How else could it be

Productivity, Generality and Causal Efficacy and the Language of Thought Hypothesis

I. Introduction

People go around believing things and evaluating sentences as true or false, planning courses of action and expressing surprise at certain discoveries accordingly. So it’s pretty clear that somehow various interactions with the environment have an impact on our brains and that this in turn exerts an influence on how we act and react. But it’s not clear how this ‘storing’ works. Are there distinct things in our brains which correspond to each of the sentences which we believe? Do these neural correlates of sentences mirror the logical structure of the sentences they correspond to, so that there are neural correlates of concepts as well and e.g. believing ‘Jim is hungry’ involves having things corresponding to HUNGRY( ) and JIM suitably composed in one’s brain? Should the correlates themselves be considered as making claims about the world and having truth-conditions so that the bit of my brain which stores my belief that Jim is hungry will itself express a claim which is true or false in the same way that ink marks on paper do?

In this paper I will consider a model of belief (which I will call ‘the language-of-thought-hypothesis’ without meaning to imply that everyone who believes there is a language of thought would accept the model as I characterize it) on which the answer to all three of these questions is yes. We might be motivated to accept this model by noting how it explains certain notable features of human thought, namely productivity, generality and causal efficacy. However, I will argue that there are other models of belief which explain these same features in an equally natural way but do not posit any structures in the brain corresponding to individual beliefs. Thus in advance of actual empirical research into how our brains do store information, considerations of productivity, generality and causal efficacy give us no reason to think that there is (or is not) a language of thought.

II. Productivity, Generality and Causal Efficacy

I’ll start by briefly describing each of the three relevant features of human thought. Productivity refers to the finite human brain’s capacity to understand (or believe) a countably infinite number of different propositions. Human thought is productive in that we can think infinitely many things. For example, if we disregard concerns about the length of the human life and attention span then a person can use very limited cognitive resources to understand and have beliefs about any member of an infinite series of sentences like ‘Its raining.’, ‘Its not the case that its raining.’, ‘Its not the case that its not the case that its raining’… or of phrases like ‘missile’, ‘anti-missile’, anti-anti-missile-missile’…

Generality expresses another feature of the human mind’s capacity for understanding: that when a person acquires a concept they are generally able to apply that concept to everything (which they understand) in the category which that concept applies to. So for example, when a person comes to understand how ‘It is not the case’ works in sentences like ‘It’s not the case that it’s raining’, they don’t just understand it in that one case, but they also see what it would mean for this phrase to be tacked on to other sentences they understand.

Finally, by the ‘causal efficacy’ of beliefs, I simply mean the common-sense view that beliefs cause or play a role in causing people to do things: that a belief that water was in the lounge could cause a thirsty person to go the lounge, and the like.

III. Language of Thought

I’ll now try to show how a language-of-thought picture of belief accounts for the three features of thought, before going on to suggest an alternative picture, which I will call ‘mental model(s)’ and showing how it accounts for the same features.

Let’s start off with the language of thought. Suppose I answer ‘yes’ to the three questions phrased in the introduction, so there are things in the brain which correspond to both to sentences and their constituent concepts and believing a sentence is a matter of relating in the right way to the mental entity corresponding to it (let us say, ‘having the corresponding mental-sentence in one’s belief box’). This suggests a kind of molecular model of belief, in which a person’s mind is stoked with (or with the materials to make) a bunch of concept-correlates and object-correlates. As they go around forming beliefs, and finding things out like, say, that there is tea in the kitchen they add all the sentences which they have come to believe (e.g. that there is tea in the room to the west of the dining room, that there is coffee or tea in the kitchen, that there is something to drink in the house) to their belief box as follows. For each of these sentences, they stick the mental entities corresponding to the concepts in the sentence together in the way corresponding to the way they are combined in the sentence, to create a mental entity made out of these concept-correlates whose physical (or functional) structure mirrors the logical structure of the sentence, and then they put this sentence-correlate into their belief box. So belief formation is a matter of stocking up ones belief-box with mental entities corresponding to each of the sentences one is coming to believe.

