am. I mean, I’ve told this story many times. We will suppose that they get so good at writing the program, I get so good at shuffling the symbols, that my answers are indistinguishable from a native Chinese speaker. I pass the Turing test for understanding Chinese. All the same, I don’t understand a word of Chinese. And there’s no way in the Chinese room that I could come to understand Chinese because all I am is a computer system. And the rules I operate are a computer program.

And— and this is the important point— the program is purely syntactical. It is defined entirely as a set of operations over syntactical elements. To put it slightly more technically, the notion some implemented program defines an equivalence class that is specified completely independently of any physics and, in particular, independent of the physics of its realization.

The bottom line is if I don’t understand the questions and the answers on the basis of implementing the program, then neither does any other digital computer on that basis because no computer has anything that I don’t have. Computers are purely syntactical devices. Their operations are defined syntactically.

And human intelligence requires more than syntax. It requires a semantics. It requires an understanding of what’s going on. You can see this if you contrast my behavior in English with my behavior in Chinese.

They ask me questions in English. And I give answers in English. They say, what’s the longest river in the United States? And I say, well, it’s the Mississippi, or the Mississippi-Missouri, depending on if you count that as one river.

They ask me in Chinese, what’s the longest river in China? I don’t know what the question is or what it means. All I got are Chinese symbols. But I look up what I’m supposed to do with that symbol, and I give back an answer, which is the right answer. It says, it’s the Yangtze. That’s the longest river in China. I don’t know any of that. I’m just a computer.

So the bottom line is that the implemented computer program by itself is never going to be sufficient for human understanding because human understanding has more than syntax. It has a semantics.

There are two fundamental principles that underlie the Chinese room argument. And both of them seem to me obviously true. You can state each in four words. Syntax is not semantics. And simulation is not duplication. You can simulate— you’re going to have plenty of time for questions.

How much time we got, by the way? I want to—

John Bracaglia: We’ll leave time for questions at the end.

John Searle: I want everybody that has a question to have a chance to ask the question.

Anyway, that’s the famous Chinese room argument. And it takes about five minutes to explain it. Now you’d be amazed at the responses I got. They were absolutely breathtaking in their preposterousness. Now let me give you some answers.

A favorite answer was this: You were there in a room. You had all those symbols. You had a box. You probably had scratch paper on which to work. Now, it wasn’t you that understood. You’re just a CPU, they would say with contempt, the Central Processing Unit. I didn’t know what any of these words meant in those days. But it’s the system that understands.

And when I first heard this, I mean, the room understands Chinese, I said to the guy. And he said, yes, the room understands Chinese.

Well, it’s a desperate answer. And I admire courage. But it’s got a problem. And that is the reason I don’t understand is I can’t get from the syntax to the semantics. But the room can’t either. How does the room get from the syntax of the computer program of the input symbols to the semantics of the understanding of the symbols? There’s no way the room can get there because that would require some consciousness in the room in addition to my consciousness. And there is no such consciousness. Anyway, that was one of many answers.

One of my favorites was this: This was in a public debate. A guy said to me, but suppose we ask you, do you understand Chinese? And suppose you say, yes, I understand Chinese. Well? Well, OK, let’s try that and see how far we get. I get a question that looks like this. Now, this will be in a dialect of Chinese some of you won’t recognize. Unknown to me, that symbol means, “Do you understand Chinese?” I look up what I’m supposed to do. And I give them back a symbol that’s in the same dialect of Chinese. And it looks like that. And that says, “Why do you guys ask me such dumb questions? Can’t you see that I understand Chinese?”

I could go on with the other responses and objections, but I think they’re all equally feeble. The bottom line is there’s a logical truth. And that is that the implemented computer program is defined syntactically. And that’s not a weakness. That’s the power. The power of the syntactical definition of computation is you can implement it on electronic machines that can perform literally millions of computations in a very small amount of time. I’m not sure I believe this, but it always says it in the textbooks, that Deep Blue can do 250 million computations in a second. OK, I take their word for it. So it’s not a weakness of computers.

18:09 Semantics and the definition of understanding

Now, another argument I sometimes got was, “Well, in programs, we often have a section called the semantics of natural understanding programs.” And that’s right. But, of course, what they do is they put in more computer implementation. They put in more syntax.

Now, so far, so good. And I think if that’s all there was to say, I’ve said all of that before. But now I want to go on to something much more interesting. And here goes with that.

Now how we doing? I’m not— everybody seems to understand… There’s going to be plenty of time for questions. I insist on a good question period.

So let me take a drink of water, and we go to the next step, which I think is more important.

A lot of people thought, well, look, maybe the computer doesn’t understand Chinese, but all the same, it does information processing. And it does, after all, do computation. That’s what we define the machine to do. And I had to review a couple of books recently. One book said that we live in a new age, the age of information. And in a wonderful outburst, the author said everything is information. Now that ought to worry us if everything is information.

