Podcast: Play in new window | Download (Duration: 42:32 — 58.5MB)
Oxford philosophy professor Timothy Williamson talks to Mark, Wes, and Seth about his new book, Overfitting and Heuristics in Philosophy.
Sponsor: Get a $1/month e-commerce trial at shopify.com/pel.
How can we best apply the insights of philosophy of science to philosophy itself? Beginning with his earlier book The Philosophy of Philosophy, Tim has argued that the way philosophy uses alleged counter-examples to refute philosophical theories is naïve. The philosophy of science has moved past Popper’s theory that a single experiment against a theory amounts to a refutation; there’s always a choice whether to change the theory fundamentally, to revise just some peripheral aspect of it relevant to the experiment, or deny the validity or relevance of the experiment. Tim claims this should be applied to philosophy too: When a thought experiment produces alleged counter-evidence to the theory, we should try to figure out the source of the intuitions make the thought experiment look tempting. The first chapter of Tim’s new book outlines several heuristics that people use to generate intuitions and how these heuristics can at times produce bad data.
In the arena of perception, for instance, a heuristic tells us (and Thomas Reid does too!) that when we see something, we take this perception as veridical. This usually works, but in cases of perceptual illusion (a straight stick looking bent when placed in water, a desert mirage), it doesn’t. This doesn’t mean we throw away the heuristic and become skeptics, but we have to be careful and try to diagnose what went wrong with the perceptual-environmental system in cases where the heuristic results in faulty belief.
That was an example of where the failure of a heuristic might produce a skeptical overreaction. A more straightforward case of the heuristic giving rise to a philosophical paradox is what Tim calls “the persistence heuristic,” which is our assumption that things in general don’t change, and in particular that small differences don’t matter. Per the “ship of Theseus” (the sorties paradox), if you change out one plank of a ship, certainly it’s still the same ship, but applying this principle repeatedly could mean you’re calling “the same ship” something that has entirely new planks. Yes, it’s useful for us to overlook small differences in keeping track of the world, and true in most circumstances that a minute change doesn’t matter, but we need to understand that this is merely a heuristic, a rule of thumb, and not a reliable philosophical intuition.
A case where a heuristic could potentially come into play is Williamson’s topic from our ep. 348: knowledge as justified true belief (JTB). Gettier had come up with alleged counter-examples where in unusual circumstances a belief could be true and justified, yet still not knowledge. One might argue that some heuristic gone wrong is responsible for these Gettier cases, and we should just ignore them and say that the JTB theory of knowledge is a good one for all ordinary circumstances. Tim does not pursue this strategy in this case: It’s hard to see which heuristic would be at fault, and JTB is already according to Tim an inelegant patchwork theory that is guilty of “overfitting,” the idea which is central in the second chapter of Tim’s book.
In science, if one makes a theory more and more complex in light of data countering the original theory, then that theory ceases to have much predictive power. If your wildly jagged graph of expected results matches point-for-point the data you actually achieve, then it’s clear you’ve jiggered this “model” to take all your data at face value, whereas it’s actually very likely that some data will be junk: outliers to be dismissed. In ethics, a rule like “you should not lie” might be tweaked and tweaked to include all the intuitively plausible individual cases where a lie seems like it would be morally OK, but a better approach would be to come up with some other way of phrasing the theory, e.g. “you should not lie maliciously” that approaches the simplicity of the original theory.
Another example of overfitting is Paul Grice’s increasingly complicated theory of what “linguistic meaning” amounts to (as discussed in our ep. 325). Tim is convinced that giving these kinds of analyses is actually not what analytic philosophy should be doing. Either such philosophers are giving psychological theories, in which case it’s implausible that the governing psychology could actually be that complicated, or they’re trying to describe something metaphysical (about meaning, knowledge, truth, causation, or whatnot). But why assume that these very important metaphysical aspects of our lives are actually analyzable in more basic terms?
We also talk about the disquotational heuristic (interpreting someone’s words according to our meanings even if the speaker’s knowledge makes their meanings different than ours) to talk about Frege’s sense-reference distinction. Another heuristic Tim gets into in his book involve the metaphor of weighing (e.g. reasons), such that intersectionality properly points out that the relation between anti-female discrimination and anti-black discrimination is not simply additive as the heuristic might tell us. Another involves how we interpret conditionals where the antecedent is false: We interpret “If the earth doesn’t exist, then water is wet” as more reasonable to believe than “If the earth doesn’t exist, then 2+2=5” even though these are both equally (trivially) true given that the earth does exist.
