In the latest _Philosophical Studies_, Jonathan Schaffer launches a series of objections to Interest Relative Invariantism. I suspect most of these will end up being clashes of intuitions, though maybe I’ll write something more about them later. What I want to focus on here are Jonathan’s arguments concerning knowledge claims involving embedded questions. Jonathan claims these support contextualism, but I don’t really understand how that could be true. What he does seem to raise is a problem about how to understand embedded questions.

Here is a small variant of a case that Jonathan uses.

bq. CIA Headquarters have been burgled. Frank is the thief, though the CIA don’t know that. Indeed, Frank isn’t even on their list of suspects. After an investigation to find out what was stolen, the CIA have discovered that the only thing stolen is the director’s coffee cup.

A and B are having a conversation about the break-in. It is common ground among them that Frank was the thief, and the only thing he stole was the coffee cup. Further, it is common ground that the CIA were doing an investigation into what was stolen. But only A knows that the investigation has completed with the CIA discovering that only the coffee cup was stolen. First consider A making any one of the following claims.

(1) The CIA know that Frank stole the coffee cup.
(2) The CIA know what Frank stole.
(3) The CIA found out what Frank stole.
(4) The CIA discovered what Frank stole.
(5) The CIA were investigating what Frank stole.
(6) The CIA know who stole the coffee cup.

My instinct is that (1) is clearly false, as is (6) but it’s plausible that A could truly utter any of (2) through (5). I don’t think these four are clearly true, but I think they have a true reading. In a context where Frank’s identity as the thief is given, and its importance to the CIA not being considered (because we are talking for now about the investigation into what was stolen) we can bracket the fact that the CIA don’t know Frank is the thief. In other contexts I think it would be natural to say (2) is false.

Similarly, consider B asking one of the following two questions.

(7) Does the CIA know that Frank stole the coffee cup?
(8) Does the CIA know what Frank stole?
(9) Does the CIA know who stole the coffee cup?

The answer to (7) and (9) is clearly no, but the answer to (8) may well be yes. (At least in circumstances where A could truly say (2).) On the other hand, if A was asked (8) by Frank’s lawyer, or his brother, who is worried about Frank’s legal jeopordy, the proper answer may well be no. As we suggested above, if A was talking to these people, it is odd, and arguably false, for her to utter any of (2) through (5).

Jonathan suggests all of this supports some kind of contextualism, and it may well, but it isn’t contextualism about ‘knows’. I think it’s true that (2) to (5) are context-sensitive, but I don’t think that’s because of the verb ‘knows’. If it were, we could generate some context-sensitivity in (1), and we can’t. The real context-sensitivity is in the embedded question. Of course, all context-sensitivity in the sentence shows is that _some_ term in the sentence is context-sensitive. And as I think Jonathan himself shows, the semantic value of the embedded question is context-sensitive.

Having said that, it’s tempting to make the following argument.

# If what Frank stole is the coffee cup, then the CIA knows what Frank stole iff the CIA knows that Frank stole the coffee cup.
# Whether ‘The CIA knows what Frank stole’ is true is context-sensitive.
# The coffee cup is what Frank stole.
# ‘The CIA knows what Frank stole’ is true iff the CIA knows what Frank stole.
# ‘The CIA knows that Frank stole the coffee up’ is true iff the CIA knows that Frank stole the coffee cup.
# So, whether ‘The CIA knows that Frank stole the coffee cup’ is context-sensitive.

That’s a bit rough, but hopefully you get the idea. It’s a very tempting argument, because every premise has at least a good ring of plausibility. But the conclusion is false. And the above example shows why; the very first premise fails in our example. I suspect in any argument for contextualism about knowledge that claims from contextualism about knowledge wh- claims, the equivalent premise linking the two types of knowledge claims will be false. Certainly Jonathan needs a premise like 1 for his argument to work, but as far as I can see he doesn’t defend one, and his very arguments provide reason to think that the premise would be false.

But this all raises a further question. If the CIA don’t need to know that Frank stole the coffee cup in order to know what Frank stole, what do they need to know? My best guess is that the following is true.

bq. ‘S knows what N F-ed’, where F is a transitive predicate, iff for some x, x is what N F-ed, and there is some description ‘the D’ denoting N such that S knows that the D F-ed x.

That’s a mouthful, but what it comes to in our case is that the CIA must know that Frank, under some description or other, stole the coffee cup, though they need not know that the description denotes Frank. And they do know that – they know that the thief stole the coffee cup, and ‘the thief’ denotes Frank. Obviously this would need a bit of work in order to get it to a general theory of embedded questions, but I think this is on the right track.

If it is on the right track, it explains where the contextualism comes from. The quantifier over descriptions in the paraphrase above is, like all quantifiers, potentially restricted by contextually supplied parameters. Which kinds of descriptions of N are allowed will be determined by context. When talking to someone who cares particularly about what the CIA knows about Frank, perhaps the only description that matters is ‘is identical with Frank’. In other contexts, descriptions like ‘the thief’ matter as well.

So the big epistemological point is that these cases provide no support for contextualism about knowledge that claims. But the semantic question, about just what the semantic value of embedded questions is seems hard and interesting. I certainly don’t know that my solution here works, but I hope it is on the right track.