And I read another book. This was an optimistic book. I reviewed— this for “The New York Review of Books”— a less optimistic book by a guy who said computers are now so smart they’re almost as smart as we are. And pretty soon, they’ll be just as smart as we are. And then I don’t have to tell this audience the next step. They’ll be much smarter than we are. And then look out because they might get sick of being oppressed by us. And they might simply rise up and overthrow us all. And this, the author said modestly— I guess this is how you sell books— he said this may be the greatest challenge that humanity has ever faced, the upcoming revolt of super-smart computers.

Now, I want to say both of these claims are silly. I mean, I’m speaking shorthand here. There’ll be plenty of chance to answer me. And I want to say briefly why.

The notion of intelligence has two different senses. It has an observer-independent sense where it identifies something that is psychologically real. So I am more intelligent than my dog Tarski. Now, Tarski’s pretty smart, I agree. But overall, I’m smarter than Tarski. I’ve had four dogs, by they way— Frege, Russell, Ludwig, and Tarski. And Tarski, he’s a Bernese mountain dog. I’m sorry I didn’t bring him along, but he’s too big for the car. Now, he’s very smart. But he does have intelligence in the same sense that I do. Only he happens to have somewhat less than I do.

Now, my computer is also intelligent. And it also processes information. But— and this is the key point— it’s observer-relative. The only sense in which the computer has intelligence is not in an intrinsic, but it’s in an observer-relative sense. We can interpret its operations in such a way that we can make— now, watch this terminology— we can make epistemically objective claims of intelligence even though the intelligence in question is entirely in the eye of the beholder.

This was brought home forcefully to me when I read in the newspapers that IBM had designed a computer program, which could beat the world’s leading chess player. And in the same sense in which Kasparov beat Karpov so we were told Deep Blue beat Kasparov. Now that ought to worry us because for Karpov and Kasparov to play chess, they both have to be conscious that they’re playing chess. They both have to know such things as “I opened with pawn to king four, and my queen is threatened on the left-hand side of the board.” But now notice, Deep Blue knows none of that because it doesn’t know anything.

You can make epistemically objective claims about Deep Blue. It made such and such a move. But the attributions of intelligent chess playing, this move or that move, it’s all observer-relative. None of it is intrinsic in the intrinsic sense in which I have more intelligence than my dog, my computer has zero intelligence—absolutely none at all. It’s a very complex electronic circuit that we have designed to behave as if it were thinking, as if it were intelligent.

But in the strict sense, in the observer-independent sense in which you and I have intelligence, there is zero intelligence in the computer. It’s all observer-relative. And what goes for intelligence goes for all of the key notions in cognitive science. The notions of intelligence, memory, perception, decision-making, rationality— all those have two different senses, a sense where they identify psychologically real phenomena of the sort that goes on in you and me and the sort where they identify observer-relative phenomena.

But in the intrinsic sense in which you and I have intelligence, the machinery we’re talking about has zero intelligence. It’s no question of its having more or less. It’s not in the same line of business. All of the intelligence is in the eye of the beholder. It’s all observer-relative.

24:24 Is it even relevant if a computer has consciousness?

Now, you might say— and I would say— so, for most purposes,
it makes no difference at all. I mean, if you can design a car that can drive itself, who cares if it’s conscious or not? Who cares if it literally has any intelligence?

And I agree. For most purposes, it doesn’t matter. For practical purposes, it doesn’t matter whether or not you have the observer-independent or the observer-relative sense. The only point where it matters, if you think there’s some psychological significance to the attribution of intelligence to machinery which has no intrinsic intelligence.

Now, notice the intelligence by which we— the mental processes by which we attribute intelligence to the computer require consciousness. So the attribution of observer-relativity is done by conscious agents. But the consciousness is not itself observer-relative. The consciousness that creates the observer-relative phenomena is not itself observer-relative.

But now let’s get to the crunch line then. If information is systematically ambiguous between an intrinsic sense, in which you and I have information, and an observer-relative sense, in which the computer has information, what about computation? After all, computation, that must surely be intrinsic to the computer. That’s what we designed and built the damn things to do, was computation. But, of course, the same distinction applies. And I want to take a drink of water and think about history for a moment.

When I first read Alan Turing’s article, it was called “Computing Machinery and Intelligence.” Now why didn’t he call it “Computers and Intelligence”? Well, you all know the answer. In those days, “computer” meant “person who computes.” A computer is like a runner or a piano player. It’s some human who does the operation. Nowadays nobody would think that because the word has changed its meaning. Or, rather, it’s acquired the systematic ambiguity between the observer-relative sense and the observer-independent sense.

Now we think that a computer names a type of machinery, not a human being who actually carries out computation. But the same distinction that we’ve been applying, the same distinction that we discovered in all these other cases, that applies to computation in the literal or observer-independent sense in which I will now do a simple computation. I will do a computation using the addition function.

And here’s how it goes. It’s not a very big deal. One plus one equals two.

Now, the sense in which I carried out a computation is absolutely intrinsic and observer-independent. I don’t care what anybody says about me. If the experts say, well, you weren’t really computing. No, I was. I consciously did a computation.

When my pocket calculator does the same operation, the operation is entirely observer-relative. Intrinsically all that goes on is a set of electronic state transitions that we have designed so that we can interpret computationally. And, again, to repeat, for most purposes, it doesn’t matter. When it matters is when people say, well, we’ve created this race of mechanical intelligences. And they might rise up and overthrow us. Or they attribute some other equally implausible psychological interpretation to the machinery.

