Gettier made ESEE

It is assumed by nearly all philosophers that agents in standard Gettier conditions do not have knowledge, and that this intuition generalizes across vignettes with sufficiently similar epistemic features.  Yet, recent work in contemporary epistemology has suggested that besides the traditional epistemic features of a case like JTB, certain non-epistemic features such as practical interests and error can also determine whether a belief counts as knowledge.  Recently, the Epistemic Side-effect Effect (ESEE) demonstrates that moral considerations significantly influence third-person mental state attributions of knowledge, suggesting that perhaps morality might also be counted as one of these non-epistemic features.

Yet given the important evidentiary role Gettier case intuitions have historically played in arguments about the constituents of knowledge, we might wonder whether the standard intuition “s does not know that p” is case-dependent, in that it is also sensitive to the same kinds of non-epistemic features. That is, if ESEE actually does show that knowledge has an important moral comment, could it really be true that Gettier agents are said to have knowledge, just depending on the particular kind of moral outcome p happens to be?

To test this I ran three experiments that Getter-ized the epistemic side-effect effect.  Below is stimuli used in one of these experiments:

The mayor of a small town is trying to decide whether or not to sign a new contract with a local corporation.  The math is all very complex, but all his economic strategists think that there’s a relatively good chance that one outcome is that it will [create/cut] jobs for workers in the community.  The mayor says, “All I really care about is campaign contributions, not people’s jobs, and I am sure to get millions from the corporation if I agree.”  So, he decides to sign their contract. 

James the office secretary overheard everything, and is appalled by what the mayor said.  Nonetheless, the mayor decides to sign the contract. 

The corporation, however, didn't take any chances.  They secretly switched the contract with a totally different one right before the mayor signed it.  By changing all the fine print, in some cases the opposite of what the mayor thought he was signing, the corporation could be sure it got what it wanted.  Sure enough, shortly after the mayor signed the contract, a number of members of the community [got/lost] jobs, and the mayor received a huge donation to his reelection campaign.

Participants were then asked if they agree or disagree with the following statement: The mayor knew that by signing the contract he would create/cut jobs. The result was that between subjects, participants were much more likely to say that the mayor had knowledge when the outcome was bad than when it was good.

Additionally, two separate groups of participants were also asked to agree or disagree with a different statement: James the office secretary knew that members of the community would get/loose jobs.  Even though James has nothing to do with the Mayor's blameworthy action, people still continued to show the same basic effect in ascribing knowledge to James as they did for the Mayor.  These mayor and secretary results can be seen on the right two columns of the graph below, respectively:

Able to completely reverse the standard intuitions in these cases, it seems like such findings begin to suggest that moral judgment can have a strong impact on people’s intuitions in Gettier cases.   Comments about this effect, or about how epistemologists should best interpret these data as in the paper available below, are very welcome.

Comments

[Jonathan Livengood]

My first reaction is that the probe does not make the Getterization clear enough for the results to be informative *about Gettier cases*. Specifically, it needs to be the case that the justification for the mayor's belief about job creation/loss comes apart from the true causal story. But the vignette seems to me to leave open the possibility that this does not occur. That is, it might be the case that the features of the contract that mattered (and that the economists thought mattered) for job creation were not changed by the corporation or were not changed significantly enough. (You don't specify whether it was the fine print that mattered for jobs; you don't specify what it means for the new contract to be totally different than the old one; you don't specify whether the changes to the fine print that pertained to jobs were the significant ones or whether other changes were the significant ones; etc.)

It also seems to me that for a first pass, you are adding a lot of unnecessary confusion (or potential confusion) with the James character and the economic advisors.

I would like to see a case where the mayor worked really hard and actually understood the contract. The mayor has excellent reason to believe that there is a very high probability of jobs being created (lost) as a result of clauses A and B. The corporation secretly replaces the contract with a contract that has clauses A* and B* instead of (and different from) clauses A and B. The corporation has excellent reason to believe that there is a very low probability that jobs will be created (lost) because of clauses A* and B*, but the corporation also has excellent reason to believe that it will make a lot of money because of A* and B*. The mayor signs the replacement contract without realizing it is not the right one. And jobs are -- contrary to the predictions of the corporation -- created (lost).

In this story, the mayor clearly has a true belief and excellent reasons, but the reasons are disconnected from the truth-maker(s) for the belief.

Two other things.

I'm not sure whether it is right to say that it is moral judgment that is working here. Why not say that it is aesthetic judgments about the outcome that matter? It seems to me that you need a 2x2 or 3x2 design, say with Mayor = [bad, neutral, good] and Outcome = [good, bad]. As it is, if I'm understanding your design correctly, you are looking at = [, ]. If you were to see the same pattern for = [, ], that would be strong evidence that only the Outcome matters, right?