Buy the book. You can read the two chapters we read here and here.
Transcript
0:08 Mark: You're listening to the Partially Examined Life, a podcast by some guys were at one point set on doing philosophy for a living, but then thought better of it. Our question for episode 354 is something like how can we apply insights from the philosophy of science to philosophy itself? We're speaking today to Oxford philosophy professor Timothy Williamson about the first chapters of his new book, Overfitting and Heuristics in Philosophy. More about the text and the podcast, please see partiallyexaminelife.com this is Mark Linsenmayer trying to graph the employment statistics for the analysis industry in Madison, Wisconsin.
0:41 Seth: This is Seth Paskin in Austin, Texas.
0:43 Wes: This is Wes Alwan trying to get fit but not over fit in Cambridge, Massachusetts.
0:48 Tim: Hi, I'm Tim Williamson and I am an Oxford philosopher.
0:51 Mark: Well, thanks so much for coming on the show. We really had enjoyed covering you in the third person in your absence a while ago. And we like to when we do that so to see what the philosopher is doing now. And you had pointed out two different books, plus it seems like you're always doing many, many things. And so this. Can you say a little about how this Overfitting and Heuristics in Philosophy book is fitting in with your overall philosophy of philosophy, your various strands of research?
1:21 Tim: My very general approach to the philosophy of philosophy is that philosophy is not so different from other disciplines, other sciences in a very broad sense of science that includes mathematics and history as well as the natural and social sciences. And so I mean, that's an approach which I call anti exceptionalism. And so in this new book, part of what I'm doing is arguing that some specific ideas which are, you know, quite well known in a lot of the sciences are about certain kinds of intellectual pathology are applicable to philosophy as well.
And overfitting is an example of that. It's a kind of process which pathology, which scientists are very aware of in their own disciplines, where people can take a wrong turning, where they, they just start producing more and more complicated theories to fit the data. And generally you can fit any given set of data with some kind of theory, but then new data come in and that doesn't fit the theory so they complicate the theory still more. And then this just iterates that further data comes in and the complicated theory has to be complicated more.
And I mean it doesn't lead to a theory that in the long run makes good predictions and it leads to a kind of instability. And so that's something that scientists have learned to avoid and they generally avoid it by trying to keep their theories pretty simple. I mean, a very specific reason for that is that if you were in this process of complicating your theories, you're actually adjusting them to data which may contain errors. And so you're kind of ensuring that these errors are built into your own theory as well if there are errors in the data, which very often there are.
And so just talking to scientists, you find that that's a pathology that they're aware of and that they know that they need to avoid. Whereas in philosophy, I think philosophers haven't really generally understood that this kind of thing is a problem, and they've just been willing to go for more and more complicated theories and models, you know, in order to accommodate what we can call the data. I mean, the data really, in many cases are coming from lots and lots of. Of kind of mini thought experiments which philosophers do to test theories to see whether they make reasonable predictions in hypothetical cases.
And so they're taking our kind of supposedly pre theoretic judgments about the hypothetical cases as something like data points which the theory has to fit. And so they've gone down this route of more and more complicated theories, new data coming in from new hypothetical cases which make trouble and leading to further complications of the. And so on. And so part of what I'm doing is just arguing that we need to handle the data in a more sophisticated way, as would be common in the sciences. And that includes, you know, understanding a bit more about where the data are coming from and where there may be errors in the data, which we shouldn't be trying to fit our theories to.
So, I mean, that's as well, that's the general project. And then the other half of the title, the heuristics bit, has to do with. With a view that I develop about where our judgments in hypothetical cases are typically coming from, which is from heuristics that we're using, which are kind of reliable but not infallible, relatively easy to apply methods that we pretty much built into our ways of thinking. And one can show that the kinds of heuristics that we're using cannot be completely reliable. And so that's a specific reason for thinking that there will be some errors in our data, that some of the judgments that we make on hypothetical cases will be wrong.
And so there's something mistaken about a methodology that just insists that we've got to fit all the data with our theories.
5:34 Mark: All right, so I'm hoping we can get to some concrete examples, but yeah. Seth, do you want to give some
5:38 Seth: introductory Yeah, I just want to ask Tim a clarifying question and then just make kind of a comment about the program. So my familiarity with overfitting is in the field of machine learning. So I'm assuming that what you're speaking of, at least by analogy, is similar to what I'm familiar with, with the idea of, you know, you're using a machine learning algorithm to analyze a data set for whatever reason, whether it's correlation or trending or whatever sort of thing. And as you say, overfitting in that world means that you lose the predictive accuracy of the algorithm as you account for more and more of the data.