In commercial computers, the computation is observer-relative. Now notice, you all know that doesn’t mean it’s epistemically subjective. And I pay a lot of money so that Apple will make a piece of machinery that will implement programs that my earlier computers were not intelligent enough to implement. Notice the observer-relative attribution of intelligence here. So it’s absolutely harmless unless you think there’s some psychological significance.

Now what is lacking, of course, in the machinery, which we have in human beings which makes the difference between the observer relativity of the computation in the commercial computer and the intrinsic or observer independent computation that I have just performed on the blackboard, what’s lacking is consciousness.

All observer-relative phenomena are created by human and animal consciousness. But the human and animal consciousness that creates them is not itself observer-relative. So there’s an intrinsic mental phenomena, the consciousness of the agent, which creates the observer-relative phenomena, or interprets the mechanical system in an observer relative fashion. But the consciousness that creates observer relativity is not itself observer-relative. It’s intrinsic.

30:00 The problem of using an obsolete vocabulary

Now, I wanted to save plenty of time for discussion. So let me catch my breath and then give a kind of summary of the main thrust of what I’ve been arguing. And one of the things I haven’t emphasized but I want to emphasize now, and that is most of the apparatus, the conceptual apparatus, we have for discussing these issues is totally obsolete. The difference between the mental and the physical, the difference between the social and the individual, and the distinction between those features which can be identified in an observer-relative fashion, such as computation, and those which can be identified in an observer-independent fashion, such as computation.

We’re confused by the vocabulary which doesn’t make the matters sufficiently clear. And I’m going to end this discussion by going through some of the elements of the vocabulary.

Now, let me have a drink of water and catch my breath. Let’s start with that old question, could a machine think?

Well, I said the vocabulary was obsolete. And the vocabulary of humans and machines is already obsolete because if by machine is meant a physical system capable of performing certain functions, then we’re all machines. I’m a machine. You’re a machine. And my guess is only machines could think. Why?

Well that’s the next step. Thinking is a biological process created in the brain by certain quite complex, but insufficiently understood neurobiological processes. So in order to think, you’ve got to have a brain, or you’ve got to have something with equivalent causal powers to the brain. We might figure out a way to do it in some other medium. We don’t know enough about how the brain does it. So we don’t know how to create it artificially. So could a machine think? Human beings are machines.

Yes, but could you make an artificial machine that could think? Why not? It’s like an artificial heart. The question, can you build an artificial brain that can think, is like the question, can you build an artificial heart that pumps blood. We know how the heart does it, so we know how to do it artificially. We don’t know how the brain does it, so we have no idea.

Let me repeat this. We have no idea how to create a thinking machine because we don’t know how the brain does it. All we can do is a simulation using some sort of formal system. But that’s not the real thing. You don’t create thinking that way, whereas the artificial heart really does pump blood.

So we had two questions. Could a machine think? And could an artificially-made machine think? Answer to question one is obviously yes. Answer to question two is, we don’t know yet, but there’s no obstacle in principle. Does everybody see that?

Building an artificial brain is like building an artificial heart. The only thing is no one has begun to try it. They haven’t begun to try it because they have no idea how the actual brain does it. So they don’t know how to imitate actual brains.

Well, OK, but could you build an artificial brain that could think out of some completely different materials, out of something that had nothing to do with nucleo-proteins, had nothing to do with neurons and neurotransmitters and all the rest of it? And the answer is, again, we don’t know. That seems to me an open question.

If we knew how the brain did it, we might be able to define— I mean, be able to design machines that could do it using some completely different biochemistry in a way that the artificial heart doesn’t use muscle tissue to pump blood. You don’t need muscle tissue to pump blood. And maybe you don’t need brain tissue to create consciousness. We just are ignorant. But notice there’s no obstacle in principle.

The problem is no one has begun to think about how you would build a thinking machine, how you’d build a thinking machine out of some material other than neurons because they haven’t begun to think about how we might duplicate and not merely simulate what the brain actually does. So the question is, could a machine think, could an artificial machine think, could an artificial machine made out of some completely different materials, could those machines think?

And now the next question is the obvious one. Well, how about a computer? Could a computer think?

Now, you have to be careful here. Because if a computer is defined as anything that can carry out computations, well, I just did. This is a computation. So I’m a computer. And so are all of you.

Any conscious agent capable of carrying out that simple computation is capable, is both a computer and capable of thinking. So my guess is— and I didn’t have a chance to develop this idea— is that not only can computers think— you and me— but my guess is that anything capable of thinking would have to be capable of carrying out simple computations.

But now what is the status of computation? Well, the key element here is the one I’ve already mentioned. Computation has two senses, an observer-independent sense and an observer-relative sense. In the observer-relative sense, anything is a computer if you can ascribe a computational interpretation.

Watch. I’ll show you a very simple computer. That computer just computed a well-known function. s equals one-half gt squared.

And if you had a good-enough watch, you could actually time and figure out how far the damn thing fell. Everybody sees. It’s elementary mathematics.

