NTV and Belief/Indirect Speech Reports

NTV ("no truth value," from Bennett) is the thesis that indicative conditionals (if A then B) do not express propositions and thus lack truth values. This thesis is often defended on the basis of certain considerations regarding Gibbard's standoff cases (in which A->B and A->~B are both equally assertable for well-informed agents), evidence that the Ramsey test seems to predict the assertability conditions for indicatives, and evidence from embedding.

It is embedding that is my concern here - I will set aside the other considerations for now. Because NTV holds that utterances of indicatives do not express propositions, NTV-theorists must offer a theory that departs from orthodoxy about what it is to believe and assert an indicative conditional. And this departure from orthodoxy ramifies throughout their semantics, overturning our simple semantics for indirect speech/belief reports. If John does not assert any proposition when he utters,

(1) If Sue comes, Eric will come too

then when Mark tries to report what John said by uttering (2),

(2) John said that if Sue comes, Eric will come too

the complement of the that-clause must not be a proposition, but something more complicated involving conditional probabilities of John's beliefs. Further, when Mark tries to report what John believes from his utterance, the complement of the that-clause in his belief report cannot be a proposition either:

(3) John believes that if Sue comes, Eric will come

but rather something about the conditional probabilities of John's beliefs. Now, orthodoxy tells us that when we compute the semantic value of a sentence like (2) or (3) the complement of the that-clause is a proposition and the value of the whole sentence is True iff John bears the proper relation (saying or believing) to that proposition. NTV predicts that, whatever the complement of the that-clause may be, it is not a proposition, so our orthodox semantics for indirect speech and belief reports must be reworked for indicative conditionals.

The reworking could be done in one of two ways, both of which seem unsatisfying. First, we could hold that the relation expressed by "believes" and "says" sometimes involves relating an individual to a proposition and other times does not. This piecemeal semantics is unsatisfying since there does not seem to be any linguistic evidence that the denotations of these words behave this way (and there is evidence against it, see below). Second, we could hold that "believes" and "says" are ambiguous - sometimes they denote relations between individuals and propositions and other times they denote relations between individuals and credences (or relations between credences). This response is also unsatisfying since, if these words are genuinely ambiguous we should be able to generate non-propositional readings of them when they take something other than an indicative conditional as a relative clause. But I simply don't see how this may be made to work. Is there a reading of (4) below in which it means something like, "John expressed a high credence for the proposition that Sue is at the party"?

(4) John said that Sue is at the party

So, it seems that NTV's departure from orthodoxy brings with it a problem of how to embed indicatives in belief and indirect speech reports. But even if a new semantics could be satisfactorily integrated with the orthodoxy, there is still reason to doubt such a semantics because there is evidence that indicatives do not behave like paradigm cases of sentences that lack propositional content. Consider a paradigm case of an utterance in which no proposition is expressed:

(5) Shut the door!

Imagine John uttered this sentence. Now, it is bad to report John's speech as follows:

(6) *John said that shut the door

It seems that we cannot indirectly report utterances of sentences that do not express propositions (sentences that "lack truth values"). Yet (2) is natural and grammatical. This contrast case highlights the problem above from a different angle - it doesn't seem that indicatives are semantically special in any way to warrant being treated like non-proposition-expressing sentences. They embed in belief/indirect speech reports just as proposition-expressing sentences.

Another piece of evidence comes from the fact that we say things like,

(6) a. That's true
b. You spoke truly

in response to utterances of indicatives. If NTV is correct, then (6a) must be false (or truth-valueless) - no proposition was expressed of which truth may be predicated! Again, NTV predicts that the following two dialogues are just as bad:

John: Shut the door!
Mark: *That's true

John: If Sue comes, Eric will come too
Mark: That's true

But in the second Mark's response is clearly acceptable while in the first his response is marked. The reason the first is marked is because the "that" in Mark's response picks up as a referent John's command, but commands do not have propositional content and so cannot be true (or false). Compare,

John: Shut the door!
Mark: That's what I want too

Here, the "that" in Mark's response refers to the state of affairs that would result from John's command being obeyed. But states of affairs not have truth values and thus it is a category mistake to predicate truth or falsity of them, which is what Mark does in his first response. So it is confirmed by linguistic intuition and orthodox theory that indicatives behave differently than standard non-propositional sentences.

I present all of this here simply to note a semantic problem for the NTV theory of indicatives. This theory is supported by many other considerations, which perhaps outweigh the problems I note here. However, the problems here suggest that we look hard for a solution to the problems originally motivating NTV that assigns indicatives propositional contents (in context).

More Troubles...

Just another quick thought, this one about a different (though related) paper by von Fintel and Gillies ("CIA Leaks" - ms). In this paper they propose a supposedly natural conversation in which a might-claim proves resistant to retraction. Just a brief note about background: relativists like John MacFarlane use retraction data as evidence that what is being denied when one responds to an epistemic modal claim is not the embedded proposition (sometimes called the prejacent) but rather the entire proposition itself.

