On revisionist reporting

Friends of singular thought typically assume that in order to have a singular attitude towards an object, one must either stand in a special acquaintance relation to the object, or have a special kind of mental representation for it. Both of these views face a challenge from our practice of attitude reports: we can seemingly attribute attitudes with singular content even if neither condition is satisfied.

In a well-known example from Sosa 1970, the army generals decide that the shortest man should go first. The Sergeant tells Shorty: 'they want you to go first'. Here the generals need not be acquainted with Shorty, and it is doubtful that they must have a "mental file" for him.

The kind of case I discussed in the previous post is even more challenging. We can report the detective as believing that each villager might be the murderer, even though she is not acquainted with any of the villagers and doesn't even know how many there are. It would be very odd if she nonetheless had a dedicated "mental file" for each villager.

In response, friends of singular thought should probably drop the assumption that the truth of an attitude report whose matrix expresses a singular proposition indicates the presence of a corresponding singular thought. That is, they should hold that 'S believes that x is F' can be true even though S does not believe* the proposition expressed by 'x is F', where belief* is an interesting type of psychological state that is under discussion when philosophers talk about singular thought.

The singularist is then in the same boat as the descriptivist: she has to explain what we mean when we say 'S believes that x is F', if not that S believes* the proposition that x is F.

To this end, singularists could, for example, endorse the descriptivist semantics that I have defended in the previous posts, but restrict it to a certain class of attitude reports. Openshaw 2018 proposes something along these lines: he suggests that 'S believes that x is F' can be true even though S only believes* a proposition of the form 'the G is F', where 'the G' is a description that picks out x.

(Openshaw does not explain how these truth-conditions come about, why they only arise for very specific descriptions 'the G', and how the case generalises to cases like the detective's, where there is no relevant definite description.)

A rather different proposal with the same apparent goal is presented in a fascinating new paper by Kyle Blumberg and Harvey Lederman (henceforth B&L). In short, B&L argue that belief reports (and desire reports, etc.) sometimes report not what the subject actually believes* (or desires* etc.), but what she would believe* (etc.) if she had further information.

B&L's main example resembles the detective case: Pete believes that every six-year-old can easily learn to play tennis; Ann is six years old; although Pete knows nothing specifically about her, we can (in a suitable context) report him as believing that Ann can easily learn to play tennis. According to B&L, this is because Pete would believe* that Ann can easily learn to play tennis if he knew that she is six.

The guiding idea is attractive: when we discuss a certain question – say, whether Ann could learn tennis – we are often interested in what a third person's beliefs would imply for this question, even if that person lacks some relevant information that's available to us. It might therefore be useful to have a shortcut for 'S believes something that, together with what we know, implies that –'.

However, as B&L note, this isn't quite how "revisionist" belief reports appear to work.

First, suppose we falsely believe that Ann is five, and we also falsely believe that Pete thinks every five-year-old can easily learn to play tennis. (In fact, Ann is still six, and Pete believes that every six-year-old can easily learn to play tennis.) We say, 'Pete believes that Ann can easily learn to play tennis'. Arguably, what we say is true, even though Pete does not have a belief* which, together with what we know, implies that Ann can easily learn to play tennis.

B&L therefore suggest that what we report are S's beliefs* enriched not by what we know, but by the true answer to a contextually relevant question (here, Ann's age).

Second, B&L argue that the "enrichment" is not simply a matter of conjunction, since S's beliefs* may be incompatible with the answer to the relevant question. As an example, B&L consider a case in which Pete believes* that every six-year-old can easily learn to play tennis, while he also believes* that Jonny is the only 6-year-old. In a suitable context, we may still report him as thinking that Ann can easily learn to play tennis, and not as thinking that Ann can easily lift a car, although both of these follow from Pete's beliefs* conjoined with the information that Ann is six. B&L therefore appeal to an AGM-style operation of belief revision. Roughly, as I said above, 'Pete believes that Ann can ϕ' is true if Pete would believe* that Ann can ϕ if he were to learn that Ann is six.

I think it is plausible that we sometimes use revisionist reporting. But I suspect the phenomenon is not as common as B&L assume. I prefer a descriptivist analysis for their main examples.

