In the April edition of Analysis, Hilary Kornblith proposes a reliabilist solution to the bootstrapping problem. I’m going to argue that Kornblith’s proposal, far from solving the bootstrapping problem, in fact makes the problem much harder for the reliabilist to solve. Indeed, I’m going to argue that Kornblith’s considerations give us a way to develop a quick reductio of a certain kind of reliabilism.
Let’s start with a crude statement of the problem. The bootstrapper, call them S, looks at a device D1 that happens to be reliable, though at this stage S doesn’t know this. We assume that S is a reliable reader of devices. S then draws the following conclusions.
(1a) D1 says that p1 at t1 (since at t1 D1 appears to say p1).
(2a) p1 is true at t1 (since D1 says that p1 at t1).
(3a) D1 is accurate at t1 (by deductive inference from (1a) and (2a)).
Note that S need not, and in the story does not, know that the grounds for (1a) and (2a) are good grounds. S does, we’ll suppose, know that (1a) and (2a) entail (3a). If reliabilism is true, then it seems S knows, or at least justifiably believes, that (1a), (2a) and (3a) are true, since each belief was produced by a reliable process. S then repeats the process a few more times, each time going through a version of the following triptych of inferences.
(1z) D1 says that p26 at t26 (since at t26 D1 appears to say p26).
(2z) p1 is true at t26 (since D1 says that p26 at t26).
(3z) D1 is accurate at t26 (by deductive inference from (26a) and (2a)).
From (3a) through (3z), S infers (4).
(4) D1 has been accurate the last 26 times I’ve used it.
And from (4), S infers (5).
(5) D1 is generally accurate.
It isn’t clear whether Kornblith thinks the problem for reliabilism is that it lets S infer (4) or that it lets S infer (5). As we’ll see, the response he offers is a reason for S to not infer either (4) or (5).
In any case, it isn’t clear that the inference from (4) to (5) is in any sense reliable. It’s true that S has a lot of information that D1 has worked well. On the other hand, if D1 were disfunctional, S would not have that information. In fact, if D1 had been unreliable at, say, t8, then S would not have any evidence about how accurate D1 was at t8, since the relevant step, (2h) as it turns out, would fail. Since S’s information harvesting technique can only produce evidence when D1 works, and not when it doesn’t work, it isn’t clear that the fact that the evidence base includes only cases where D1 works is evidence that D1 is generally reliable. So I’ll focus here on the worry that S shouldn’t be able to infer (4) by using D1 itself. (The points in this paragraph draw on some considerations raised by Jonathan Weisberg (forthcoming).)
Kornblith’s point is that the process S uses to get to (4) is, overall, an unreliable process. For imagine that S then goes through the same reasoning with D2, which is in fact unreliable.
(6a) D2 says that p1 at t1 (since at t1 D2 appears to say p1).
(7a) p1 is true at t1 (since D2 says that p1 at t1).
(8a) D2 is accurate at t1 (by deductive inference from (1a) and (2a)).
(6z) D2 says that p26 at t26 (since at t26 D2 appears to say p26).
(7z) p1 is true at t26 (since D2 says that p26 at t26).
(8z) D2 is accurate at t26 (by deductive inference from (26a) and (2a)).
(9) D2 has been accurate the last 26 times I’ve used it (deductive inference from (8a)-(8z)).
Kornblith says that S uses the same process to get to (9) as to get to (4). And that process is clearly an unreliable process, since it doesn’t distinguish between (4) and (9). So S isn’t justified in believing (4), since the process that produced (4) is unreliable.
I don’t think this is properly responsive to the problem. The worry was that S was using a reliable process to derive (4), and hence S’s belief that (4) was unjustified, contrary to a strong intuition that it is not. Kornblith’s response is that S is using an unreliable process to derive (4), and so S’s belief is unjustified. But what the reliabilist needs is that S is not using a reliable process, not merely that S is using an unreliable process. The former claim does not follow from the latter, if S is using more than one process. And, worryingly for the reliabilist, that seems to be exactly what’s going on here.
If I’m preparing beans and rice to eat, there is a process I run through to prepare the meal. That process has some subprocesses; there is at least one for the beans and one for the rice. Let’s assume that I make the beans first, then the rice. Then when I finish preparing the rice, I finish two processes at once, the rice-preparing process and the meal-preparing process. There’s nothing particularly special about this case; any particular action can be the conclusion of any number of processes. For similar reasons, any particular belief can be the termination of any number of cognitive processes. Some of these may be quite short processes, some of them longer processes.
That seems to be what’s going on in S’s case. When S reaches step 4, two processes terminate. One is the very long, and very unreliable, process of figuring out whether D1 is reliable by comparing D1’s outputs with what we know about the world via D1. Another is the very short, and very reliable, process, of drawing logical conclusions from what S has come to know through (2z).
Let’s assume, with Kornblith, that the long process is really a process. And let’s also assume the following two conditionals, (R1) and (R2), which are characteristic of a certain kind of reliabilism.
(R1) Any belief is justified if it is the outcome of a reliable process.
(R2) Any belief is unjustified if it is the outcome of an unreliable process.
I claim these assumptions lead to a contradiction. For S’s belief in (4) is the outcome of two processes, the last two steps of which overlap, one of which is reliable, and the other of which is unreliable. So it is both the outcome of a reliable process, and the outcome of an unreliable process. That is no contradiction, any more than it is a contradiction to say that the belief is the outcome of both a long process and a short process. What is a contradiction is to say that the belief is both justified and unjustified. But that’s just what follows from our earlier conclusion, plus the conditionals connecting reliability to justification. So I conclude that (R1) and (R2) can’t both be true.
To be sure, that doesn’t mean that reliabilism is doomed in all shapes, since there are plenty of reliabilist theories that don’t endorse both (R1) and (R2). Alvin Goldman (forthcoming), for instance, says that justification requires reliability and evidential support. That implies that (R2) is true, but (R1) is false. Of course, any theory that rejects (R1) has no need for a fancy solution to the bootstrapping problem, since they are not required to say that (2a) was a legitimate step.
So I conclude that Kornblith’s attempt to save reliabilism from the bootstrapping problem in fact leads strong versions of reliabilism, those that accept both (R1) and (R2), into contradiction. Weaker versions of reliabilism, including versions which qualify (R1), avoid the contradiction, but also have more direct means of avoiding the bootstrapping problem in the first place.
Goldman, Alvin (forthcoming) “Toward a Synthesis of Reliabilism and Evidentialism? Or: Evidentialism’s Problems, Reliabilism’s Rescue Package,” to appear in T. Doughterty, ed., Evidentialism and Its Discontents, Oxford University Press.
Kornblith, Hilary (2009) “A Reliabilist Solution to the Problem of Promiscuous Bootstrapping”, Analysis 69: 2, 263-267.
Weisberg, Jonathan (forthcoming) “Bootstrapping in General“, presented at the 2009 Rutgers Epistemology Conference (and I imagine forthcoming in PPR)