So if this is a computer, then anything is a computer because being a computer in the observer-relative sense is not an intrinsic feature of an object, but a feature of our interpretation of the physics of the phenomenon.

In the old Chinese room days, when I had to debate these guys, at one point, I’d take my pen, slam it on a table, and say that is a digital computer. It just happens to have a boring computer program. The program says: stay there.

The point is nobody ever called me on this because it’s obviously right. It satisfies a textbook definition.

You know, in the early days, they tried to snow me with a whole lot of technical razzmatazz. “You’ve left out the distinction between the virtual machine and the non-virtual machine” or “You’ve left out the transducers.” You see, I didn’t know what the hell a transducer was, a virtual machine. But it takes about five minutes to learn those things.

Anyway, so now we get to the crucial question in this. If computers can think, man-made computers can think, machines can think, what about computation? Does computation make a machine a thinking process? That is, is computation, as defined by Alan Turing, itself sufficient for thinking? And you now know the answer to that.

In the observer-relative sense, the answer is no. Computation is not a fact of nature. It’s a fact of our interpretation. And insofar as we can create artificial machines that carry out computations, the computation by itself is never going to be sufficient for thinking or any other cognitive process because the computation is defined purely formally or syntactically. Turing machines are not to be found in nature. They’re to be found in our interpretations of nature.

Now, let me add, a lot of people think, ah, this debate has something to do with technology, or there’ll be advances in technology. I think that technology is wonderful. And I welcome it. And I see no limits to the possibilities of technology. My aim is this talk is simply to get across, you shouldn’t misunderstand the philosophical, psychological, and, indeed, scientific implication of the technology.

38:41 Questions from the audience

John Bracaglia: Thank you, John.

John Searle: I’m sorry I talk so fast, but I want to leave plenty of time for questions.

John Bracaglia: We will start with one question from Mr. Ray Kurzweil.

Ray Kurzweil: Is this on?

Well, thanks, John. I’m one of those guys you’ve been debating this issue for 18 years, I think. And I would praise the Chinese room for its longevity because it does really get at the apparent absurdity that some deterministic process like computation could possibly be responsible for something like thinking.

And you point out the distinction of thinking between its effects and the subjective states, which is a synonym for consciousness. So I quoted you here in my book “Singularity is Near,” about the equivalence of neurons and even brains with machines. So then I took your argument why a machine and a computer could not truly understand what it’s doing and simply substituted human brain for computers, since you said they were equivalent, and neurotransmitter concentrations and related mechanisms for formal symbols, since basically neurotransmitter concentrations, it’s just a mechanistic concept.

And so you wrote, with those substitutions, “The human brain succeeds by manipulating neurotransmitter concentrations and other related mechanisms. The neurotransmitter concentrations and related mechanisms themselves are quite meaningless. They only have the meaning we have attached to them. The human brain knows nothing of this. It just shuffles the neurotransmitter concentrations and related mechanisms. Therefore, the human brain cannot have true understanding.”

So — [Laughter]

John Searle: There’s something interesting variations, again, on my original.

40:50 How to recognize consciousness

Ray Kurzweil: But the point I’d like to make, and that I’d be interested in your addressing, is the nature of consciousness because, I mean, you said today, and you wrote, the essential thing is to recognize that consciousness is a biological processes like digestion, lactation, photosynthesis, or mitosis.

We know that brains cause consciousness with specific biological mechanisms. But how do we know that a brain is conscious? How do you know that I’m conscious? And how do you—

And how do we know if a computer was conscious? We don’t have a computer today that seems conscious, that’s convincing in its responses. But my prediction is we will. We can argue about the time frame. And when we do, how do we know if it’s conscious of it just seems conscious? How do we measure that?

John Searle: Well, there are two questions here. One is, if you do a substitution of words that I didn’t use and the words I did use, can you get these observed results? And, of course, you can do that. That’s a well-known technique of politicians. But that wasn’t the claim.

What is the difference between the computer and the brain? In one sentence, the brain is a causal mechanism that produces consciousness by a certain rather complex and still imperfectly understood neurobiological processes. But those are quite specific to a certain electrochemistry. We just don’t know the details. But we do know, if you mess around in the synaptic cleft, you’re going to get weird effects.

How does cocaine work? Well, it isn’t because it’s got a peculiar computational capacity. Because it messes with the capacity of the postsynaptic receptors to reabsorb quite specific neurotransmitters, norepinephrine— what are the other two? God, I’m flunking the exam here. Dopamine. Gaba is the third.

Anyway, the brain, like the stomach or any other organ, is a specific causal mechanism. And it functions on specific biochemical principles.

The problem of the computer is that it has nothing to do with the specifics of the implementation. Any implementation will do, provided it’s sufficient to carry out the steps in the program. Programs are purely formal or syntactical. The brain is not.

The brain is a specific biological organ that operates on specific principles. And to create a conscious machine, we’ve got to know how to duplicate the causal powers of those principles. Now, the computer doesn’t in that way work as a causal mechanism producing higher level features.

Rather, computation names an abstract mathematical process that we have found ways to implement in specific hardware. But the hardware is not essential to the computation. Any system that can carry out the computation will be equivalent.