Here is a case where the descriptivist analysis isn't available, and which I would treat as revisionist.

We're riding at a steady 10 km/h on a path through the desert, looking for an oasis. Jones said the oasis should be about 40 km from our starting point. After 3 hours, you ask how much longer we have. I reply, 'Jones thinks it's one more hour'. This is a somewhat strange thing to say, but it might be acceptable. If so, I would treat it as a case of revisionist reporting. I suspect all genuine cases are as strange as this one.

Let's compare the B&L interpretation of problematic singular attitude reports with the descriptivist interpretation that I prefer.

Recall that on the "Quine/Kaplan model", 'S believes that x is F' can be true because there is a suitable role R such that (i) S believes* that the R is F, and (ii) in fact, x is the R. The relevant singular content (that x is F) is here entailed by the subject's beliefs* (that the R is F) together with the true answer to a certain question: who is the R?

So the Quine/Kaplan model looks like a special case of B&L-type revisionist reporting. To be sure, the simple model can't handle the Pete example, but the generalised model from my previous post can: roughly speaking, I suggested that Pete thinks of the group of six-year-olds that they can easily learn tennis, and we distribute this belief about the group (in our reporting) over the group members. Here, too, the relevant singular content (that Ann can learn) is entailed by a belief* (that all six-year-olds can learn) together with certain facts (that Ann is a member of the group of six-year-olds).

If my preferred interpretation is a special case of the revisionist interpretation, why do I think revisionist reports are unusual? Because I think this special case arises through a different mechanism, not through the general mechanism of revisionist reporting.

In any case, there are some differences in how the two accounts treat the cases to which they can both be applied.

For example, my descriptivist model does not involve any "revision" of the subject's belief state. As a result, it has trouble with the Jonny example.

Recall: Pete thinks every six-year-old can easily learn tennis, and also that Jonny is the only six-year-old. In a suitable context, we may still report Pete as thinking that (six-year-old) Ann can easily learn tennis. Here, the class of six-year-olds in Pete's belief worlds are all singleton classes whose sole member is a boy called 'Jonny'. To apply the model from my previous post, we would have to assume that the class of actual six-year-olds is tracked as that singleton class across Pete's belief worlds. In the relevant context, we could therefore also say that Pete believes that Ann is called 'Jonny'. But that seems false.

There's something really nice about B&L's model here. Compare the original Jonny case with a case in which Pete thinks Jonny is the only six-year-old, that Jonny is unusually strong, and that he can easily learn tennis. Here Pete thinks every six-year-old can easily learn tennis, but only because he thinks Jonny can, and he's the only six-year-old. In this case, we would not say that Jonny thinks Ann can easily learn tennis. So the report about Ann is sensitive to why Pete thinks every six-year-old can learn tennis – to whether he thinks this happens to be the case for the class of actual six-year-olds, or whether he thinks this reflects some general fact about six-year-olds. In B&L's account, the difference is reflected in how Pete would respond to learning that Ann is six.

I'm not sure how to handle this problem. I don't see a way to get around it by tweaking the proposal from the previous post. I could, of course, also treat it as genuine case of revisionist reporting. However, it seems to me that when we report Pete as thinking that Ann can easily learn tennis (although he thinks Jonny is the only six-year-old, and therefore that Ann is not six), we are really reporting only a part of Pete's belief* state: namely, his belief* that there's something about six-year-olds in general that allows them to easily learn tennis, ignoring his belief* that Jonny is the only six-year-old. I don't think this is the same as reporting his beliefs* revised by the information that Ann is six. But I don't know how to make the relevant notion of belief parts precise.

I'll shelve this problem for now. Here are some advantages of the descriptivist interpretation over the revisionist interpretation.

First, the descriptivist account has a more narrow scope. It does not apply to the desert example above, or to many other cases for which the revisionist account could be applied. This is an advantage, I think, because all the cases to which the descriptivist explanation applies sound a lot better than the other cases. This would be odd if the same interpretation rules are in play either way.

