(This arose from a discussion with Thomas Hodgson who's doing a PhD on context-sensitivity at Arché, and earlier from some discussions with François Récanati. I thought I might just as well post if for those interested. I'm sure that what I call here the confusion problem is discussed in the literature, but I'm not able to give a reference to a clear statement of the problem.)

Background

Let's start assuming a thinker (mind/brain/computer) has a Language of Thought that is not context-sensitive. I assume that externalist dependence of content on the environement (Twin-Earth style) does not count as context-sensitivity proper. (See introduction of Récanati's Perspectival Thought on that.) So to clarify: our starting point is a LoT such that for every thinker in the community we're looking at, each LoT word has the same semantic value all its instances/occurences.

The case of EGO

One might first want to add an indexical LoT word corresponding to "I" to our langage. Call it EGO. That's a word whose semantic value is given by the rule:

  • A token of EGO refers to the thinker of the token.

Or:

  • "EGO Fs" tokened by a thinker S is true iff S Fs.

This will hopefully allow us to explain the difference between first-person and non first-person thoughts about oneself in cases like Perry's messy shopper and Kaplan's pants on fire.

Now the case of the indexical "I" seems to me special, because it is the only case in which what I'll call the confusion problem doesn't arise. As it happens, there will never be thoughts in the same mind in which EGO takes different values. In my mind, all my EGO thoughts will systematically refer to me. So as far as my mind/brain is concerned, it can treat EGO as a proper name. (Of course it will be a special proper name, because of how it connects with perceptual information and action, e.g. if coldness on my skin generates the thought "EGO is cold"; but still like a proper name semantically.) Of course EGO is not a proper name, because it occurs in other minds with a different semantic value, but my mind/brain will never be confronted with such thoughts, or with a pair of thoughts in which EGO takes different values. So it does not need to "care" about the context-sensitivity of EGO.

(Things are different if you think we use thoughts of the form "Peter thinks EGO is rich" to represent other's people first-person thoughts - i.e. EGO is a logophoric pronoun / Kaplanian monster; but let's leave that aside and assume EGO works like "I".)

The Confusion Problem

By contrast, add to our eternal LoT an indexical word YOU. Let's say (leaving time aside) the rule for YOU is this:

  • "YOU Fs" tokened by a thinker S when X is the person most perceptually salient to S is true iff X Fs.

Suppose I'm now looking at Alice. The following LoT sentence gets produced:

YOU is a girl.

At the next instant I'm watching Bob. The following LoT sentence gets produced:

YOU is a boy.

Now I have to sentences in my belief box:

YOU is a girl. YOU is a boy.

My little propositional-logic machine works well and generates:

YOU is a girl and a boy.

And if my mind/brain gets there, I'm confused: my mind/brain somehow fails to take proper account of the shifting values of indexical LoT words. That's the confusion problem.

You get the same problem with time. Suppose you have a LoT present tense, or to make it simpler, a NOW indexical. At t1 I token:

NOW it rains

at t2:

NOW it doesn't rain

And I get by inference to:

NOW it rains and it doesn't.

Just to make things clear, one can't reply:

But the mind/brain wouldn't get confused like that! Precisely, because they are indexicals, NOW and YOU have changed their semantic value meanwhile. So the mind/brain would have noticed the change, and would not draw the inferences.

For how does a mind/brain can notice or represent a change of semantic value? By background assumption, only by tokening LoT sentences that represent the difference. But if the mind/brain has only one word, YOU, for representing several values, it won't be able to represent/notice the difference.

Eternalisation and Inferential Isolation

At bottom, there are only two ways a mind/brain can avoid contextual confusion. To avoid the confusion, you have to prevent the inference from happening; you can do this either by (a) making sure that you don't have the same word in the two premisses, or (b) making sure that no inference is drawn from the two premises together. (You can also do both.) That gives two (mutually compatible) options:

  1. eternalisation.
  2. inferential isolation.

Eternalisation

Here's an exemple of what I call "eternalisation". Seeing Alice, I token:

YOU is a girl.

Before I end up in a context where YOU has a different value, that is, before switching my perceptual attention elswhere, I generate a new mental word A, and "copy" all YOU-thoughts into A-thoughts:

A is a girl.