Now, the second question is about how do you know about consciousness. Well, think about real life. How do I know my dog Tarski is conscious, and this thing here, my smartphone, is not conscious? I don’t have any doubts about either one.

I can tell that Tarski is conscious not on behavioristic grounds. People say, well, it’s because he behaves like a human being. He doesn’t. See, human beings I know, when they see me don’t rush up and lick my hands and wag their tails. They just don’t. My friends don’t do that. But Tarski does.

I can see that Tarski is conscious because he’s got a machinery that’s relatively similar to my own. Those are his eyes. These are his ears. This is his skin. He has mechanisms that mediate the input stimuli to the output behavior that are relatively similar to human mechanisms. This is why I’m completely confident that Tarski is conscious.

I don’t know anything about fleas and termites. You know, your typical termite’s got 100,000 neurons. Is that enough? Well, I lose 100,000 on a big weekend. So I don’t know if that’s enough for consciousness. But that’s a factual question. I’ll leave that to the experts.

But as far as human beings are concerned there isn’t any question that everybody in this room is conscious. I mean, maybe that guy over there is falling asleep, but there’s no question about what the general— it’s not even a theory that I hold. It’s a background presupposition. The way I assume that the floor is solid, I simply take it for granted that everybody’s conscious. If forced to justify it, I could.

Now, there’s always a problem about the details of other minds. Of course, I know you’re conscious. But are you suffering the angst of post-industrial man under late capitalism? Well, I have a lot of friends who claim they do. And they think I’m philistine because I don’t. But that’s tougher. We’d have to have a conversation about that. But for consciousness, it’s not a real problem in a real-life case.

46:03 The difference between running a simulation of a brain and a thinking maching

Audience: So you’ve said that we haven’t begun to understand how brains work or build comparable machines. But imagine in the future we do.

So we can run a simulation, as you put it, of a brain. And then we interface it with reality through motor output, sensory input. What’s the difference between that and a brain, which you say you know is producing consciousness?

John Searle: In some cases, there is no difference at all. And the difference doesn’t matter. If you have got a machine— I hope you guys are, in fact, building it because the newspapers say you are. If you have got a program that’ll drive my car without a conscious driver, that’s great. I think that’s wonderful. The question is not, what can the technology do?

My daddy was an electrical engineer for AT&T. And his biggest disappointment was I decided to be a philosopher, for God’s sake, instead of going to Bell Labs and MIT as he had hoped. So I have no problem with the success of the technology.

The question is, what does it mean? Of course, if you’ve got a machine that can drive a car as well as I, or probably better than I can, then so much the better for the machinery. The question is, what is the philosophical psychological scientific significance of that?

And if you think, well, that means you’ve created consciousness, you have not. You have to have more to create consciousness. And for a whole lot of things, consciousness matters desperately.

In this case of this book that I reviewed, where the guy said, well, they got machines that are going to rise up and overthrow us all, it’s not a serious possibility because the machines have no consciousness. They have no conscious psychological state. It’s about like saying the shoes might get up out of the closet and walk all over us. After all, we have been walking on them for centuries, why don’t they strike back? It is not a real-life worry.


47:59 Causal similarities of brains and current or future artificial systems

Audience: The difference that I’m interested in— sorry, the similarity I’m interested in is not necessarily the output or the outcome of the system, but rather, that is, it has the internal causal similarity to the brain that you mentioned.

John Searle: Yeah, that’s a factual question. The question is, to what extent are the processes that go on in the computer isomorphic to processes that go on in the brain? As far as we know, not very much.

I mean, the chess-playing programs were a good example of this. In the early days of AI, they tried to interview great chess players and find out what their thought processes were and get them to try to duplicate that on computers. Well, we now know how Deep Blue worked. Deep Blue can calculate 250 million chess positions in one second. See, chess is a trivial game from a games theoretical point of view because you have perfect information. And you have a finite number of possibilities. So there are x number of possibilities of responding to a move and x number of possibilities for that move. It’s interesting to us because of the exponential problem. And it’s very hard to program computers that can go very many steps in the exponents, but IBM did.

It’s of no psychological interest. And to their credit, the people in AI did not claim it as a great victory for— at least the ones I know didn’t claim it as a victory for AI because they could see it had nothing to do with human cognition.

So my guess is it’s an interesting philosophical question— or psychological question— to what extent the actual processes in the brain mirror a computational simulation. And, of course, to some respect, they do. That’s why computational simulations are interesting in all sorts of fields and not just in psychology, because you can simulate all sorts of processes that are going on.

But that’s not strong AI. Strong AI says the simulation isn’t just a simulation. It’s a duplication. And that we can refute.

50:06 Attempts to refute the Chinese Room argument

Audience: Could you prove to me that you understand English?

John Searle: Yeah, I wouldn’t bother. [Speaking with British accent] When I was in Oxford, many people doubted that I did. I happened to be in a rather snobbish college called Christ Church. And, of course, I don’t speak English. I never pretended to. I speak a dialect of American, which makes many English people shudder at the thought.

Audience: So you’ve said you understand English, but how do I know you’re not just a computer program?