Then, given this bunch of objects in the belief box corresponding to all the sentences a person believes, the various manifestations of belief will be caused by (the presence of) certain sentence correlates. In the case of evaluating English sentences, the brain could simply determine what the correlate of the sentence would be and then check whether there was anything in the belief box matching that. In the case of other manifestations of belief, like acting accordingly or being surprised, a more complicated story can presumably be told about how the mental entities corresponding to relevant beliefs would cause these manifestations, so that the mental object corresponding to a thirsty person’s belief that there is water in the kitchen would play a part in the chain of physical causes leading to their going into the kitchen.

This language of thought theory does explain the productivity, generality and causal efficacy of belief, at least if we ignore certain quibbles about productivity and efficacy which I’ll consider in the next paragraph. It explains productivity by saying that the human mind has a limited number of brain structure types corresponding to concepts, but it has the ability to combine and recombine tokens of these types indefinitely. Given that my brain can make a certain type of mental object corresponding to the concept NOT, and it can attach this concept correlate to any sentence and I have the structure corresponding to the sentence ‘It is raining’ I will be able to form any member of the sequence of sentences mentioned above by attaching one of these concept correlates to my mental sentence the appropriate number of times. The same molecular theory of beliefs also explains generality. For if I am able to believe ‘John is taller than James’ this means that I can make (and put in my belief box) a complex of mental objects of the types corresponding to JOHN, JAMES and IS TALLER THAN( , ). But if I can produce tokens of the latter type then I could just as well combine them with JAMES and JOHN in the other order, or with any other pair of suitable concepts which I possess. Finally, it explains the causal efficacy of belief (or at least allows for such explanation) in that it identifies believing something with the existence of a certain kind of object in the brain. This object (the neural correlate of the sentence believed) might then do things like producing or inhibiting electrical signals or activating other regions of the brain etc. so one can easily imagine a chain of relatively low-level physical causes leading from this object corresponding to a belief to the actions which we’d intuitively like to say the belief caused.

Now for the quibbles. First, I said that the language of thought theory sketched above explains how our finite minds could be capable of believing infinitely many things. This true in the sense that it explains (as I hope I’ve just shown) how there can be infinite collections of sentences/thoughts such that you could believe any member of them. What it doesn’t explain is how a person can believe infinitely many different things. This is a problem because there seem to be many cases where we do want to attribute each member of an infinite collection of beliefs to a person. So for example I believe that fewer than 50 people have been to the moon. I also believe that fewer than 51 people have been to the moon and so on for every integer after that. It would, of course, be perverse for me to report on my beliefs by saying that I believe that fewer than 1 million people have been to the moon, but I certainly count as believing this. If, for example, we were debating the claim that everyone knows that fewer than 1 million people have been to the moon, I would not be a counter example. But now, if the theory sketched above is right, and believing a sentence is a matter of having a corresponding mental entity in my belief box this becomes extremely problematic: how could the infinitely many physical correlates of these infinitely many sentences all fit into my brain, much less into my belief box?

I say that this is only a quibble, although it seems to me a decisive reason to reject the version of the language of thought theory just sketched, because there’s an obvious rejoinder. Maybe your belief box only contains a subset the sentences which you believe, and you count as believing all the other sentences because of their relationship to this subset. So, I might have in my belief box just the one sentence ‘Fewer than 50 people have been to the moon’ and count as believing ‘Fewer than n people have been to the moon’ for every n greater then 51 merely in virtue of this. This proposal that a person believes a sentence if they have a correlate of some relevant (potentially different) sentence raises interesting questions about relations of analyticity and our ignorance of various logically necessary truths which I will consider elsewhere.

Secondly I said that the language of thought theory allows for a nice mechanistic explanation of the common sense view that our beliefs can (help) cause our actions. It does so, however, in a way the following slightly odd, and certainly not common-sensical consequence. Suppose I walk into a room where I see a bottle of water. I start feeling thirsty so I reach over and open the bottle. Here we clearly see the causal efficacy of belief in action: my belief that there was water in the bottle together with my desire for something to drink caused me to reach over and open the bottle. But now, if the language of thought theory is right and beliefs are stored in objects with sentence form, which help cause action, we can ask which belief(s) were causally involved in producing this action. Did my belief ‘There is water in that bottle’ or my belief ‘There is something to drink in that bottle’ or both of these cause me to act?