Then I wipe out the YOU-thoughts. I'm now looking at Bob, and token:

YOU is a boy.

My belief box now contains:

A is a girl. YOU is a boy.

No risk of confusion. Later on, I introduce a new mental word B, and copy all the current YOU-thoughts in B-thoughts and wipe them out. My belief box will contain:

A is a girl. B is a boy.

Again, no confusion.

On this picture, the mind tracks the different values of a context-sensitive LoT word by having different words for the different values. So there's a backup of eternal words for the context-sensitive one. (In fact, if the system works smoothly, it will become unclear whether we should ascribe it context-sensitive words at all; why not say that the story above could equally be described as my directly tokening the thought "A is a girl" and then "B is a boy"?)

Inferential isolation: Wiping Out

As for inferential isolation, I see two different way it might go. The common idea is that even though the mind/brain tokens context-sensitive sentences, it isolates them from one another to avoid confused inferences.

The first way I call "Wiping Out": the context-sensitive sentences are wiped out with changes of context. Whenever I switch my perceptual attention, or whenever a instant passes, I wipe out of my mind any YOU-thoughts, or NOW-thoughts, respectively. Looking at Alice, I token for instance:

YOU is a girl.

YOU wears red shoes.

and infer:

YOU is a girl and wears red shoes.

Someone is a girl and wears red shoes.

But when I switch my attention to Bob, all those sentences except the last (which does not contain context-sensitive terms) are simply wiped out. So no confusion occurs.

I wouldn't be surprised if working memory and some animals operated with the Wiping Out technique. For instance you may picture working memory as tokening sentences like:

Car Passes-by.

Dog barks.

Which you can take to be context-sensitive sentences without explicit indexicals, given the following semantics:

  • "Car passes-by"/"Dog barks" tokened at time t and place p is true iff a car passes by p at t/a dog barks at p at t.

The working memory does whatever useful processing it does in terms of these smaller sentences. (For instance it may invoke background information sentences or rules like "All dogs like bones" to draw inferences.) But when changing mental topic the whole working/memory blackboard is wiped-out, but for a few sentences that are eternalized before being stored.

Inferential Isolation: Situation Boxes

The second way is by the mind having various "boxes" for inferentially isolating sentences. At a given context, all context-sensitive LoT sentences are put in a Box, but when context changes a new Box is generated and the previous one set aside. A Box is a sub-part of the belief box (or hypothesis box, or desire box, etc.) such that its context-sensitive sentences don't get out, that is, cannot be used out of Box to draw inferences. This is akin to the distinction between Local and Global variables in programming. It can be extended by having different levels of locality, i.e. boxes within boxes. I call those boxes "Situation Boxes", because typically they will contain a bunch of info about a particular situation.

The idea will get clearer with the following rough representation. When looking at Alice, I first generate a specific empty Situation Box. Call it 1 and write it thus:

[ ]1

Then I token a context-sensitive sentence in the box:

[ YOU is a girl ]1

Now switching my attention to Bob, I set box 1 aside and generate another one, in which a sentence is generated:

[ YOU is a boy ]2

I now have in my belief box:

[ YOU is a boy ]2 [ YOU is a girl ]1

Given the way boxes work, I cannot infer the conclusion "YOU is a boy and a girl", so confusion is avoided.

Note that it is different from the "eternalisation" strategy. For I need not have a different box for the value of each indexical, just one box per set of values corresponding to a situation. For instance, I can both isolate the values NOW and YOU with the same boxes, with a belief box like this:

[ YOU is a girl. YOU talks NOW ]1

[ YOU is a boy. YOU doesn't talk NOW ]2

Within the box, you can make inferences. And in particular, you can infer non-indexical sentences. For instance, one starts with:

[ YOU is a girl. YOU talks. ]1

and generates:

[ YOU is a girl and talks ]1

and then:

[ there is a girl who talks ]1

But (ignoring time) the later is not context-sensitive, so it can get out of the box:

there is a girl who talks

And it can be used within Box 2 or in the general box. Again, confusion is avoided.