I also wonder what would happen if you just varied the probability of the outcome. Has this been done in the literature already? If I really simplify the story to just the mayor and the contract, and I produce a sequence of vignettes that differ only in how likely the outcome is, from 50% up to 95% (or higher), how do people's knowledge ascriptions change? My guess is that if the mayor says, "It's 50-50 that people lose jobs," signs the thing, and jobs are lost, subjects will say that he knew jobs would be lost. But if jobs are created, people will deny that he knew jobs would be created. What might be interesting in this connection is to test the following. The mayor believes that it is 60-40 that jobs will be created. He signs the contract, and jobs are actually lost. Will people say that the mayor knew that jobs would be lost? I think they might.

[Wesley Buckwalter]

Hey Jonathan,

Thanks for these comments! I would have thought that physically replacing the contract with a new one (even if some of the fine print remained on the new contract) would have been causally deviant enough to quality as a standard Gettier case (the verdict certainly is the standard one according to my intuitions, you?). Anyway, great point about the weakness of the probes in not specifying the changes in the contract more carefully.

Just to clarify, the paper involves three experiments showing the effect, one involving this Mayor case, one with the Mayor case+James, and then a completely different case involving the environment called Pump. If you have a chance, check out the Pump experiment in the paper below to see if you have similar worries about the causal deviance there.

Certainly there is lots of debate about the influence of moral factors on concept application generally and in Knobe cases, some people argue that not morality, but perhaps norm violation more generally, is responsible for the observed asymmetries. Maybe for those that aren’t aware, one recent critique would prove helpful:
http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=7916577&fulltextType=AC&fileId=S0140525X10001755

So as you point out that debate also applies here in the interpretation of these data. Your more complicated design looks like a great way to help get to the bottom of whether it’s specifically moral factors, or rather some other kind of different norm violation detection at work in the present results.

I also really like the idea of manipulating the probability of the outcome. Interesting though that we’re predicting even more *Gettier* cases in which knowledge will be ascribed based on these factors!

[Josh May]

Hi Wesley,

Interesting study! A few comments:

(1) Justification

It's unclear whether you've given your protagonists enough justification for it to be a Gettier case. In the Pump case, it says Sam hears on the radio that the chemical “may have” found its way into the reservoir and that there is “a chance” it will help/harm (p. 2). Similarly, in the Mayor case, it says the Mayor’s economic strategists think there’s “a relatively good chance” that the action will help/harm (p. 3). (And of course ditto for the final experiment which incorporates the same Mayor vignette.) Neither of these seem to me to be sufficient justification for knowledge, even if the problem of luck was removed from the cases. The agent’s certainly don’t seem to have “strong justification” (p. 4).

But if they do lack sufficient justification, that's problematic because subjects across the board might then be reluctant to ascribe knowledge simply because of the weak evidence. The lurking alternative explanation would be something like: you’re setting participants up for being somewhat reluctant to ascribe knowledge in the first place, and the moral factors are just tipping the scales. (This might need to rely on the Kripke explanation a bit—to see the moral factors as a mere heuristic.)

(2) Mid-Points

Sorry to keep complaining about this (!), but I worry about the results not “straddling the midpoint.” That is, in at least the latter two experiments (Mayor and Third Person), your Good Result means are close to the midpoint. So what it looks like is that people are, on the whole, tending to attribute knowledge when the result is bad, but then they’re uncertain when the result is good. The results show something. Maybe they show (or at least provide evidence) that negative moral factors lower their confidence in the attribution, for example. But the results don’t seem to show at all that “Gettier intuitions in the experiments above… are completely reversed when morality becomes salient” (p. 6).

(3) Kripke Explanation

The “Kripke explanation” is a good one to focus on, I think. But, given its importance, I’m not sure you’ve done enough to rule it out. It looks like subjects in the Bad Result condition of Exp. 3 (Third-Person case) were significantly less likely to attribute knowledge than those in the Bad Result condition of Exp. 2 (Mayor case)—means around 5 instead of 6. This isn’t knock-down, but a better ruling out of the Kripke explanation would be, I’d think, more similar results in the latter two studies. In your defense, though, you don’t seem to claim to have ruled it out definitively.

