Wednesday, May 20, 2009

Review of 'Knowledge, Belief and Reasons for Acting': Jennifer Hornsby

Here I assess and evaluate the first section from Hornsby's paper where she tries to support the claim that knowledge is necessary for objective reasons to occur as reasons in a reason-giving explanation of the agent's activity. In the end I argue that her argument fails to establish this and that her formulation of the principle that governs acting for objective reasons must be revised. Yet, I remain sympathetic to her suggestion and think that arguments can be supplied to support knowledge's essential role in reason-giving explanations although I leave it to future work coming up with such a principle.

In this paper Hornsby tries to find grounds for thinking that an agent’s possession of knowledge is presupposed by the agent acting for reasons and thus for the claim that acting for reasons does not come into play unless the agent has knowledge. The tendency to explore knowledge’s role in action theory is one she shares with a number of other philosophers: Jason Stanley suggests in his paper together with Tim Williamson that knowing that p is a reason for F-ing is a necessary condition for rationally F-ing whereas Stanley together with John Hawthorne takes the step a bit further and explores and defends the idea that knowing that p is a reason to F is both necessary and sufficient for rationally F-ing. On the other hand, we have people like John Gibbons who thinks that intentional action without knowledge is impossible and thus that knowledge is presupposed in some form or another whenever one says of some agent that she intentionally F-d. Hornsby seems to be on roughly the same track as Gibbons since she too is exploring the metaphysical foundation for actions rather than merely asking, like Stanley, Hawthorne and Williamson, about the norms or ethical principles that govern rational conduct. Important as the latter question is Hornsby sees herself as going further than the normative question to pose questions about the metaphysical constitution of actions.

Hornsby seems to start out by picking up a clue from Williamson’s suggestion (2000, p. 62) that knowledge sometimes must figure in the best explanation for why some agent F-d. According to him, attributions of knowledge may be a better predictor for determining someone’s actions by lending more probability to a certain way of conduct. The intuitive example is the rational burglar who risks a lot by searching the whole building for a valuable diamond. The only way to understand why a burglar would take such a risk is, according to Williamson, by attributing her with the knowledge that the diamond is in the building. Otherwise it would be hard to explain why she apparently disregards evidence to the contrary—i.e. as the time goes and her search is not successful—on pain of diminishing her rationality (like declaring her to be just plain stubborn, insensitive to evidence etc.). Another example would be the case where someone, say your mother-in-law, comes to your door ringing the door bell and you consider whether to open the door or not. The outcome of one’s deliberation should depend on whether you have reason to think that she knows that you’re home or whether she has a mere true belief to this effect. In other words, the predictive outcome—i.e. whether your mother-in-law becomes insulted or just disappointed—depends on the presence or absence of knowledge. The relevant generalization needed for explaining these cases—which, by the way, is how Williamson likes to think about the notion of causality—thus seems to depend on knowledge in certain intuitive cases of human conduct. In this sense knowledge is causally and psychologically relevant for human conduct.

However, if such cases goes through they suffice to show that there are at least some cases where attributing knowledge to the agent provides the best explanation of his or her behaviour. Thus, knowledge is sometimes needed in psychological explanation since in those cases it would severely decrease the explanatory power of psychology if we were to restrict that discipline from attributing cognitive states like knowledge to the agent. Beliefs won’t suffice to rationalize or to provide a reason-giving explanation here so, pace Stephen Stich and psychology’s restriction to autonomous behavioural description (i.e. a description of a way of acting such that if you would act in a certain way in a given setting so would your replica that shares all your current, internal, physical properties (Stich, Folk Psychology, p. 167)), we should think that what knowledge adds to beliefs is psychologically relevant. (Of course, one could reply here that those limits to psychological explanations might be just what we should expect since this discipline is, after all, not an attempt to explain everything. Agreed, however, I do think that Gibbons’ paper on this topic provides us with plenty of cases that one would like to be explained psychologically and where one nevertheless would fail to do so without attributing the agent with knowledge. Thus we should reject Stich’s restriction to autonomous behavioural descriptions. The point is that since which psychological states there are is at least partly determined by what one needs in order to explain human behaviour it seems arbitrary to exclude knowledge as psychologically irrelevant unless one can provide a more principled distinction between which of these states are psychologically relevant and not: again, the failure of such attempts is discussed in both Gibbons (2001) and Williamson (2000, chapters 2 and 3)).

