What
I propose is that similarity sometimes plays no part at all, and that
whether it does also varies with sentence and context. When it plays
no part, the truth of the counterfactual in question requires that
all
A-worlds are C-worlds. (To state the modified theory elegantly, we
could speak of 'all relevant A-worlds', defining 'relevant' using
'most similar' but adding that sometimes all A-worlds will be
relevant.)
The
argument for this modification involves what could be called
categorical counterfactuals. Consider the following sentence,
uttered in the context of teaching someone how to use the word
'bachelor':
(A)
If I had spoken to a bachelor this morning, I would have spoken to an
unmarried man this morning.
Intuitively,
the truth of this hinges on the fact that bachelors are necessarily
unmarried men. Lewis's analysis, without my proposed
modification, although it gives the right truth-value, gives the
wrong truth-condition and thus distorts the meaning of (A); it
is false to say that the truth-condition for this sentence is that
all the most similar A-worlds are C-worlds - on any understanding
of similarity.
The
modified theory handles (A) much better: this is one of those cases
where similarity plays no part, and so (A) is true iff all
worlds where I spoke to a bachelor this morning are worlds where I
spoke to an unmarried man this morning. This seems right.
(A
note on the structure of Lewis's theory as formally developed with
systems of spheres: this can remain as is, but in the case of
categorical counterfactual conditionals the “innermost” sphere
will contain all worlds, and so it would be misleading to call the
worlds in this sphere 'the most similar A-worlds'.)
Reference
Lewis,
D. 1973. Counterfactuals. Basil Blackwell: Oxford.