There's a limit and there's a mathematical ratio that you can implement to determine, you know, there are ways to calculate how much in the sense of overfitting, you're doing. So I think I understand conceptually what you're going after, although I can't say that I followed every single paragraph in the first two chapters of your book with the science, with the logical notation and such. That's not my bailiwick, but I think the question you're raising, I absolutely am on board with the idea that of questioning the analytic philosophical methodology of relying on intuitions and the idea that we might want to question whether intuitions count as data.
So you have two big questions that come out quite clearly in chapter two. Are those data in the same sense that sciences have fields or have realms of discourse or experience that they use as data? And then is the methodology that is being used appropriate for that set of information if it is in fact data? And I think those are absolutely fundamental questions that you're right, haven't really been asked or at least have not been addressed in the mainstream literature. So I found the project very interesting. So we'll see. I'm interested to hear more about what you have to say about, because it feels to me if we can't take what we consider to be philosophical intuitions or the way in which we approach that sort of thing in philosophy as data, then we have to seriously question the methodology and go back to the great methodological thinkers like Descartes in philosophy and talk about, you know, they were very careful about talking about how we should approach this sort of thing, although it typically was not evidence related some in the same way as it was maybe conceptually related.
8:07 Tim: Yeah, I very much don't want to go back to Descartes because, I mean, he had a strategy that perhaps looked promising in his own time, but I think has turned out basically to be hopeless, which Is, I mean, Descartes strategy is to try to find methods which are completely fail safe, which just can't go wrong, and then build up science using those methods. And you know, it's a very familiar thing that of course that he doesn't succeed in doing that himself because he cheats almost straight away by bringing in a dodgy proof of the existence of God and then using that to try to certify all sorts of other methods.
And I think something that has become pretty clear in the period since Descartes is that it's just hopeless to look for ways of doing things which can never go wrong. So that as it were, you know, to get it, to try to get into a state where we're just so good at thinking that we will never make any mistakes. I think what's a much more realistic aim is simply to, you know, obviously we don't want to be slapdashed, we don't want to be making lots of unnecessary mistakes, but to recognize that it's inevitable, inevitable that we will make mistakes from time to time and to be good at identifying when we have made mistakes and then correcting for the mistakes.
You know, and I think that applies to, as with the kind of data that we're, we use in any science that, I mean you, what you can't expect is to have some kind of method which guarantees that there will never be any errors in the data, in the stuff that you're treating as evidence. But what the sciences have generally done is to, as we're refine their methods as they go along to deal with certain kinds of errors that have become a serious problem, but without any aspiration to some kind of Cartesian level of certainty. So I do not think we should be going back to Descartes.
That would fundamentally be a disaster. I don't think on a different aspect of what you were saying, I wouldn't describe the data that we're getting when we think about thought experiments as intuitions because I think that mystifies things too much. As if philosophers were claiming to have some special faculty of intuition that somehow gives us some weird hotline to the truth or something like that.
10:40 Seth: Wait, we don't have that?
10:41 Tim: Unfortunately not. I mean, I think what's going on with these thought experiments is that they're just examples. I mean, they're hypothetical cases and the fact that they're hypothetical isn't such a big deal because it's just often more, you know, more convenient to think of hypothetical medical examples than to have to have to construct them. Yourself. So that, for example, a standard case in epistemology is somebody who forms the belief that it's 3 o' clock by looking at a clock that says 3 o', clock. But it turns out that the clock stopped. But nevertheless, they got the right answer because they happened to look at it at the very time that was showing on the stopped clock. And so seems pretty clear that that isn't a case where you actually know what the time is, even though you have a true belief about the time
11:31 Mark: and you're using a knowledge gathering method that is in general reliable, per Goldman.
11:36 Tim: Right.
11:36 Mark: Looking at clocks is what we do.
11:39 Tim: Yeah, I mean, that was an example that Bertrand Russell put forward. And just hypothetically, but you know, if you could be bothered, you could have a real life case where you could set up, you know, a clock somewhere and then you could secretly observe all the people going past and maybe see somebody setting their watch by this clock and so on. And in that case, you know, you'd be just be making a comment in real time about somebody that, oh, he doesn't really know that it's three o'. Clock. And you know that it's not an intuition, it's just.
