Action-guiding counterfactuals

Suppose Alice is essentially a non-liar and essentially knows all about human affairs as well as about her essential properties, but she does not have any significant powers to affect human affairs except by answering questions. You ask Alice whether there is poverty among humans.

She thinks to herself that since she essentially knows all about human affairs and is incapable of lying, therefore this is true for her:

1. Were I to say “There is no poverty among humans”, there would be no poverty among humans.

Since it’s a lot better that there be no poverty, she says there is no poverty among humans.

That’s absurd. Yet (1) seems to follow from the following plausible premises:

2. If p entails q, and it is contingent whether p holds, then were p to hold, q would hold.

3. It is contingent whether there is poverty among humans.

For that Alice says that there is no poverty among humans entails that there is no poverty among humans, since she is essentially incapable of lying and essentially knows whether there is poverty among humans.

It seems that Alice should deny (1) in favor of:

4. Were I to say “There is no poverty among humans”, I’d be lying.

But that’s very odd, because it’s a counterfactual with a contingent antecedent (Alice can say “There is no poverty among humans”: she says it in possible worlds where there is no poverty among humans) and an impossible consequent. And it seems like all such counterfactuals should be false.

What’s going on? Here’s a suggestion. Counterfactuals are highly context-dependent in what one keeps fixed. Richard Gale once illustrated this point with the dark joke: “What would Queen Victoria be doing if she were alive today? Clawing at the inside of her coffin!” In the context of action-guidance, we need to keep fixed the true “dependency hypotheses” (familiar from causal decision theory). Present facts are among the dependency hypotheses to keep fixed when deliberating, except in special cases like where you are deliberating about how to use a time machine. Thus, we keep fixed that there is poverty, and (4) is correct while (1) is false when said in an action-guiding context.

If we say that the dependency hypotheses count as part of the antecedent, we can keep a version of (2), though I don’t know that that’s exactly the right way to do the semantics.

But things are a bit more complicated. Essential facts about one’s traits are surely also among the dependency hypotheses. But now the true dependency hypotheses include:

1. There is poverty

2. Alice knows whether there is poverty

3. Alice doesn’t lie.

What would be true if we were to combine (i)–(iii) with Alice saying “There is no poverty”? I have no idea! Poof, logic would explode, and everything would be true? Or would there be something more specific true? I don’t know. In any case, it is unclear that (4) is true and (1) isn’t when we think about it this way.

Here is a thought. There is a hierarchy among the dependency hypotheses. Facts about the agent’s character—even necessary facts—are lower down in the hierarchy than other dependency hypotheses. In cases of conflict, we keep fixed the higher up dependency hypotheses at the expense of the lower down ones. Thus, we keep fixed (i) at the expense of (ii) and (iii). Maybe that helps. But it seems rather ad hoc.

And consider this puzzle. You ask Alice whether she can lie. Now the relevant dependency hypotheses are just that Alice knows her essential traits and that she is essentially a non-liar, and these seem all on par. So if we had these dependency hypotheses along with Alice saying she can lie, there is no telling what would eventuate. And in particular it is unclear how Alice can reason to the conclusion that she should say “I can’t lie” rather than “I can lie.”