Second, though relatedly, the descriptivist model fills an important gap in the B&L account (for the cases where they both apply) – namely how to determine the "contextually relevant question". According to B&L, when we report Pete as believing that Ann can easily learn tennis, we enrich (or revise) Pete's beliefs* by information about Ann's age. So Ann's age is the "contextually relevant question". But obviously this question need not be "contextually relevant" in any ordinary sense. As B&L acknowledge, their account requires tight constraints on how to select the contextually relevant question, but they don't explain what these constraints are, how they come about, and how thy are motivated by the pragmatic purpose of revisionist reports. The descriptivist interpretation implies that the relevant question is always 'who is R?', where R is the "suitable role". (For cases like Pete's, where the simple Quine/Kaplan model doesn't apply, we only need a partial answer: R is being six-year-old, and a partial answer to 'who is R?' is 'Ann'.)

Third, the descriptivist model does not assume that the relevant role (i.e., the "relevant question", in B&L terms) is settled by transparent features of the conversational context, as B&L assume. Consider a variant of Sosa's Shorty example. The sergeant, knowing that Shorty is the shortest man in the platoon, tells the lieutenant that the generals want Shorty to go first. The lieutenant goes to Shorty and says, 'The generals want you to go first'. Here the "relevant question" has to be 'who is the shortest man?'. But nobody in the conversational context (of the lieutenant's utterance) may have any idea that this is how the generals picked out Shorty.

Fourth, relatedly, the descriptivist model allows the suitable role to vary from subject to subject. 'All the generals want Shorty to go first' can be true even if the different generals pick out Shorty in different ways: the first general wants the shortest man to go first (and Shorty happens to be the shortest man), the second wants the oldest man to go first (and Shorty happens to be the oldest man), and so on. Here we would need to revise the first general's desires by the information that Shorty is the shortest man, the second by the information that Shorty is the oldest man, and so on. How is that supposed to work? How can "the contextually relevant question" vary for each individual over which we quantify?

Fifth, the descriptivist account applies to puzzles about de re epistemic modality. For example, it explains why the detective can say 'every villager might be the murderer', even though she is not acquainted with the individual villagers. The revisionist model is not applicable here, since it essentially concerns third-person attitude reports.

Sixth, revising the subject's beliefs can have unintended effects. Consider another well-known example, due to Andrea Bonomi: Swann wants to kill his wife's lover; unbeknownst to him, that person is the chief of the army. In a suitable context, we can say: 'Swann wants to kill the chief of the army'. But suppose Swann were to learn that Odette's lover is the chief of the army. It's quite possible that he would no longer want to kill him. (Similarly, Oedipus would no longer want to kill his father if he knew he was his father.) The B&L account seems to get these cases wrong.

Relatedly, suppose my oldest daughter is six. I might say, 'Pete believes that my oldest daughter can easily learn tennis'. On the B&L account, we here update Pete's beliefs by the information that my oldest daughter is six. But suppose Pete believes* that I don't have any children, that I am a happy person, and that having children makes everyone unhappy. If we revise his beliefs* with the information that my oldest daughter is six, we therefore get a belief* state according to which I am unhappy. According to B&L, in the context in which 'Pete believes that my oldest daughter can easily learn tennis' is true, 'Pete believes that I am unhappy' is also true. But that sentence is unambiguously false.

Seventh (or eighth?), how exactly should we understand the relevant revision of a belief state, especially if the move is supposed to help friends of singular thought? If Pete has never heard of Ann, how can we revise his belief* state with the proposition that Ann is six? Will the revised state involve a singular belief* about Ann – a person he has never met and for whom he has no identifying information? Isn't this just what we wanted to avoid? Besides, how does the revision work if Pete has two "files" for Ann, in a "double vision" case?

Speaking of double vision cases, B&L want a fairly classical logic of revisionist belief, but I think double vision cases call for something less classical. Perhaps revisionist reporting is not supposed to apply to standard Ortcutt cases, but it's easy to think of quantified cases in which, say, S believes that all Fs are Gs, and that no H is G, and some x happens to be both F and H. Or suppose Pete thinks* all six-year-olds can easily learn tennis, but also that Ann, who he thinks* is 5, can't easily learn tennis. In such a case, I'm inclined to say that Pete believes both that Ann can easily learn to play tennis and that she can't easily learn to play tennis, while he does not believe the conjunction of these contradictory claims.


