On Anderson on Epistemic Modals and Fallibilism

Tonight [post delayed due to maintenance on CD] in my fallibilism seminar, we discussed Charity Anderson’s forthcoming paper, “Fallibilism and the Flexibility of Epistemic Modals.”  It’s a really great paper, though written from a diametrically opposed view as my own (well, not entirely, since she does say that Concessive Knowledge Attributions are expressions of fallibilism, which now seems to be a majority view —Lewis, Dougherty, Rysiew, Littlejohn, Anderson–over against Stanley 2005).  She offers a model of a shifting modal base for epistemics to relieve tension between fallibilism and the Knowledge-might principle

Know-Might Principle (KM): If S knows p, then ‘it might be that ~p’ is false for S

Here are a few reflections about how a Jeffreyan radical probabilist like me will think about the issues.

1. First, it seems really, really odd to me that someone who *endorses* fallibilism as the thesis that “knowledge is compatible with a chance of error” *also* endorses KM.  For, it seems obvious to me, if, in the relevant usage, there is a chance that not-p, then not p is a possibility.  So that’s something I just don’t get.  I understand motivations for infallibilism.  It’s an attractive view in many ways (when wedded to a slick error theory).  But I don’t understand the impetus of so many to give fallibiism with one hand and take it away with another.

2. Her basic idea is that—assuming the standard account of epistemic possibility (yuck!)–not only does the relevant domain of knowledge change from person to person, the modal base changes from one time to another for an individual.

For example, when someone says

(*) Sure, it could be a painted mule, but I know it’s not

what’s going on is that the modal base is not the speaker’s total evidence, but a proper subset of it, specifically, lacking that it is a zebra.  Shifty modal bases strike me as waaaay too much machinery for the problem, like swatting a fly with a sledgehammer, when we can explain the data by a probabilistic account of epistemic possibility—<e>p iff Pr(p) > 0.  But there’s another oddity here.

Because of the nature of epistemic support relations, if you have all of the support for p, you have all of the support for all p supported.  So when someone makes a CKA—Kp & <e>~p–, taking p out of the picture isn’t enough, for that leaves in all the epistemic support for p, and so is consistent with K-level J that p, and that’s doing all the work (Kp is, as it were, a mere puppet: the E for p is the hand).  So you’d have to take out all the E for p as well (or face the seemingly insurmountable problem of drawing some non-arbitrary line), and that’s just…well, weird.  And it seems a really drastic maneuver bound to cause problems.

While I was going on about what problems this would cause and the arbitrariness involved, Chris Tweedt and Allie Thornton pointed out that I had misread her view.  I was assuming the evidence base shifted between the first conjunct of the CKA and the second.  Chris pointed out the passage where this is clearly not what she is saying:  “for many propositions we know, there is some body of propositions, Kp, such that we can know p on the basis of K-p even though ‘might ~p’ is true relative to K-p.”  I had rolled right over this, having heard the paper given as a talk.  But the first conjunct of the CKA—“Kp”—is evaluated on the same evidence base as the second conjunct—“<e>~p”—and the latter requires subtracting not just p, but all the evidence for p, then clearly the first conjunct won’t be true, and no harmony has been shown between CKA’s and KM.

Well, the guy is stacking the bar stools now and the other guy is mopping the floor, so I guess other points will have to wait.  This is a very thought-provoking paper which anyone interested in epistemic modals should read.


On Anderson on Epistemic Modals and Fallibilism — 5 Comments

  1. Unfortunately, when I posted this, the site was still in some flux and it posted without comments enabled. They came on of their own accord, but only after the discussion had ensued on Facebook. If you are reading this, we’re probably facebook friends already, so I invite you to join the discussion there until Charity replies here. 🙂

  2. “Shifty modal bases strike me as waaaay too much machinery for the problem, like swatting a fly with a sledgehammer, when we can explain the data by a probabilistic account of epistemic possibility”

    I wonder if the probabilistic account won’t need to have a shifty threshold to make sense of ordinary-language cases where “might” is denied even though the probability is non-zero.

  3. Alex, I don’t think so. No one can sensibly deny a might-statement when context is explicit that the chance is salient. And if it is not salient, then the denial can be of warranted assertability.

    I originally toyed with a context-sensitive probabilistic account:

    p iff Pr(p) > k, where k was set by context. But the strict account has the advantage of automatically satisfying the desideratum that Regularity is respected by epistemic possibility.

    Consider the following two dialogues.


    A: Dang it, I’ve looked everywhere for my keys!

    B: Mightn’t they be on the moon?

    A: Don’t be stupid.


    A: I know I have hands.

    B: Mightn’t you be a brain in a vat?

    A: Well, sure, I *might* be, but I know I’m not.

    A’s final statement in II is fine because it’s clear from context that minute and ordinarily negligible probability is made salient (mention-worthy) by the invocation of a skeptical hypothesis.

    Yet A’s denial (we can make it explicit if we wish) in I is also fine because it’s clear that in that context such possibility has not already been invoked by context (we can assume that B really is just being uncooperative). And not that A’s opening false statement in I is also passed over as assertible because context makes it clear that the proposition asserted (as opposed to what is “said”) is true.

    We could naturally extend I in this way:


    B: Well, what is your evidence that there are no aliens that could have taken them to the moon.

    A: Fine, they *might* be on the moon, but we both know they *aren’t*, so get to looking.

  4. I’m not feeling the force of either oddity.

    Suppose Kratzer’s canonical semantics is correct. Then shifty modal bases are not machinery above and beyond the existing machinery. That is, nothing in Kratzer’s semantics requires that the modal base supplied by the conversational background be the agent’s total knowledge. So Anderson’s point is merely that the knowledge in the modal base can be a proper subset of the agent’s knowledge.

    In fact, it seems far more odd to require that the agent’s total knowledge always be supplied by the conversational background. When I am asserting that so-and-so might be the author of Waverly, my knowledge about where my car is parked is irrelevant both conversationally and epistemically. It has no bearing on whether so-and-so might have authored Waverly. Insofar as we want to accurately model epistemic modals, we’ll probably want our semantics to allow intra-subjective flexibility anyway.

    It also doesn’t seem so strange to knock out knowledge of p, but leave the knowledge-level justification for p. Maybe you want evidence for p around to help order the space of accessible worlds that are supplied by the conversational background. So the modal base is one’s knowledge, but the ordering source for the worlds is one’s evidence. Then you’ll still get the felicity of “might p” utterances by knocking out p, but you will also get the felicity of more gradable modal utterances like “might p, but probably not p” where the probability operator tracks the ordering induced on the worlds by one’s evidence.

  5. Hi Trent,

    Thanks for these comments on my paper. Here are a few quick thoughts in response:

    I agree that when there is a chance that not-p, there is a possibility that not-p. The point I try to make is that the truth of assertions of ‘chance that not-p’ and ‘possibility that not-p’ can be evaluated with respect to one’s total knowledge or a subset of one’s knowledge (just as they can be evaluated with respect to my knowledge or with respect to your knowledge).

    I certainly don’t think that KM, as I qualify it, is infallibilist. But no need to repeat the entire argument of the paper here…

    I don’t understand why you think the K-p base relevant for CKAs would involve subtracting all the evidence for p. In fact, I was thinking quite the opposite. The idea is that when you remove p (and other propositions that obviously entail p, such as ‘p and q’), what remains is a basis which grounds knowledge that p and yet ‘might ~p’ is compatible with this base.

Leave a Reply

Your email address will not be published. Required fields are marked *