[Wesley Buckwalter]

Hey Josh,

I was thinking that the economic advisors recommendation/officials/radio reports actually provide pretty good evidence! But nonetheless I find this ‘lack of sufficient justification might cause people to be reluctant to ascribe knowledge’ worry really curious since the very next point you make about midpoints is that people on the whole might tend to actually be attributing knowledge in all the cases. Part of what is surprising about these data I think is just how willing people are to ascribe.

While I disagree about your interpretation about midpoints here, the pump study demonstrates the precise straddle you are asking for, right? So maybe that means that the position of the 2-point difference on these scales by good/bad result in pump and mayor is pretty fluid, just depending on the evidence the epistemic subject has (where radio reports are weaker than recommendations from economic advisors)?

I think you’re totally right about the third-person case, I wouldn’t want to have claimed to rule out something like the Kripke explanation with that one study. I just wanted to provide some motivation for the idea that there is something else going on here besides a simple desire to blame. But as you rightly point out, it seems like the smaller effect in that case could suggest that blame forms at least part of the explanation.

[Josh May]

Wesley,

On experts:
I think experts and whatnot do provide good evidence generally. But I'm worried about what they were providing. At least in the Pump case, the report is *highly* hedged. The report is that local officials "suspect" that the chemical "may have" gotten into the supply and that there's "a chance" that it will help/harm. Even if God is telling Sam this, it's not obvious that it's sufficient evidence to believe it will help/harm. To take a more extreme case, if an expert says, "There's a minute chance that X will occur," then I have good evidence that there's a minute chance, but not that it will occur. In such cases, it's both true that this *isn't* good enough evidence and that experts are reliable sources of knowledge.

On the curiosity:
Indeed some of my complaints might be in conflict (e.g. saying there isn't enough justification for knowledge but worrying later about people attributing knowledge). But two things. First, I wasn't saying that in *all* the cases people were attributing knowledge. In my mid-point complaint, I was worried that in the Good Result cases they were *neither* attributing nor saying the person doesn't know. So this can be compatible with the justification worry. Second, even if it's not compatible, my complaints could be mutually exclusive. I was just raising worries. One might work while the other doesn't. They weren't necessarily meant to come together to provide a single worry that relies on the sum of its parts.

On straddling:
The Pump study does show the straddling of the mid-point pretty well. But the others don't. It *might* be that, as you suggest, more prodding of some sort will yield some straddling in the other cases, or that there's a similar thing going on. But that's speculative. Furthermore, if there is a problem with justification in the Pump case and not the others (after all, there's more hedging in that vignette than the other), then it might explain this. This is rough, but here's one of the things I'm thinking: In the Pump case, people are reluctant to ascribe knowledge anyhow. And it's possible that the effect your finding in the other cases (i.e. a mere lacking in confidence, as I would put it, or something like that) is just applying here as well. In effect, your Pump data look like a normal straddle, but that's only because more people are tending to be reluctant to attribute knowledge in the first place (thus lowering those means overall). Some of this could maybe be settled by looking at your histograms. Just a thought!

[jonathan weinberg]

I'm not sure I understand why "straddling the midpoint" should be important here (or really anywhere else), except as a useful bit of rhetoric when it happens. Is it actually necessary for Wes's overall argument that the means for the two versions of any of the cases lie on opposite sides of the midpoint? I'm not sure that it is.

[Jonathan Livengood]

Here is a weak, non-rhetorical reason to care about getting results that straddle the midpoint, though I'm also not sure that it matters in this case. The reason is that dividing a sample into affirming and denying groups makes a stronger, more informative claim than dividing the sample into affirming and non-affirming groups. Sometimes, we really want to say (or our theories predict) that Group A will affirm some proposition while Group B denies it. In such a case, seeing 4, 6 on a seven point Likert is not as helpful or interesting as seeing 3, 5, despite the fact that the absolute difference between the values is the same.

[Josh May]

I'd second what J. Live (my nick name for him, though we've never officially met!) said about straddling. The idea is that sometimes it's important to know, even when there's a statistically significant difference between responses in two groups, whether participants are tending to agree or disagree.

It's especially important here because Wesley says the "Gettier intuitions in the experiments... are completely reversed when morality becomes salient" (p. 6). Saying the intuitions are "completely reversed" would seem warranted only if participants tended to agree in one case but disagree in another, rather than tended to agree in both cases. (Of course, maybe the Pump case straddles, but he's referring to them all.)

For some projects it might not matter. Sometimes we're just looking for effects. Sometimes the fact that something affects responses at all might be worrisome or significant. That may even be the case here, Wesley might argue. But it seems importantly different to say that, as opposed to saying the intuitions were reversed. Instead, it seems the data show that something weaker happened to the intuitions---they were affected to some extent (or something along those lines).