What Hornsby wants to do is to take these ideas a bit further and actually put knowledge into the constitution a specific kind of agency, namely the agency that goes by the label acting for reasons; or, as she puts it, “until it is allowed that our knowing things explain our acting, our acting for reasons is not in view”. In order to get there we need to make a couple of preliminary distinctions. An intuitive and much-discussed distinction in reason-explanations goes between the subjective sense in which something is a reason for F-ing and the objective sense in which something can be a reason. The latter is, according to Hornsby, a matter of fact that obtains regardless of whether the agent is considering that reason as a reason for F-ing. The former is another kind of fact: namely the fact that the agent considers some p (whether true, false or plain stupid) to be a reason for F-ing. The following schemas cash out this distinction:

OBJ: A reason for A to F was that p: p
SUBJ: A had a reason to F: she believed that p: Bp

The distinction can be appreciated with an example: suppose A goes skating on the edge of a pond and clearly avoids skating in the middle of it. Suppose further that the ice is too thin for skating in the middle of the pond. Now, according to Hornsby, that fact is a reason for A to skate on the edge of the pond as well as avoiding the middle of it. For short, call that fact P. A may now be acting for a reason in the objective sense, thus skating in the middle of the pond for the objective reason that P. On the other hand, there’s another way in which A may have a reason for skating at the edge of the pond and that is the subjective sense in which she believes the ice in the middle of the pond is thin. What’s more, her belief that the ice is thin might also be the fact that causes her to skate on the edge. Hornsby’s point is that both are reasons for the agent to skate at the edge of the pond given that she desires or wants to remain safe and dry. When we explain human actions we need both and we need to show how they are related. Hornsby’s claim in this paper is that in order to achieve both ends—i.e. explaining human conduct and how those different action explanations are related—it is required that we credit agents with knowledge. Let’s see if she can establish this claim.

The reason why we need subjective explanations may seem obvious to some, especially to the Humeans in action theory, like Donald Davidson, who thinks that a reason figuring in a reason-giving explanation or rationalization must “lead us to see something the agent saw” (Davidson 1965, p. 3); but it might be worth rehearsing its intuitive appeal. Hornsby resurrects Bernard Williams’ example where someone makes a mixture of petrol and tonic because he wants to drink gin and tonic and believes the petrol being gin (despite the smell...). In such a case there was no reason for making this particular mixture. After all, what the agent wanted was something of a totally different kind; what’s more, the mixture could be quite toxic and dangerous to the agent’s health (no suggestion here that gin and tonic is particularly healthy either...). But all this does not mean A had no reason to making this mixture since A still had her reasons for making the mixture as she did. The distinction that we need to make here is the difference between two existential claims, namely the fact that there is no reason for A to F is non-identical to the claim that A had no reason to F. Or in symbols:

~$p (p is a reason for A to F) ¹ ~$q(q is A’s (or her) reason for F-ing)

In other cases one’s belief, although true, might be rather silly and not even remotely connected to what one wants to achieve by acting. Hornsby calls these cases benighted agency. For instance, if I believe truly that my 30th birthday is 10.02.2009 and I take that to be a reason for me to make a first bid at 30.100.220,09 $ for a flat in Queens, there’s a sense in which I’m clearly benighted in my activity. Such examples are supposed to show the need for the subjective sense in which something is a reason for F-ing: they may be true, false, silly or just plain stupid, still they are the reasons for which A, as a matter of fact, is F-ing since the agent takes those as reasons for F-ing and acts on them as such. What happens it those cases the agent acts on her belief that P is reason for F-ing and then, from A’s perspective, there would be an objective reason to F if P were true.

To account for the objective sense in which something is a reason for A to F we only need to point to the fact that P is reason for A to F, i.e. OBJ. However, to make that reason figure in a rationalization or reason-giving explanation of A’s activity it is not enough to just cite this objective fact. After all, there might be a perfectly good reason for me not to be writing this blog post at this moment; but that does not make it a good reason to cite, as it stands, in a rationalization for why I am struggling to write it. Would a belief do the trick? That is to say, would it capture the objective sense in which something figures as a reason in a rationalization of one’s activity if the agent truly believed that P was a reason for her to F and acted on this belief? According to Hornsby, the answer is clearly no and the true belief that P is a reason to F do not add up to all we want from the rationalization. The true belief could be result of a mere happy conjecture or just the result of a lucky happenstance in which case Hornsby thinks that “inasmuch as the skater’s belief could have been false, the skater’s believing what she did can hardly provide her with the reason that there was for her to keep to the edge.” For instance, if A was told that the ice was thin by an otherwise reliable friend who for some reason was out to trick her from skating at the centre of the pond, it will be true that the ice was thin (unbeknownst to the friend’s knowledge) and A will have and act on this true belief. Yet, there is a sense in which the friend’s attempted trickery ruins the way in which we expect the agent to be connected to her reasons for acting; it requires something more than just a Gettierized, justified, true belief that P is a reason to F to be provided with that reason in the objective sense.