It's basically just a report on this guy's situation that he doesn't know what the time is because he's relying on this clock. And the thing is, you know, what philosophers have found is that you don't actually need to do these experiments a lot of the time because you can, you know, just by imagining the case, you can get the same answer. You know. So I don't think that what we're talking about with hypothetical cases is intuition in any special interesting sense. All we're talking about is judgments that we can make in these cases. They're judgments about whether somebody knows something.
I mean, we're making those judgments all the time. You know, when we, you know, I say to you that you don't know what shoes I'm wearing because you can't see my feet during this podcast. I mean, it's about knowledge, but it's just a completely straightforward judgment about the situation. And then it turns out that we can do the same thing without actually having to set up the real life situation by imagining it and get corresponding answers. And that's really what's happening with philosophical thought experiments. And of course, I mean, some of them would be unethical to set up in real life because they involve putting people in situations where either one person is killed or five are killed and that sort of thing.
So I Think we're just getting judgments about what would be the case in these possible situations. And normally speaking, the kind of judgments that we're making would be things that we ourselves know. And so they would. It would be fine to use them as evidence. But there are tricky cases where things can go wrong, where there's some kind of glitch in our cognitive system, or where we're using a heuristic that usually gives the right answer, but can give the wrong answer in tricky cases. And if that's what's going on, then some errors are going to creep into the data.
But I don't think that there's some huge problem about whether the evidence that you can get from just from considering hypothetical cases is. I mean, whether that's genuine evidence. I mean, of course, it's not quantitative evidence and it's fallible, but quantitative evidence is fallible, too. So I think the problem is that we've got quite a good capacity for working out what would be the case in hypothetical situations. I mean, we're using that all the time when we're making decisions, by the way. I mean, you know, when we make decisions, we have to think hypothetically about what will happen if I do this versus what will happen if I do that.
And, you know, and often we know what will happen in those cases. So this kind of ability to make judgments about hypothetical cases and in fact to have quite a lot of knowledge about what would be the case in hypothetical cases, it's not in principle problematic, it's just that it can sometimes go wrong. And then if we tailor our philosophical theories to fit the results, that they're going to go wrong too.
14:58 Wes: So should we talk about heuristics? Mark, you were just about to say something.
15:01 Mark: Yeah, I'm wanting to. I have a list of sort of different philosophical problems that we've done in past episodes that we could sort of apply the idea of heuristics, the idea of overfitting to.
15:09 Wes: But yeah, yeah, let's talk about heuristics and how this applies to philosophy. So you describe a heuristic as a rule of thumb for solving a problem. An example is smelling. So it's a reliable method, but it's not. It's generally reliable, but it's not perfectly reliable. So it's prone to error.
15:25 Tim: Sometimes smelling something to smelling something to see if it's bad. Yes.
15:30 Wes: Yeah. Perfectly good most of the time. Perfectly good method for seeing if food is bad, but not fully a reliable test.
15:38 Mark: I would smell for the COVID I'd see If I could detect the COVID on it.
15:43 Wes: Visual illusions are another example. So we have heuristics built into are cognitive processes that we're not even conscious of. Right. So we're making visual judgments and for the most part that works fine. But visual illusions show us that those heuristics in certain cases fail or can be taken advantage of by a certain situation. And then we get into more philosophical examples of heuristics like, and this gets more abstract very quickly. Like persistence, for instance.
16:16 Tim: Yeah, so the persistence heuristic is very roughly, it's just saying that you can ignore small differences, don't bother with them. So that normally speaking something that you learned yesterday, as if it's true about if you learned yesterday or you confirmed yesterday, let's say that Vladimir Putin is the president of Russia, then you describe yourself as remembering that he's the president of Russia, even though you're saying that a day later when you know things could have changed, he could have amazingly resigned in the meantime or been assassinated or whatever.
And so what you're doing there is effectively just ignoring the passage of time. You're treating this piece of data that you acquired yesterday as something that is still good today, even though there is an imaginable possibility where it's gone wrong. And so, I mean, that's something that databases in effect are have to do all the time, because you can't be constantly checking everything. And what you need usually is information about the present, not just information about the past. And this also applies to treating things that differ in only some very, very trivial way as effectively the same.