# on 25 September 2020, 15:49

Hi Wolfgang!

Thanks so much for your nice words about our paper, and for the interesting comments on it. We really enjoyed thinking through your post, and had a couple of replies we thought we might share.

Before we get to those though, we just wanted to pause on (and note with satisfaction!) how much it seems you agree with us in your post. Our paper could be thought of as having two goals: first, to focus attention on the phenomenon of revisionist reporting and offer a way of handling it; second, to argue that revisionist reporting is pretty widespread and includes cases of exportation. From your post, it seems you agree with us that revisionist reporting is interesting and hard. It also seems that you concede that you don't know how to account for these reports, so even though you sketch an alternative to our theory, it seems you're at least open to the idea that for the cases you agree are indeed revisionist reports, our mechanism is a good contender. Basically, then, it seems you feel you achieved our first goal, and your disagreements center just on our second point: on which cases should be deemed revisionist reports, and (relatedly) whether our mechanism is a good way to handle the cases you see as not being genuinely revisionist reports.

You concede that you can't handle our "Johnny" case, basically because it features revision, but there's a different kind of case that we don't yet see how you would handle and that was also very important to us, our "Running" case, from page 3 of the paper. This case was actually our main reason for giving up on a view in the spirit of the one you develop in your post 7 in this series. To recall the case (quoting from the paper): I tell Mary that some of my friends have entered the race. Mary thinks that all of my friends are fit and strong, so she says ‘A friend of yours will win the race’. Mary has no idea which of my friends entered. Now suppose that we’re at the start looking at the racers lined up. Of my friends, only Bill, Ted and Ann have entered the race, and they’re standing to my left, mixed in with some other racers. I say:

(2) Mary thinks that a person to my left will win the race.

The key points about this example are (a) the indefinite "a person" has to be read narrow scope with respect to "thinks" and (b) there is no way of putting individual concepts for the "people to my left" in bijection with individual concepts for my friends. Presumably they differ widely in cardinality, and there's no salient restriction of either set. (b) feels like a technical glitch that could be massaged away, but (a) is a deep issue: the only kinds of inferences the account in post 7 can handle are ones which involve a move from a universally quantified sentence (that would naturally be used to report the person's belief*) to a sentence featuring a singular term (used in what we would call the revisionist report). We don't see how it can handle this kind of case, with an indefinite.

Now on to more detailed responses to your seven points:

On the first: you claim that the revisionist data sound less good to you than the cases of export. To us, the sentences all sound acceptable, and hence it's natural to think they're all true. It seems you agree, but are proposing a finer-grained difference between the cases. We're not sure what that difference comes to, and in any case we don't hear such a difference; it'd be helpful for us to hear more about what you take the difference to be.

On the second, you claim that your descriptivist account is better than our account because it's more constrained. We don't think that's true. First, you've given some more constraints on what a role is and which ones are salient than we gave in our paper on which questions are salient, but by your own admission those constraints are themselves very context-dependent. If we spot you a function from people to roles, then your account may seem a bit more constrained than ours, but we think getting such a function out of a context is a mechanism that's also pretty unconstrained as it stands. (Compare: if you spot us which question is determined by context, our account is very constrained!) Second, as it stands we don't see how your account will handle examples like our (25), taken from Sudo 2014, which involve cases where the revision seems to concern a property rather than an individual. There are descriptivist treatments of these cases in the literature (e.g. a conference paper by Christopher Baron 2015), but going in that direction seems to make the descriptivist account much less constrained than it might have seemed at first sight.

On the third point, we think there's been a bit of a misunderstanding. We did not say that the contextually salient question is determined by "transparent features of the context", and we agree with the reasons you give here that such an account wouldn't work. We don't think the relevant features are transparent. We actually think we're on pretty similar footing to you here. It's true that you've given more qualitative constraints than we did in our paper, but we don't see that being a dramatic difference between your theory and ours: both of us at this point can at best give some vague descriptions of what will happen in some contexts, but neither of us has anything like a predictive account.