Of course, any interesting or empirical belief “could have been false” so I guess we should read Hornsby’s suggestion here as saying something to the effect that a mere belief “could easily have been false” in order to retain a charitable reading. By that I mean that one’s belief, although true, could have been true as a matter of epistemic luck—as shown by the Gettierized case—and that Hornsby thinks this presence of epistemic luck suffices to block the agent from being provided with the reason there was for her to skate at the edge of the pond. If that reading is correct we can begin to appreciate the intuitive connection between agency and epistemology/knowledge: we could say that something goes missing in this case—i.e. the possibly Gettierized scenario—and that “what one needs for one’s true belief to provide one with a reason for skating on the edge of the pond is that the belief be not only true but also epistemically reliable (i.e. holding true in all of one’s epistemic alternatives)”. Here the reliability relation could be defined as an ordinary accessibility relation in modal logic that is defined as function from the world one is in (@) to the possible worlds one for all one knows have been in (i.e. the set of those worlds that are consistent with all one’s evidence in @). P is the reason for A’s F-ing then (i.e. the reason because of which A’s F-ing) only if P is (a) true; (b) believed; and (c) reliably based. In the Gettier case condition (c) fails and we will have to say that A kept to the edge of the pond not because the ice was thin but because he believed (correctly) that the ice was thin. So his true belief does not provide him with an objective reason for acting because it fails to be reliably based; so adding the true belief to the objective reason merely gives you another subjective sense in which P is a reason to F.

Now if A were to know that the ice was thin and thereby acting on her knowledge she would satisfy the reliability condition—after all, knowledge requires being reliable—and thus there would be no obvious reason to deny that the fact that the ice is thin now provides A with an objective reason for skating on the edge. So the presence of knowledge is enough to provide A with an objective reason for which she acted. But Hornsby makes the further claim that knowledge is also necessary or that a condition for F-ing for the reason that p is that one knows that p. Her Gettier case obviously does not establish that knowledge is necessary; reliability, for all that’s been said and done so far, could possibly be supplied for by other means. Yet knowledge is a plausible candidate and also one that frequently occurs in normative and rational evaluations of people’s activities (for evidence see: Stanley and Williamson; Stanley and Hawthorne). We might also think that the necessity requirement could be established via lottery considerations, i.e. cases where the requirement of justification needed for acting with an objective reason is pressed increasingly towards probability 1 (= knowledge); but I won’t go in this now. Suffice it to say that more is needed—and can probably be provided—in order to support Hornsby’s main claim that: “We act for reasons in virtue of our having knowledge of relevant facts. As agents, we rely upon our often being, so to speak, the conduits of facts.”

We should note, in passing, that Hornsby’s suggestion does not preclude that p could still be a reason for A to F even though he is unaware of it or fails to know it and merely believes it. The point is rather that as soon as he acts his F-ing can only count as F-ing for the objective reason that p if A also knows that p. Failing that he would merely be F-ing for his subjective reason in accordance with his objective reason, i.e. by acting for the correct belief that p was a reason for F-ing. So the fact that p is a reason for A to F can only occur as an explanation of A’s actual F-ing if A knows that p is a reason for A to be F-ing. In this sense, Hornsby’s suggestion is completely on a par with Davidson’s requirement that a reason can only figure in a rationalization of someone’s behaviour if it shows or “leads us to see something the agent saw” (1965, p. 3); when we explain that A knows that P is a reason for A to F just is a way to come to understand something the agent saw. Davidson also lists knowledge as one of the possible cognitive attitudes we can list and combine with a pro attitude (desires, wants, etc) to yield the primary reason that rationalizes A’s intentional behaviour.

Anyway, Hornsby goes on to suggest that (E) captures what she thinks established by her Gettier case:

(E) Where ‘x F-d because p’ gives a reason-explanation (x F-d because p iff x F-d because x knew that p).