So that, for example, if you imagine a heap of sand, and then you imagine, you know, one grain being taken away from it, then it seems like a no brainer that it will still be a heap of sand there. And that's just the kind of application of the persistence heuristic that you can ignore small differences, but what you've in effect done is, if you think of it as a general principle, is to commit yourself to the principle that when you take a grain away from a heap, the result is still a heap. And if you iterate that, if you apply that over and over again, you obviously get into some kind of absurdity when you, if you start with a heap with 10,000 grains and then you keep subtracting grain after grain, as the ancient Greeks realized, you end up by this principle saying that the heap only got three grains left and 2 and 1 and 0.
And that's actually a philosophical paradox. And people have suggested that we need to change logic and change our theory of truth and falsity and go to fuzzy logic with degrees of truth or all sorts of things just to, you know, to handle this kind of case. But I think that all that's going on is that we're applying this heuristic of ignoring small differences where it's not meant to be applied, where it leads to these absurd conclusions. And I think, you know, the reason it feels like a paradox is because we're not really aware of some particular line, some particular cutoff point for heaps.
And so the fact that we're using a heuristic never really is made salient until it's too late. And so it's this kind of argument that leads to the paradox. I mean, it feels quite intuitive, but the reason it feels quite intuitive is because it's using a heuristic that we use all the time. You know, it's not, I mean, it's not 100% reliable heuristic.
19:26 Wes: So the heuristic is that if you have a heap and you take away a grain, it's still a heap, but if you iterate, then the heuristic fails at a certain point.
19:35 Tim: Yeah, but the general heuristic is just to ignore small differences, right?
19:39 Seth: Yeah.
19:40 Mark: So this has come up in, like, if you have a brain and you replace neuron by neuron, you replace it with silicon. Does consciousness go away at some point, for instance, or is there some gradual fading of consciousness or something that anytime we're dealing with just one little bit at a time, this heuristic is coming into play? It doesn't necessarily. Realizing that is a heuristic and not an a priori truth or something like that doesn't necessarily really solve the philosophical problem in question. Right. In this case, I don't know that it tells you what consciousness is.
Will it still be there between, you know, the full biological brain and the full silicon brain that mimics it? It's every structural thing, but it seems like it can help. I don't know. Do you have any. That's not an example specifically that you brought up, but it came to mind because we've talked about that in terms of David Chalmers and things. Do you feel like that this heuristic provides any insight to that?
20:29 Tim: Well, I mean, somebody might try to use an example like that to, quote, prove that a silicon brain would be conscious, you know, if they were claiming, you know, if they thought that just, you know, difference of one molecule couldn't be enough to make the difference between consciousness being present and Consciousness not being present. And so of course, once you've realized how similar this is to these other kinds of reasoning, you know, like the one about the heap, you're going to be very suspicious of an argument like that, and in particular suspicious of the assumption that a difference of one molecule can't make the difference between consciousness and lack of consciousness.
21:14 Seth: So is the issue here that we're dealing in the realm of heuristics? And the error is that we think we're working in the realm of laws and that we're trying to apply a stringent set of requirements for the heuristic or for the behavior to create an infallible generalization, or at least a scientifically valid generalization, when in fact we should just embrace the fact that that's not the way these things work. And the way that we treat them with heuristics is the appropriate methodology. And we should spend more time working out how heuristics work as opposed to thinking that what we should be doing is the traditional philosophical enterprise.
21:54 Tim: Yeah, because the thing with heuristics is that you can be using it without realizing that it is a heuristic. And I don't think people have fully come to terms with the fact that an awful lot of our thinking is of that kind. And so, you know, when we're using a heuristic in a case like this, it can easily feel like sort of intuitively compelling as an argument. And, you know, because there is no kind of health warning on a heuristic that this, using this heuristic could be dangerous for your health or something. And so it feels as though what we're doing is just kind of informal but compelling argument. But what we're doing is actually just is using a heuristic. And so we have to be a bit more suspicious of these arguments, you know, that seem compelling in this kind of way.
22:40 Mark: I wonder if there's another philosophical example. So frege on sense versus reference to an episode we did many years ago, that you've got another one of these heuristics that is supposed to explain why this is a problem. That is, somebody thinks the evening star was up during the evening, but it was not up during the morning. And then the same person who doesn't know that the morning star and the evening star are both, in fact, Venus says the morning star was up during the morning, but it was not up during the evening. And if you just don't take into account the fact that they didn't know this, then you would say, oh, this person believes that the planet Venus was both up and not up in the morning and was both up and not up in the evening.