The two different sentences have different structure and correspond to different items in the brain so it is a perfectly legitimate questions which detailed brain-scans and the like should be able to reveal to us which of these two mental entities were involved in the causal chain that lead to me opening the bottle. What’s weird about this, is that nothing in the realm of introspection of folk psychology – nothing less then careful reaction-time studies or inspection of the brain – will suffice to answer the question. In the case above I don’t say either sentence to myself, I just feel thirsty and reach over, so my experiences don’t tell me anything about which sentences had causal power. And in this case my dispositions to act don’t really bear on the question either, for even if we know that I was so thirsty that whatever beverage I believed was in the bottle I would be disposed to drink it, one can still imagine (on this picture) different ways in which this disposition could be realized: my brain might be set up so that for each beverage different beverage X, having the belief ‘There is X in that bottle’ sets off a causal chain leading to my drinking it, or it might instead be the belief ‘There is some beverage in that bottle’ which touches off this causal chain.

I’m not saying that there couldn’t be neurological facts of this kind to be discovered. But the fact that the language of thought theory entails that there are such facts, settleable only by looking closely at the brain, does somewhat tarnish it’s claims to validate the common man. I think that the common man would agree that my belief that there was water in the bottle caused my action, and that my belief that there was something to drink in it did, and likewise for any belief which matched up with my dispositions (e.g. he would not say that my belief that there was liquid in the bottle caused my action, because if I knew only that it was liquid or if I knew that it was liquid and that it was dishwashing detergent I would not have opened and drunk from it). The language of thought hypothesis agrees with the common man that at least one of these beliefs caused the action buts says that it’s an as yet unknown matter of neurobiology which one(s) did.

IV. Mental Model(s)

I now want to argue that, even if the quibbles above can be ironed out (I suspect they can), we should not take the productivity, generality and causal efficacy of belief to be evidence for the language of thought theory of belief, by describing an alternative theory of belief which provides an equally natural explanation of these three features without postulating any neural correlates of sentences.

I will introduce this alternative ‘mental models’ picture (and its contrast to the language of thought picture) with a toy example. Imagine a little machine with wheels to move itself around a room, and two kinds of input devices: a system of magnets which tell it when it’s next to certain objects A B C D and E, and a connection with a keyboard where a person can input certain ‘sentences’. Imagine that the machine is programmed so that if you put in a room containing certain of these objects it explores around noting when it runs into one of these objects or a wall and then after this it can respond to queries like ‘A, n, n, n, e, e, B?’ by sending back one signal, call it ‘yes’ if going three units north and then two units east from A would lead it to B, and ‘no’ otherwise e.g. if there’s a wall inbetween or if A isn’t northeast of B.

Now, there are at least two ways we can imagine designing such a robot to explore its surroundings and answer these queries. One way (which I think is analogous to the language of thought picture of belief) would be for the robot to keep a list of all the query sentences it ‘knows’ to be true, and update this list each time it moved into a new square by adding all the new sentences which it now knows to be true, and then to answer queries after it has finished exploring by sending ‘yes’ if the query is on its list and ‘no’ otherwise. So for example it starts out in the place where A is and moves two units north and encounters B it will add ‘A, n, n, B’ and ‘A, n, e, n, w, B’ and ‘A, s, n, n, n, B’ and all the other sentences to its list at that time. Another way we can imagine this machine working would be for it to keep and update not a list of sentences but a map of its surroundings and then apply certain recursive rules to whatever sentence gets inputted to determine whether that sentence fits the map or not. So if it starts at A and goes two steps north and detects that it is at B, at each step it will record only one change on the internal map and end up producing something like:

Then if a sentence like ‘A, n, e, n, w, B?’ it will evaluate the sentence using rules like ‘C’ is true at a point on the map iff object C is at that point, ‘n, x’ is true at a point iff ‘x’ is true at the point one sep up from that point, and ‘A, x’ is true if there is some point on the map such that ‘A’ and ‘x’ are both true at it. Thus rather then adding all the sentences which it ‘knows’ to a list as it discovers them, we can imagine the robot storing the information it gathers in a form which has a very different structure from these sentences and then using recursive rules to evaluate whatever sentences it is queried with against this map.