Note that the Box technique can be combined with partial eternalisation. The idea is to use, say A and B as "local proper names" for two different persons in a context 1, and then reuse them as "local proper names" for two different persons in context 2:

@@[ A is a girl. B is a girl. A and B are friends ]1

@@[ A is a boy. B is a girl. A and B are not friends ]2

Does confusion occur?

Barring some system to track context-change such as eternalisation, wiping-out or boxing, a mind/brain will get confused when using context-sensitive sentences. But this is precisely an interesting prediction that the context-sensitive thought theory makes. There are cases in which one could make a case for the existence of such contextual confusions, i.e. cases in which we would fail to register shifts in contextual value of thoughts. Vagueness, knowledge attributions, alethic modals seem to me to be good candidates.

Comparisons and applications

Wiping Out and Pylyshyn's FINSTs

I guess Pylyshyn's view of FINSTs is on the Wiping-Out/Eternalisation model: at any instant, you have three FINSTs, say Fa, Fb, Fc gathering info about sthg in your visual field. When a FINSTs "jumps" to another target, its file of sentences is wiped out, except from some sentences that are eternalised (Fa replaced with a proper name A for instance) and transfered to long-term memory.

Situation Boxes and Récanati on episodic memory

You could imagine a story where perceptual-demonstrtive thoughts generate Situation Boxes that are set aside and kept as such when a context-change is registered. So what you get as a memory is not something like "A was a dog and B was a cat" where A and B are proper names, but rather something like "There was a situation with a dog x and a cat y."

This is the way in which Récanati seems to think that episodic memory stores episodes. And that has some plausibility. What you remember/store in long-term episodic memory, on his view, is not a timestamped set of sentences, maybe not even a place-stamped one, but just an non-located episode. ("It was raining, it was dark, there was a man under the lampost, ...".) (Thus Récanati remarks that you typically infer the date of the memory from its contents.) The question then is: how do we not mix up memories together? Suppose you also have a memory like: "It was sunny, there was no man around", how do you avoid inferring that "It was sunny and dark"? If they are not timestamped/placestamped, the only explanation is that they are stored in different inferential boxes that prevent such interference.

Récanati's relativism and the confusion problem

In Perspectival Thought (see my review here), Récanati distinguishes two levels of content: lekton and Austinian proposition. Roughly, lekton is a (possibly incomplete) proposition, and the Austinian proposition is a <lekton, situation> pair that is true iff the lekton is true of the situation. This allows to distinguish two levels of context-sensitivity: (a) standard or contextualist: when the lekton includes a context-sensitive word, (b) relativist: when a lekton is meant to be evaluated relative to different situations.

In my simple LoT terms, Récanati would draw the line between "lekton"/contextualist aspects of context-sensitivity and "situation"/relativist ones at the difference between (again ignoring time):

It rains

and

It rains HERE

Where the first "rains" is meant NOT to be a context-sensitive terms that means "rains-at-the-place-of-utterance", and the semantics:

  • It rains tokened by S at place P is true relative to a situation X iff it rains at X.
  • It rains HERE tokened by S at place P is true iff it rains at P.

and the context of the first utterance fixes that the thought It rains is to be evaluated w.r.t. the thinker's place. (Leaving cases where the thinker is thinking of a distant place aside.)

Récanati would draw the line here whether those LoT are within a system of Wiping Out, or Eternalisation, or Situation Boxes. (The later seems akin to what Récanati calls the "anaphoric mode".)

But I'm wondering whether one should not instead draw the line at the difference between those thoughts whose context-sensitivity the thinker deals with vs. those thoughts whose context-sensitivity the thinker does not deal with. E.g. whenever you have an appropriate Boxing system you would say that the content of my thought YOU is a girl in our case is simply that Alice is a girl, and that the content of the LoT sentence it rains is that it rains here. But where there is no appropriate tracking system (for instance, if you imagine Perry Z-landers being moved around), you would say that they have incomplete thoughts like It is raining that are true only relative to some relevant situation.

This would match relativism and predictions of confusion, in the following way:

  • whenever the truth of some thought S is only relative to situations, speakers fail to register relevant changes of situations.

And this would bring the relativist framework more in line with Perry's idea of "thought without representation".