I think I see a problem with (E): it seems to be a version of the KK principle and thus leads one to the absurd consequence that follows when one applies an S4 model for the accessibility relation that operates on the epistemic operator. That is to say, Hornsby’s suggestion can easily be shown to require much more reliability and knowledge than first assumed. Here’s why:

Read F(x, p) as ‘x F-d because p’
Read K(x,p) as ‘x knew that p’

Then according to my reading of (E) as an instance of what is troublesome in the KK principle it would follow from F(x, p) that F(x, K(x,p)): F(x, Kx(K(x,p))), and so on. In short, it follows that whenever one acts for the objective reason that p one would have to not only Kp but KKp, and KKKp, etc.. The reason why this is a problem is that, according to Hornsby, the presence of knowledge adds reliability and thus it adds a restriction on one’s epistemic possibilities: the space of epistemic possibilities shrinks with every addition or iteration of knowledge. Thus, the extent to which (E) can be shown to iterate knowledge requirements is also the extent to which one would require a higher epistemic standard whenever one acts for the objective reason that P. My allegation is therefore that (E) commits one to an impossibly strict epistemological standard in order to act for objective reasons.

The best way to deal with this objection would be if one could point to the scope of the principle since the principle’s application is supposed to be guarded by a qualification to apply only where ‘x F-d because p’ gives a reason-explanation. One could hope that this qualification could be enough to block the reiteration that would make (E) an instance of the fallacious KK principle; however, (E) is easily turned into an instance of the KK principle by Hornsby’s own words since she regards both the left-hand side and the right-hand side of the equivalence as a reason-explanation. Thus, whenever ‘A F-d because p’ is a reason-explanation it follows from the equivalence in (E) that ‘A F-d because A knew that p’ is a reason-explanation too. Since that is the case nothing stops us from reapplying (E) to ‘A F-d because A knew that p’ since (E) is a universal principle that is supposed to apply whenever something of the form ‘x F-d because p’ gives a reason-explanations: and, guess what, ‘A F-d because A knew that p’ has that form since ‘A knew that p’ is a fact too. It’s just that on the reapplication the variable ‘p’ in the schema is replaced by ‘A knew that p’ rather than ‘p’; so when we put that into the formula (E) it gives you back that ‘A F-d because A knew that A knew that p’. This process can now be repeated as many times as you want giving you an infinite number of iterations of knowledge as a general requirement for objective reason-explanations. In other words, to have reason-explanation for F-ing of the form ‘x F-d because p’ requires not only that one knew p but also that one knew that one knew p, knew that one knew that one knew p, and so ad infinitum. Again, the trouble is that knowledge adds further reliability thus restricting the epistemic possibilities: in short, one would need to know so much that it rules out any epistemic possibilities and thus one would need to know exactly which world one resides in order to act because that P. And no one, except for an omniscient being, knows that much which is an intolerable consequence of Hornsby’s suggestion. At least, so it seems to me.

A possible solution is to insist on two senses of because here: we could read ‘A F-d because that P’ as the sense in which something is objectively a reason for F-ing without being the ‘because’ that figures in a reason-giving explanation; and think of ‘A F-d because* she knew P was a reason for her to F’ as the ‘because’ that does figure in a reason-giving explanation. But that runs counter to the qualification or what we seek to explain, namely what we need for the citing objective sense of being a reason to F as a reason figuring in the reason-giving explanation for F-ing. As far as I can see, Hornsby is in real trouble here and I see no easy way out of it. (Note: it won’t help replacing the biconditional with a conditional, either; since the consequences only hinges on one direction of the biconditional).

Bracket this problem and I’m in sympathy with Hornsby’s general idea that, as she says, adding knowledge to the soup is an elegant way in which something can be a reason for F-ing at the same time as one can account for the agent’s motivation. Knowledge is factive and therefore Kp entails p; also knowledge is arguably a cognitive or mental state and thus can figure as a causal factor in the explanation of one’s acts. We still need a story about the subjective sense in which something is a reason for F-ing; or rather, we need to connect such a story to the one about knowledge. That is Hornsby’s topic for section 2 of this paper but I will return to that project in a blog post that is soon to follow.

Davidson, D, 1963, ‘Actions, Reasons, and Causes’
Gibbons, J. 2001, ‘Knowledge in Action’
Hornsby, J. 2007?, ‘Knowledge, Belief and Reasons for Acting’
Williamson, T. 2000, Knowledge and Its Limits


  1. Note that the symbols I used in the original text fell out on this blog post: however, '$' is just the existential quantifier whereas F-ing is just a replacement for the customary greek letter Phi. All else should be in place.

  2. Hey Heine,
    I'm planning to attend the Action Theory reading group, beginning this Friday. Finally got around to reading the Hornsby paper yesterday and I've posted a short response to your summary over at The Space of Reasons.

    Hope to see you on Friday!