The general point I mean this example to suggest is that even though the brain clearly manages to store information in such a way that it can later evaluate English sentences as being known to be true, false, unknown etc. relative to that information it may not do this by using any kind of entities with the same structure as these English sentences. The brain might store information using some kind of data structure which does not look like a complete list of English sentences known to be true (e.g. a map or a table) and then use recursive rules to evaluate English sentences relative to this structure as needed. So believing some sentence S wouldn’t be a matter of having some object with the same structure as S somewhere in you head. Rather it would consist in having whatever structure the brain uses to store information relevant to S configured in such a way that when the sentence S is evaluated relative it, S comes out to be true.

This picture can, I claim, explain the productivity, generality and causal efficacy of belief and indeed it can do so without raising the kind of quibbles about productivity and causal efficacy which the language of thought view does.

The productivity and generality of thought both arise (on this view) from the idea that our brains are capable of applying recursive rules to evaluate sentences in natural language. If, for example, your brain has the capacity to evaluate ‘It is not the case that it’s raining’ this is because it has some part which can turn an evaluation of ‘x’ into an evaluation of ‘it is not the case that x’. And if it has this capacity then, as generality requires, it will be able to apply it to evaluate other sentences like ‘It is not the case that John is hungry’. And, as productivity requires it will be able to the same sentence again and again as required, and so to evaluate ‘It is not the case that it is not the case that it is not the case that it is raining’ or any other member (ignoring issues of boredom and death) of the infinite series of sentences we considered. The theory also explains how a belief can figure in a series of mechanistic causes leading up to an action. Even though there is no object in the second robot which I described that corresponds to or shares the structure of the sentence ‘A, n, e, n, w, B’ there is a simple and interesting physical fact about the machine (how its map is configured) in virtue of which it ‘believes’ this sentence and the fact that its map is so configured must figure in the chain of causes which lead up to the robot’s sending back ‘yes’ when queried with this string. On this picture one’s belief in a sentence is a matter of one’s brain’s storage mechanism being configured a certain way and this belief causes appropriate actions in the same sense that the approach of a speeding baseball could cause a window to break or the absence of gasoline from a car could cause it to stop. It’s hard to see how the common-sense view that our beliefs cause our actions would (or could) demand any more.

Finally, this picture doesn’t raise the kinds of puzzles about productivity and efficacy which we considered for the language of thought. There’s no issue of how it can be the case that I actually believe an infinite number of sentences if you don’t think that believing a sentence is a matter of having something in your head. On the mental model’s picture what makes it the case that I believe a sentence S is that my mental storage mechanism is such that were my brain to evaluate S relative to it, S would come out to be true. And there’s no problem about a finite brain having an infinite number of dispositions (even something as simple as the table I’m writing on has, for each of an infinite number of objects, either the disposition to break if that object were put on it or the disposition not to break).

And the second issue of which of the various beliefs which are adequately related to my dispositions are actually involved in causing my actions (whether, in the example mentioned above by belief that there was water in the bottle or my belief that there was something to drink in the bottle caused me to reach for it) doesn’t arise because the way in which beliefs cause actions is already supposed to be dispositional. My belief B causes act A in that (something like) in all the closest possible worlds[1] in which my brain’s relevant storage device is so configured that it evaluates the sentence ‘B’ as true, I perform act A, and I don’t perform act A in the closest possible worlds in which the device is not so configured. Thus, once we’ve specified that whatever beverage I thought was in the bottle my total belief state would have caused me to open in, and that in some of the closest possible worlds where I didn’t believe there was water in the bottle, and those in which I didn’t think there was a beverage in there I would not have drunk, there is no inscrutable question about which of these beliefs actually did the causing. And the sentences ‘I opened it because I believed there was water in it’ and ‘I opened it because I believed there was something to drink in it’ both come out true, just as the commonsense tells us.

V. Conclusion

In this paper I have argued that the productivity generality and causal efficacy of belief don’t give us reason to think that there are entities in the brain corresponding to beliefs by suggesting an alternative way that information could be stored in the brain which explains these features without positing such entities.

[1] I am using possible worlds terminology here only for the sake of concision and clarity. There is (so far as I know) no connection between this particular way of expressing counterfactuals and the views about belief attribution here advocated.