This. This person has conflicting beliefs. And the heuristic that you appeal to there is that we walk around with these mental maps of the state of the world. It's hard enough for us to keep track of the world if we had to add keeping track of everybody else's mental maps. So there's this disquotational. Is that the name of the heuristic?
23:40 Tim: Disquotational heuristic, yes.
23:41 Mark: Where we just assume that the words somebody says, we just convert those to what we understand the meanings to be. So we understand both the meaning of evening star and morning star to be Venus because we are astronomically educated and attribute that. And we can see why we've then made the mistake in thinking that this person holds an explicit contradiction. So does that have we done away with or. It seems like we. I was never clear whether that Frege. It seemed like we were setting up that we need a difference between the referent of two terms, which are both Venus and the sense of two terms, which because you don't know, because not everybody is just by linguistic competence would know that the morning star and the evening star are the same.
Then there is something. There's some additional thing, but it wasn't necessarily. Here's a problem that needs to be solved. And now that we understand the disquotational heuristic, we whoosh, we've solved the problem. Like what is adding the idea of a heuristic add to if, you know, Frege was thinking in these terms at that point.
24:38 Tim: Yeah. So Frege used these examples to say that we have to distinguish between reference and sense. And sense is something like what the word means for us or something. And the theory that, you know, and these are meant to be cases where the names Hesperus and Phosphorus, they have the same reference. They both refer to the planet Venus, but they different senses. One of them presents it as the evening star and one as the morning star. So originally, I think Frege's theory looked like a pretty good try at making sense of what's going on. But there is people, notably Saul Kripke, I think, shown that Frege's approach doesn't really get to the heart of what's going on.
And one reason for that is that you can have these kinds of examples where in a way there isn't any difference in meaning. You can have these examples simply because there's a difference in word. So that one case, that is the words furze and gorse, which are in fact synonyms, they're just two words for the very same natural kind, you know, kind of shrub. So there is no difference between them. But for even for somebody who's kind of quite familiar with these words, there is a little bit of kind of room for doubt. Not because you associate, you have different mental associations with the words apart, but, you know, you could just think to yourself, well, look, I've been assuming that furze is gorse, but maybe there is some subtle difference between them.
Maybe they're not actually the same thing. You know, you wouldn't have to have some special theory about what the difference was or anything like that. It's just enough that, you know, once you've got the two words, you can start having doubts about whether furze is gorse, you know, as you would put it, but without having any doubts about whether the furze is furze, and so that your level of confidence is different for the two sentences. Furze is furze and furze is gorse, even though there's actually no difference in meaning between the sentences whatsoever, even for you.
Because, as I say, you don't have different mental associations. It's simply a difference of word. And so I think that Frege's whole project of trying to sort of construct a level of sense that's intermediate between words and the things out there in the environment that the words refer to doesn't really get at any kind of stable intermediary level. Because the kind of phenomena that he's appealing to, these kind of doubts about whether it's the same thing or not, all you need to create those doubts is a difference in the words. You don't need some difference in their meanings.
27:27 Wes: So should we connect over heuristics to overfitting and see how that might help us in philosophy? I mean, I take it to be the heuristics. We produce these interesting paradoxes or counter examples when the heuristic fails. So it's like taking a bad data point seriously. And then we tried to build out these epicycles or complexities into our theory to account for that data point. And it sounds like you're suggesting, well, this is actually just a place where the heuristic doesn't work and we shouldn't be trying to fit that in. So how does that help us with, say, something like the Gettier problem, so the Gettier problem?
28:05 Tim: Well, that's traditionally the examples of what is supposedly justified. True belief without knowledge and so there was this popular theory that analyzed knowledge as justified true belief.
28:19 Mark: And we can just point at the clock example from Russell that you already gave, that the person has a belief that is, that's three o'. Clock. It is true, that is three o'.
28:27 Seth: Clock.
28:27 Mark: And they're justified in that they're looking at a clock, which is what one normally does to justify these things. Okay, go ahead.
28:34 Tim: I mean, then, I mean, we wouldn't say that the way that they formed their judgment was irrational or anything like that. So, you know, it's true that one way you could try to use the arguments that I'm giving in the book is to say, well, maybe the justified true belief analysis was okay after all. And there's some kind of heuristic going on here which is saying that this isn't knowledge, but that there's something wrong with the heuristic. So in this case, that's actually not the moral I want to draw. I don't think there is an error in the gettier cases, but I think what the arguments of the book do suggest is that we shouldn't be resting everything on a single type of counterexample like this.