On the fourth point, the issue about plural subjects is an interesting thought, but we're not moved by the idea that it's a deep problem for us. First, it would be helpful to hear if you have a case in mind that requires different questions for different people. In all the cases we could come up with, it always works to use the same question for both people (by taking the meet of the two questions); in the cases where we couldn't use the meet the data degraded to our ears. Second, even supposing you do convince us this kind of complexity is needed, it's not clear that this would be a big problem for our theory, since we could just take the parameter supplied by context to be a function from individuals to questions.

On the fifth point, we think this issue is super interesting, and as we said in the paper (second paragraph of section 7) we recognize descriptivists carve up the data differently than we will be inclined to. But one of us (Harvey) also thinks that currently popular descriptivist proposals have trouble even with phenomena related to Frege's puzzle, and so we think there's independent reason to reject them. But that discussion might be something that it would be better to fill in another day!

On the sixth point, as we say in section 7, we need a different mechanism to handle transparent readings of subject-denoting terms. The Swann example seems to be handled by that; we don't think it's a revisionist report. Your second example here is something we thought about a lot in writing the paper. Basically when a person believe*s p, then it's harder than in ordinary cases of revision to find contexts where it makes sense to revise and report them as believing ~p. We think this probably has to do with the function of revisionist reporting. But anyway we don't see a problem for the theory from this example; it seems to confirm a general pattern.

And finally -- on the seventh point: you seem to have a clearer idea of what counts as singular thought than we do, and we're not sure what to make of this objection. We don't see a problem with thinking that a person could revise their beliefs by a "singular proposition" even if they don't know anything else about the person. But we're not totally sure what's at stake here.

Anyway, thanks again so much for your nice words about the paper, and these thoughtful comments! It was fun to come back to this material.

Harvey and Kyle

# on 08 October 2020, 11:39

Thanks Harvey and Kyle for these extensive and really interesting comments!

I hadn't seen the problem with the running case. You're right that there's no straightforward way my account can handle it. I've played around with some extensions that might do the trick, but I'm not fully convinced by them. So this is another point in favour of your model.

As you correctly note, I don't actually disagree with your model (except in matters of detail), but only about where it should be invoked. This is related to the first point I made in the post: I think revisionist interpretations are unusual, require strong contextual cues, and even then tend to sound strange -- more so than the exportation cases for which I'd give a descriptivist explanation.

Here the running sentence is interesting, because it does sound problematic to me. For what it's worth, I've asked my partner, and she thinks the sentence has no true reading in the described scenario. She easily saw the true reading of simple exportation sentences (like "the detective thinks you are a peasant"). So I'm not convinced the case presents a serious problem for my account, since I only want to capture the mechanism behind the better-sounding ordinary exportation reading.

I'm surprised you find the different cases equally acceptable. Looks like we'd have to poll a few more people. If your sample is more representative than mine, I'd have to reconsider.

Just one other comment on your response to my points 3 and 4. I see that to some extent, you could technically get around these problems. But I'd like to understand why we use a mechanism that works in this way. I found your initial explanation appealing: we're often interested in what somewhat else's beliefs would imply for the topic we're discussing if they knew certain facts that we're taking for granted. I'd like to keep the revisionist mechanism as close as possible to this intuitive function. In order to deal with some exportation cases, you instead have to move far away from it. The "contextually relevant question" will often be a question of which we have no idea what it is; it will depend on details of the subject's (unknown) belief state, on which subjects we quantify over, etc. But why are we interested in what someone's beliefs would imply for the topic we're discussing if they knew the answer to /this/ question (which we don't even know)? Here I think the descriptivist treatment offers a more attractive explanation.

# on 26 October 2020, 16:51

Thanks Wolfgang. That's very interesting about your and your partner's judgments about the running case. Our consultants agreed with us that it was fine. We will have to ask more people!

Add a comment

Please leave these fields blank (spam trap):

No HTML please.
You can edit this comment until 30 minutes after posting.