John Searle: Well, it’s the same question as Ray’s. And the answer is all sorts of ways. You know, if it got to a crunch, you might ask me. Now I might give a dishonest answer. Or I might give an honest answer.

But there’s one route that you don’t want to go. And that’s the epistemic route. The epistemic route says, well, you have as much evidence that the computer is conscious as that we have that you are conscious. No, not really.

I mean, I could go into some detail about what it is about people’s physical structure that makes them capable of producing consciousness. You don’t have to have a fancy theory. I don’t need a fancy theory of neurobiology to say those are your eyes. You spoke through your mouth. The question was an expression of a conscious intention to ask a question.

Believe me, if you are a locally produced machine, Google is further along than I thought. But clearly, you’re not.

51:49 The definition of consciousness

John Bracaglia: We’re going to take a question from the Dory.

John Searle: Is he next?

John Bracaglia: We had some people—

Audience: Almost.

John Bracaglia: We had some people submit questions ahead of time.

John Searle: OK.

John Bracaglia: So we’re going to read those as well.

John Searle: OK. All right. Right.

John Bracaglia: So the first question from the Dory is, what is the definition of consciousness you’ve been using for the duration of this talk?

John Searle: OK. Here goes.

John Bracaglia: Please be as specific as possible.

John Searle: It is typically said that consciousness is hard to define. I think it’s rather easy to define. We don’t have a scientific definition because we don’t have a scientific theory. The commonsense definition of any term will identify the target of the investigation. Water is a clear, colorless, tasteless liquid. And it comes in bottles like this. That’s the commonsense definition. You do science and you discover it’s H2O.

Well, with consciousness, we’re in the clear, colorless, liquid, tasteless sense. But here it is. Consciousness consists of all those states of feeling or sentience or awareness that begin in the morning when you awake from a dreamless sleep. And they go on all day long until you fall asleep again or otherwise become, as they would say, unconscious. On this definition, dreams are a form of consciousness. The secret, the essence, of consciousness is that for any conscious state, there’s something it feels like to be in that conscious state.

Now, for that reason, consciousness always has a subjective ontology. Remember, I gave you that subjective-objective bit. It always has a subjective ontology. That’s the working definition of consciousness. And that’s the one that’s actually used by neurobiological investigators trying to figure out how the brain does it. That’s what you’re trying to figure out. How does the brain produce that? How does it exist in the brain? How does it function?

53:43 Could a true AI emerge organically?

Audience: I’d like to propose a stronger boundary on your observation that we do not know how to build a thinking machine today. Even if we knew how to build it, because, I mean, our thinking machine was built by the process of evolution, I’d like to propose— well, what do you think about stating that, actually, we may not have the time? And that it actually may not matter.

The reason we may not have the time is the probabilities that need to happen, like the asteroid falling and wiping the dinosaurs and whatnot, may not happen in the universe that we live in. But if you subscribe to the parallel universes theory, then there is some artificial consciousness somewhere else.

John Searle: Yeah. OK, about we may not have the time, well, I’m in a hurry. But I think we ought to try as hard as we can. It’s true. Maybe some things are beyond our capacity to solve in the life of human beings on Earth. But let’s get busy and try.

There was a period when people said, well, we’ll never really understand life. And while we don’t fully understand it, but we’re pretty far along. I mean, the old debate between the mechanists and the vitalists, that doesn’t make any sense to us anymore. So we made a lot of progress. There was another half to your question.

Audience: It may not matter because all universe—

John Searle: Oh, yeah. Maybe consciousness doesn’t matter. Well, it’s where I live. It matters to me.

Audience: Philosophically speaking.

John Searle: Yeah, but the point is there are a lot of things that may or may not matter which are desperately important to us: democracy, and sex, and literature, and good food, and all that kind of stuff. Maybe it doesn’t matter to somebody, but all those things matter to me in varying degrees.

Audience: Your artificial heart analogy that you mentioned. I think you included the idea that it’s possible, just like with the artificial heart, that we use different materials and different approaches to simulate a heart and, in some ways, go beyond just— come closer to duplication, that we might, in theory, be able to do the same thing with an artificial brain.

I’m wondering if you think it’s possible that going down the path just trying to do a simulation of a brain accidentally creates a consciousness or accidentally creates duplication, even if we don’t intend to do it with exact same means as a brain is made.

John Searle: I would say to believe in that, you have to believe in miracles. You have to— now think about it.

We can do computer simulations of just about anything you can describe precisely. You do a computer simulation of digestion. And you could get a computer model that does a perfect model of digesting pizza. For all I know, maybe somebody in this building has done it. But once you’ve done that, you don’t rush out and buy a pizza and stuff it in the computer because it isn’t going to digest a pizza. What it gives you is a picture or a model or a mathematical diagram. And I have no objection to that.

But if my life depended on figuring out how the brain produces consciousness, I would use the computer the way you use a computer in any branch of biology. It’s very useful for figuring out the implications of your axioms, for figuring out the possible experiments that you could design. But somehow or other that the idea that the computer simulation of cognitive behavior might provide the key to the biochemistry, well, it’s not out of the question, it’s just not plausible.