I mean, of course there are all sorts of variations on the theme, but we should at least consider the point possibility that there's something, as it were, in our cognitive apparatus which is sending us wrong here. And so. Well, I've done some work on this where partly having a look at whether the justified true belief analysis is such a great theory anyway, because people tend to present it as, you know, here's this fantastic theory of knowledge which unfortunately was defeated by some nasty counterexamples. But wouldn't it be great if we could have it back again?
And I think when you look more closely at the justified true belief account, it's actually for different reasons, it's not really a good account at all. I mean, I've done some work using mathematical models to kind of model the way that the justified true belief account would work. And it actually comes out as a very weird account and not one that would make knowledge be a particularly useful category or anything like that. Actually, it's not such a great account. I mean, in fact, I think it's when you're teaching, you know, you're told that one way of attracting students attention and getting them to take notes is if you say there are three points that you need to know about such and such, and then you see them all kind of getting more alert.
And so it's a kind of, there's Something about three bullet points that seems to appeal to human psychology. And then the justified true belief account, it's kind of, it's attractive because it's a three bullet point analysis of knowledge, but it's, but I mean, there isn't much more to be more that's good about it than that. So, you know, I think in a way it's not an attractive view. And you know, for rather different reasons I've argued actually that we shouldn't be trying to analyze knowledge in more basic terms. And in fact that, you know, the whole category of justification that is being cooked up here is a somewhat dodgy one.
So, you know, in that case I don't think the moral is we should go back to the justified true belief account. But definitely it made sense to have another look at the account, given the kinds of argument that I'm making. Just relying, you know, putting all your eggs in one basket and just relying on the counterexamples is a bit risky.
31:28 Wes: Well, maybe let's give a better example then of philosophical overfitting. I mean, I think you were saying this is kind of a feature of conceptual analysis.
31:38 Mark: Yeah.
31:38 Tim: And then what you find is that, I mean, you can find a lot of overfitting, not so much in the original justified true belief account, but in the kind of all the modifications.
31:48 Wes: Right, that's what I was thinking of and trying to save it. Yeah. Okay.
31:51 Tim: And so they were starting to produce, add all sorts of epicycles, some extra conditions, and then the extra conditions might be a bit too strong and then they had to be watered down with some kind of water.
32:05 Wes: We're talking about reliabilism, aren't we?
32:06 Tim: In general there were reliabilist accounts of knowledge, but I mean, there were also much more internalist accounts of knowledge which had to do with the structure of the justification rather than reliability. And I mean, none of those have really turned out to be good analyses of knowledge by the standard of the hypothetical cases. And the kind of proliferating complexity that you find in that research program is absolutely typical of overfitting and the fact that it never stabilizes. And there are some other reasons as well for in this kind of case, if what somebody's interested in is understanding the kind of natural category of knowledge is why making it very complicated seems implausible.
Because the category of knowledge, I mean, you know, do you know whether it's going to rain this afternoon? Those kind of, you know, they're just very ordinary uses. You know, the idea that when we say no, that we're no There we're referring to this kind of incredibly complicated thing that ordinary people would need a lawyer to tell them whether this analysis
33:16 Wes: applied in a given knowledge scored.
33:18 Tim: Yes. You know, there's no way that something like that is going to be one of our basic categories, and there's no point in having that kind of sort of ramshackle sort of category as one of our basic instruments for thinking.
33:32 Mark: With another example philosophical problem, we recently, within the last year, covered Paul Grice on the show. So he has a notion of meaning, right? The initial definition, meaning a uttered X with the intention of inducing a belief by means of the recognition of this intention. So already that sounds pretty complicated, that I want you to realize that I'm trying to tell you something, and I want you to have this belief of the thing that I'm telling you, and I want you to have it because you recognize I'm trying to get you to believe it. And he already has that level of complexity because of various cases where I merely let something slip and.
Okay, well, did you mean to tell me that. No, no, you just merely let it slip. Okay, so we want to restricted to. I'm talking to you, I'm giving you some information with the intention that you have the information. You recognize that we're engaged in that interplay. But then he goes on in, for instance, this article Utterer's Meaning and Intentions to give all these outlandish scenarios in which he has to keep adding things to this formula, making it longer and longer and longer so that it can fit all the individual, really weird scenarios of people having partial knowledge and overhearing things.
And there are third and fourth people involved, and I forget all the details. So this seems like a pretty classic example of overfitting. Like maybe that initial definition I gave or even something simpler might actually be better. Is that the idea?