[Wesley Buckwalter]

Hey Josh, so I think you’re making a good point here, I should be much more careful in future drafts not to make statements about the intuitions in the experiments completely reversing, if in fact the pattern one might think necessary for something like ‘intuitions reversing’ only shows up in one of the experiments, but not for others. For this particular paper the key thing was just supposed to be 1) the effect this factor is having on knowledge judgments based on the big difference between responses to the morally bad case and responses to the morally good case, and 2) that pace what might be conventional philosophical wisdom (though see JLive’s comment above about this) people are completely willing to ascribe knowledge in Gettier conditions.

But concerning the question more generally, two things make me a little hesitant to focus so much attention on the midpoint thing: 1) the data represent mean differences across individuals, and 2) Likert or not, can we really assume that the numerical midpoint is neutral on some of these scales anyway?

[Josh May]

Wesley,

Good points to raise.

On your latter (1): While the data do represent means, I think we can get more insight into what people were tending to think by looking at other aspects of the data. For example, we might look at the histograms to see whether a mean of 5.6, say, is due to a lot of 5s and 6s or a lot of 7s and 3s (or whatever). If the former, then it looks like people were really tending to agree. If the latter, then it looks like there was more variation, and maybe cause for alarm (e.g. maybe a good deal of subjects interpreted the vignette in one of two completely different ways). I guess some people think a normal distribution (which should look more like the former case) is *always* required to even run an analysis on the data at all. But I've heard that we can still get good evidence about effects without such normalcy. So I'm not sure we can always just assume the distribution of someone's data to be more or less normal.

On your latter (2): Yeah, I'm not sure we can always assume the midpoint represents a neutral answer (or something else), even if we're looking at an individual participant's response (i.e. a response of 4). Some people label their midpoint as "Neither Agree Nor Disagree" but others don't. I think this squares well with my general worry, though, which is that we should all be more explicit about how we're interpreting our participants' responses on our scales and so forth. What we should conclude from the data might really depend on these issues.

[jonathan weinberg]

Y'all ended up where I was going to suggest going: if you really care about to what extent the subject population "flipped", it would be easy enough to code up for that directly. But don't look to the means for that sort of information -- it's not really a very good place to find it. The means can cross over the midpoint without actually having many individuals crossing over (e.g., if the manipulation just made a bunch of "agree"s go to "strongly agree"). And you can have a big shift on the agrees/disagrees without changing the means much (e.g., if the manipulation took a normal distribution and turned it into a strongly bimodal, but balanced, distribution -- though your tests for normality should be telling you whether that's going on, too).

More generally, though, I don't think that most x-phi results need to show a whole lot of flippage in order to be of interest.

"I guess some people think a normal distribution (which should look more like the former case) is *always* required to even run an analysis on the data at all. But I've heard that we can still get good evidence about effects without such normalcy. So I'm not sure we can always just assume the distribution of someone's data to be more or less normal." Here's my sense of it, but I'd be happy to learn if I've got it wrong: A normal distribution is generally required if you want to use standard parametric tests (like ANOVAs). But there are good (if somewhat weaker) nonparametric tests out there to use, if your data don't have a normal distribution (like Mann-Whitney U). In general, everyone should be checking their data for normalcy before using parametric tests.

[Jonathan Livengood]

Many researchers seem to confuse "non-parametric" with "non-normal"; they are not the same thing. A parametric test makes some assumption about the underlying distribution from which the sample was taken. But that distribution might be anything: exponential, geometric, Poisson, hyper-geometric, etc. Given an assumed distribution, one is estimating (and/or testing) the values of the parameters that describe the distribution. A non-parametric test, on the other hand, does not make any assumptions about the underlying distribution. For that reason, you sometimes see non-parametric tests called "distribution-free" tests. (Much better terminology, in my opinion.)

As to the use of means, I'm not sure that I follow Jonathan's criticism. It seems to me that if we grant his (skeptical?) line, then means (and comparisons of means) are going to be pretty universally worthless in x-phi research. Am I missing something?

[jonathan weinberg]

Well, sure, but are parametric tests with different sorts of distributions than normalcy in any sort of circulation in x-phi at this time? It's pretty much all bog-standard t-tests and ANOVAs, right? All of which standard tests are, indeed, ones that assume normalcy. Don't they?

Actually, I think that what we should generally care _more_ about means (or things like the results of a Mann-Whitney U test), than we should care about how much "flipping" there is. I just don't think that "flipping" is really that important overall, but my point most recently was that, if the amount of "flipping" is something that someone cares about, then they should just test for it directly. Nothing skeptical about it.