Here's an example from MacFarlane:

A: Joe might be in Boston
B: No, he can't be in Boston. I just saw him an hour ago in Berkeley.
A: Oh OK, I guess I was wrong.

What A admits to being wrong about is not the claim that Joe is in Boston (the embedded proposition) but rather the claim that Joe might be in Boston. Consider the weirdness of the following conversation:

A: Joe might be in Boston.
B: No, he can't be in Boston. I just saw him an hour ago in Berkeley.
A: ??Ok, but I stand by what I said earlier - he might be in Boston.

von Fintel and Gillies hold that cases like the one above are not universal, and propose a case where they claim it is natural for A to not retract her original statement:

A: The keys might be in the drawer.
B: (Looking in the drawer and finding nothing) They're not. Why did you say that?
A: I didn't say that they were in the drawer, only that they might be there, and they might have been.

Is this really a case in which A doesn't retract her original might-claim? Here's a reason why not. Might have-claims are ambiguous in a way that might-claims are not. Consider the following minimal pairs:

(1) ??John might come to the party, although he won't
(2) John might have come to the party, although he didn't

There is no reading on which (1) comes out acceptable (what explains this clash, we leave for another day). However, there seems to be two readings for (2), one of which is preferred given the second conjunct. This non-epistemic reading of 'might have' renders (2) equivalent to,

(3) Although John didn't come to the party, he could have come

Where a particularly salient reading of 'could have' here is that it was within John's power to come to the party.

Anyway, my point is this. von Fintel and Gillies make an error thinking that their example conversation counts as evidence that one can resist retracting might-claims when presented with evidence to the contrary. What really seems to be going on is that what A does is retreat to a non-epistemic might have claim which is really consistent with what B says, even though the original (epistemic) might-claim is not. At the very least, their example allows this possibility and thus in the absence of some evidence that such a reading is not available in this situation, their case is not clearly an example that might-claims are resistant to retraction.

Might this be wrong?

Here's what I'm thinking about right now. Andy Egan has a paper ("Epistemic Modals, Relativism, and Assertion" - Phil Studies 2007) in which he presents a simple argument for the thesis that sentences with epistemic modals express propositions that are true/false relative to a judge. The argument rests on the intuition that when person X hears person Y utter a sentence of the form

(1) Might φ

and when X knows that ¬φ, it is appropriate for X to say any of the following:

(2) a. What Y said was false
b. That was false
c. No, it cannot be that φ

However, consider someone else (Z) that does not know that ¬φ (or equivalently, for all Z knows, φ). It seems appropriate for Z to respond to Y's utterance as follows:

(3) a. What Y said was true
b. That was true
c. Yes, might φ

Given that X and Z can appropriately make these replies to Y and ceteris paribus any defeaters (whatever those might be), it seems that this lends itself to a quick argument for relativism about epistemic modals (X accurately reported that the proposition Y asserted was false and Z accurately reported that the very same proposition is true - taken at face value, this suggests that the proposition Y asserted is true and false relative to whoever is assessing it).

What I'm thinking about is Kai von Fintel and Anthony Gillies' paper (ms) "'Might' Made Right". In that paper, they propose a case just like above, but argue that X may respond to Y as follows,

(4) That's right. Might φ

If that is right, then this takes some of the wind out of the relativist's sails, since the relativist wants to say that what is relevant to the truth of epistemic modal claims is what is known by the assessor of the sentence (in this case, X). But in this case what X knows rules out φ so it is not true (relative to X) that might(φ). How would the relativist account for the acceptability of X's response (4)?

But is this response really acceptable for X? Though I am currently in California, a friend of mine, M, doesn't know this and is trying to find me. It is compatible with what she knows that I am in Connecticut, so she says,

(5) Justin might be in Connecticut

I am evaluating her utterance, knowing that I am in fact in California. Would it be appropriate for me to say,

(6) That's right, I might be in Connecticut (?)

(Imagine we are playing some weird phone version of Marco Polo if it helps to put us in conversation with each other). (6) seems to be a very weird and misleading response for me to make, though putting stress on the might alleviates some of the badness. While I don't have strong intuitions one way or the other, here is one reason to think that (6) is bad:

(7) ??I might be in Connecticut, but I know that I'm not
(8) ??I might be in Connecticut, but I am not

Any reasonable account of 'might' must account for why these are bad (either contradictory or unassertable due to some pragmatic rule governing assertion). Now, given the set up of the case, we should all agree that the second conjuncts in (7) and (8) are both assertable for me (I know where I currently am). But then given that they are both assertable by themselves and their conjunction with (6) is not, it seems we have good reason to think that (6) is not assertable for me. I can't think of any examples of two propositions that are both assertable yet their conjunction unassertable (except for Gricean reasons involving "but" - if that bothers you, then change the "but" in (7) and (8) to "and"; the result seems even worse!).