John Bracaglia: Humans are easily fooled and frequently overestimate the intelligence of machines. Can you propose a better test of general intelligence than the Turing test, one that is less likely to relate false positives?

John Searle:
Well, you all know my answer to that is the first step is to distinguish between genuine intrinsic observer-independent intelligence and observer-relative intelligence. And observer-relative intelligence is always in the eye of the beholder. And anything will have the intelligence that you’re able to attribute to it.

I just attributed a great deal of intelligence to this object because it can compute a function, s equals one-half squared. Now this object has prodigious intelligence because it discriminates one hair from—I won’t demonstrate it, but in any case take my word for it that it does, even in a head that’s sparse with hair.

So because intelligence is observer-relative, you have to tell me the criteria by which we’re going to judge it. And the problem with the Turing test— well, it’s got all sorts of problems, but the basic problem is that both the input and the output are what they are only relative to our interpretation. You have to interpret this as a question. And you have to interpret that as an answer.

One bottom line of my whole discussion today is that the Turing test fails. It doesn’t give you a test of intelligence.

Audience: So you seem to take it as an article of faith that we are conscious, that your dog is conscious, and that that consciousness comes from biological material, the likes of which we can’t really understand. But forgive me for saying this, that makes you sound like an intelligent design theorist who says that because evolution and everything in this creative universe that exists is so complex, that it couldn’t have evolved from inert material.

So somewhere between an amoeba and your dog, there must not be consciousness. And I’m not sure where you would draw that line. And so if consciousness in human beings is emergent, or even in your dog at some point in the evolutionary scale, why couldn’t it emerge from a computation system that’s sufficiently distributed, networked, and has the ability to perform many calculations and maybe is even hooked into biologic systems?

John Searle: Well, about could it emerge, miracles are always possible. How do you know that you don’t have chemical processes that will turn this into a conscious comb? How do I know that? Well, it’s not a serious possibility.

I mean, the mechanisms by which consciousness is created in the brain are quite specific. And remember, this is the key point. Any system that creates consciousness has to duplicate those causal powers. That’s like saying, you don’t have to have feathers in order to have a flying machine, but you have to duplicate and not merely simulate the causal power of the bird to overcome the force of gravity in the Earth’s atmosphere. And that’s what airplanes do. They duplicate causal powers. They use the same principle, Bernoulli’s principle, to overcome the force of gravity.

But the idea that somehow or other you might do it just by doing a simulation of certain formal structures of input-output mechanisms, of input-output functions, well, miracles are always possible. But it doesn’t seem likely. That’s not the way evolution works.

Audience: But machines can improve themselves. And you’re making the case for why an amoeba could never develop into your dog over a sufficiently long period of time and have consciousness.

John Searle: No, I didn’t make that case. No, I didn’t make that case.

[Interposing voices]

John Searle: Amoeba don’t have it.

Audience: You’re refuting that consciousness could emerge from a sufficiently complex computation system.

John Searle: Complexity is always observer-relative. If you talk about complexity, you have to talk about the metric. What is the metric by which you calculate complexity? I think complexity is probably irrelevant. It might turn out that the mechanism is simple. There’s nothing in my account that says a computer could never become conscious. Of course, we’re all conscious computers, as I said.

And the point about the amoeba is not that amoebas can’t evolve into much more complex organisms. Maybe that’s what happened. But the amoeba as it stands— a single-celled organism— that doesn’t have enough machinery to duplicate the causal powers of the brain.

I am not doing a science fiction project to say, well, there can never be an artificially created consciousness by people busy designing computer programs. Of course, I’m not saying that’s logically impossible. I’m just saying it’s not an intelligent project. If your thinking about your life depends on building a machine that creates consciousness, you don’t sit down your console and start programming things in some programming language. It’s the wrong way to go about it.

Audience: If we gave you a disassembly of Google Translate and had you implement the Chinese room experiment, either it would take you thousands of years to run all the assembly instructions on pen and paper, or else you’d end up decompiling it into English and heavily optimizing it in that form. And in the process, you’d come to learn a lot about the relationships between the different variables and subroutines. So who’s to say that an understanding of Chinese wouldn’t emerge from that?

John Searle: Well, OK, I love this kind of question. All right.

Now, let me say, of course, when I did the original thought experiment, anybody will point out to you if you actually were carrying out the steps in a program for answering questions in Chinese, well, we’d be around for several million years. OK, I take their word for it. I’m not a programmer, but I assume it would take an enormous amount of time.

But the point of the argument is not the example. The example is designed to illustrate the point of the argument. The point of the argument can be given in the following derivation.

Programs are formal or syntactical. That’s axiom number one. That’s all there is to the program. To put it slightly more pretentiously, the notion some implemented program defines an equivalence class specified entirely formally or syntactically. But minds have a semantics, and— and this is the whole point of the example— the syntax by itself is not sufficient for the semantics. That’s the point of the example.

The Chinese room is designed to illustrate axiom three, that just having the steps in the program is not by itself sufficient for a semantics. And minds have a semantics. Now, it follows from those that if the computer is defined in terms of its program operations, syntactical operations, then the program operations, the computer operations by themselves are never sufficient for understanding because they lack a semantics. But, of course, I’m not saying, well, you could not build a machine that was both a computer and had semantics. We are such machines.

