December 14, 2002
A Problem for Process Reliabilism
The following strikes me as a pretty
persuasive argument against a thorough-going process reliabilism. Since I’m no
expert on the field, I don’t know how similar it is to existing arguments
against process reliabilism, which is to say that if this turns out to be a
boring repetition of familiar points, well at least it wasn’t intentional
plagiarism.
Process reliabilism says that the
justification of a belief is proportional to the reliability of the process
that generated the belief. This raises the generality problem, as stressed in
Conee and Feldman’s 1998 paper - what is the process by which the belief
is generated? Or, to put the point more obscurely, what are the individuation
conditions for process types being used in this formulation. At one level the
generality problem is the problem of making the basic claim of process
reliabilism contentful - if we are prepared to count gruesome enough types,
then every belief is the product of some very reliable processes, and some very
unreliable processes. But let’s assume that problem has been handled.
At another level, the generality problem
raises a tension that I think can’t be resolved for a full-blown process
reliabilist. On the one hand, we want processes to be instantiated more than
one time, or else we’ll be led to the crazy view that a belief is justified iff
it is true. So we don’t want the instantiation to be too fine-grained.
On the other hand, the definition of justification entails rather immediately
(so immediately that it might surprise you to learn how long it took me to
realise this) every belief generated by the same process is equally justified. To
the extent that justificatory status can be very sensitive to the particular
ways a belief is formed, that implies we want processes to be individuated
quite finely. I think, and I think I have an example that supports this, that
these two constraints can’t be satisfied at once. Onto the example…
DIAGNOSIS
Morgan is
displaying symptoms S. Dr Watson knows that symptoms S normally
imply that the patient has a liver disease. But he also knows that in some
cases, happily enough in all and only cases where the patient has genetic condition
C, a patient with symptoms S doesn’t have a liver disease, but in
fact has a kidney disease. Dr. Watson also knows that genetic condition C
is rare, only 1% of males and 7% of females are C. And he knows that
there’s no easy way to test for whether a patient has condition C, for
usually it has no readily observable effects. And he knows he has no other
relevant information about whether Morgan is has condition C. So Watson
concludes that Morgan has a liver disease.
How justified is Dr. Watson’s belief?
I think you don’t know enough to say yet,
because you don’t know whether Morgan is male or female. If Morgan is male,
then Watson’s belief is very well justified. If Morgan is female, then Watson’s
belief isn’t particularly well justified, for he should be taking more
seriously the possibility that Morgan has condition C. Even in that case, it isn’t a disastrous
belief, but not as well justified as in the case where Morgan is male. Since
the two possible beliefs are not equally well justified, we need to say that
they are the results of different processes.
That alone might not be a problem. Perhaps
we can find a different way of categorising beliefs such that the belief that a
male patient displaying S has a liver disease falls into a different
category than the belief that a female patient displaying S has a liver
disease, though I’m not entirely convinced that existing (pure) reliabilist
theories have the resources to do this.
The problem is that the example generalises.
If x and y are both relatively small numbers, and Watson knows
that x% of males have condition C and y% of females do,
then his conclusion that Morgan has a liver disease is more justified if Morgan
is male rather than female for any such x and y, even if they are very
close, say x = 4.5 and y = 5, or even, I’d guess, if x =
4.5 and y = 4.51.
That means that we’re going to have to posit
infinitely many different categories of belief-forming processes, just to
account for all the different possible processes via which Watson could form
the belief that Morgan has a liver disease. The problem is that when categories
belief-forming processes get so fine-grained, we will start to get some
lucky guesses counting as justified beliefs, because they are the only beliefs
ever formed by that process, and some unlucky reasoned judgments counting as unjustified
beliefs, again because of the small sample size. This I take it should be
intolerable.
One response to related problems raised in
the 1980s was to modalise the notion of reliability. Maybe I’ll come back to
that in later posts, but I think it should be pretty clear that won’t help. The
problem is that there’s too many darn worlds to possibly count successes and
failures of a process, and no other approach to summarising the data from
nearby possible worlds seems to be much use.
This is not a problem for theories of
justification that incorporate some aspects of process reliabilism, but also
build in some more traditional internalist evaluations of modes of reasoning.
Ernie Sosa’s virtue reliabilism is like this, and my theory, which is
reliabilist about observational beliefs and (sorta kinda) foundationalist about
non-observational beliefs isn’t either. But a theory that is all process
reliabilism all the time really looks like it has problems with DIAGNOSIS.