35:02 Tim: Yeah, but it also might be that we shouldn't be trying to analyze it at all, because, you know, I mean, philosophers got into this kind of way of thinking where their job was to find analyses of things in more basic terms. But it's not at all obvious that with many of these categories, I mean, meaning is one example, causation is another, that we actually have more basic terms that we can give, you know, an analysis of them into. And so it's not clear why we should be setting ourselves up with this kind of project, because the track record of these attempts at analysis is very poor, you know, and it doesn't seem that they're kind of converging on the truth.
I mean, they're just in some ways they seem to be getting further and further away from it with all these extra layers of complexity. And I think that was at the time when philosophy was following something like the so called linguistic turn, where people felt that analyzing meanings as philosophers should do. You know, they felt they should be producing these kind of analyses. But that isn't the way that most philosophers now think of what they're doing. And it might just be better to produce. I mean we can produce theories about meaning, about how meaning works and so on, but the theories don't have to take the form of somehow giving necessary and sufficient conditions for something.
In more basic terms, I mean, they can be giving important information about without doing that.
36:31 Seth: There's another point I've been waiting to bring up and I guess now's as good a time as any. Okay, so Descartes approach is wrong. You've identified the way in which this basic conceptual analysis, the program of, let's just call it early to mid 20th century analytic philosophy, is misguided. You mention in chapter one and kind of circle back to it in chapter two and in some of the footnotes that there is a program of analysis of human heuristics in the field of psychology and that psychologists and philosophers should maybe collaborate a bit more.
You touch on experimental philosophy as being somewhat misguided, which I agree with. So what is the program of philosophy then? Taking into account the idea that we've got a problem with the way in which we approach our quote unquote data, we don't have. We haven't spent the time to analyze our methodology and even make clear what our research program is. We have to take into account the way in which we heuristics aren't just theoretical, right? These are behavioral. This is a real tangible way in which human beings cope with the world. So there's a very real phenomenal way in which these heuristics manifest and there's obviously developmental way in which they get created.
For example, you have to have experience to get a persistence idea. So in a broader picture, what is it that we should be doing or where do you want to move towards from the way we think about data and our methodological approach? Or is philosophy just out of, step out of date? And we need to sacrifice, you know, we need to just turn everything over to psychology and, and the hard sciences, the social sciences?
38:21 Tim: No, I very much don't think we should turn it over to, to them. I mean, I think we can plenty of room for collaboration between philosophers and psychologists. But I Think the kinds of question that philosophers are interested in isn't one that you can. They're not ones which. That you can answer by purely psychological means. So, for example, in epistemology, well, one typical philosopher's question is whether if you know something, you know that you know it. And I don't think that's a question to be answered just by doing experiments or something.
I mean, I don't think there's any obvious way you could answer it was by doing it, experiments. I mean, it's true that the psychologists in the end are the ones who are doing tests to find out what heuristics humans are using. And what I'm doing in the book is making some speculative claims about what general cognitive heuristics we might be using. I mean, giving some evidence that we are using them and then saying how they can mislead us. But I think in the long run, those claims need to be tested by psychologically. Although it's quite difficult to have tests because there are so many different factors in play that you have to screen away from.
But all that will give us in the end, let's say, is some knowledge of our own cognitive apparatus. But what we're trying to do in philosophy is to answer questions which are not just about our own psychological apparatus, but are about reality and how our thought corresponds to reality and that kind of. Of thing, which is not really a psychologist's job. It's just that. Yeah, I mean, one way of putting it is that these heuristics, you can think of them as a bit like. I mean, the downside of them is that they in fact embody certain kinds of biases.
Could put like that. And then we have to be aware of our biases when we're investigating the world. But just being aware of our biases doesn't answer the philosophical questions any more than it answers the cycle, the physical question. I mean, the kind of questions that, you know, you might be interested in about, for example, you know, are there things which are possible even though they're not physically possible? I mean, could it be, for example, there could have been different laws of nature from the actual ones? I mean, that's not a purely scientific question because science is concerned with our laws of nature.
It's not concerned with whether there could have been a quite different set of laws of nature. And, you know, I mean, psychology can't answer a question like that.
40:54 Mark: I think we could get into more examples. You can get part two, where we continue talking with Tim right now, if you're A partially examined life supporter. Go to partialexaminedlife.com support if you want to become one. Otherwise, we'll see you next week. Thanks.
41:07 Wes: Thank you.
41:08 Seth: Thanks.
Leave a Reply