Audience: You couldn’t verify experimentally what the difference might be between semantics and what would emerge from thousands of years of experience with a given syntactical program.

John Searle: I think you can— I don’t inherit this. He does.

I think you don’t want to go the epistemic route. You don’t want to say, well, look, you can’t tell the difference between the thinking machine and the non-thinking machine. The reason that’s the wrong route to go is we now have overwhelming evidence of what sorts of mechanisms produce what sorts of cognition.

When I first got interested in the brain, I went out and bought all the textbooks. By the way, if you want to learn a subject, that’s the way to do it. Go buy all the freshman textbooks because they’re easy to understand.

One of these textbooks, it said cats have different color vision from ours. Their visual experiences are different from ours. And I thought, Christ, have these guys been cats? Have the other cats mind problem? Do they know what it’s like to be a cat?

And the answer is, of course, they know completely what’s the cat’s color vision is because they can look at the color receptors. And cats do have different color vision from ours because they have different color receptors. I forget the difference. You can look them up in any textbook.

But if in real life we’re completely confident that my dog can hear parts of the auditory spectrum that I can’t hear. He can hear the higher frequencies that I can’t hear. And cats have a different color vision from mine because we can see what the apparatus is.

We got another question? You’re on.

John Bracaglia: This will be our final question.

John Searle: OK. I’m prepared to go all afternoon. I love this kind of crap.

Audience: So at the beginning of your talk, you mentioned an anecdote about neuroscientists not being interested in consciousness. And, of course, by this time, a number of neuroscientists have studied it.

And so they’ll present stimuli that are near the threshold of perceptibility and measure the brain responses when it’s above or below. What do you think about that? Is that on the right track? What would you do differently?

John Searle: No, I think one of the best things that’s happened in my lifetime— it’s getting a rather long lifetime— is that there is now a thriving industry of neuroscientific investigations of consciousness. That’s how we will get the answer.

When I first got interested in this, I told you I went over to UCSF and told those guys get busy. The last thing they wanted to hear was being nagged by some philosopher, I can tell you. But one guy said to me— famous neuroscientist said— in my discipline, it’s OK to be interested in consciousness, but get tenure first. Get tenure first.

Now, there has been a change. I don’t take credit for the change, but I’ve certainly been urging it. You can now get tenure by working on consciousness.

Now, neuroscience has changed, that now there’s a thriving industry in neuroscience of people who are actually trying to figure out how the brain does it. And when they figure that out— and I don’t see any obstacle to figuring that out— it will be an enormous intellectual breakthrough, when we figure out how exactly does the brain create consciousness.

But in particular, that approach they’re using now— I use the example of presenting stimuli that are near the threshold of perceptibility and looking for neural correlates, do you think that’s going to be fruitful? What particular questions would you ask to find out?

John Searle: I happened to be interested in this crap. And if you’re interested in my views, I published an article in the “Annual Review of Neuroscience” with a title “Consciousness.” It’s easy to remember. You can find it on the web. And what I said is, there are two main lines of research going on today.

There are guys who take what I call the building block approach. And they try to find the neuronal correlate of particular experiences. You see a red object. Or you hear the sound of middle C. What’s the correlate in the brain? And the idea they have is if you can figure out how the brain creates the experience of red, you’ve cracked the whole problem. Because it’s like DNA. You don’t have to figure out how every phenotype is caused by DNA. If you get the general principles, that’s enough.

Now, the problem is they’re not making much progress on this what I call the building block approach. It seems to me a much more fruitful approach is likely to be think of consciousness as coming in a unified field. Think of perception not as creating consciousness, but as modifying the conscious field. So when I see the red in this guy’s shirt, it modifies my conscience field. I now have an experience of red I didn’t have before.

Most people— and the history of science supports them— use the building block approach because most of the history of science has proceeded atomistically. You figure out how little things work, and then you go to big things. They’re not making much progress with consciousness.

And I think the reason is you need to figure out how the brain creates the conscious field in the first place because particular experiences, like the perception of red or the sound of middle C, those modify that conscious field. They don’t create a conscious field from nothing. They modify an existing conscious field. Now, it’s much harder to do that because you have to figure out how large chunks of the brain create consciousness. And we don’t know that.

The problem is in an MRI, that conscious brain looks a lot like the unconscious brain. And there must be some differences there. But at this point— and I haven’t been working on it. I’ve been working on other things.

But I want somebody to tell me exactly what’s the difference between the conscious brain and the unconscious brain that accounts for consciousness. We’re not there yet.

However, what I’m doing here is neurobiological speculation. I mean, I’m going to be answered not by a philosophical argument, but by somebody who does the hard research of figuring out exactly what are the mechanisms in the brain the produce consciousness and exactly how do they work.

John Bracaglia: John, it’s been an immense, immense honor to be here with you today. Thank you so much for your time. And thank you for talking to Google.

John Searle: Well, thank you for having me.



First published on December 4, 2015.

Video source: