The Semantic Paradoxes and the Paradoxes of

Vagueness
Hartry Field∗
March 30, 2003

Both in dealing with the semantic paradoxes and in dealing with vagueness
and indeterminacy, there is some temptation to weaken classical logic: in par-
ticular, to restrict the law of excluded middle. The reasons for doing this are
somewhat different in the two cases. In the case of the semantic paradoxes,
a weakening of classical logic (presumably involving a restriction of excluded
middle) is required if we are to preserve the naive theory of truth without in-
consistency. In the case of vagueness and indeterminacy, there is no worry about
inconsistency; but a central intuition is that we must reject the factual status
of certain sentences, and it hard to see how we can do that while claiming that
the law of excluded middle applies to those sentences. So despite the different
routes, we have a similar conclusion in the two cases.
There is also some temptation to connect up the two cases, by viewing the
semantic paradoxes as due to something akin to vagueness or indeterminacy in
semantic concepts like ‘true’. The thought is that the notion of truth is intro-
duced by a schema that might initially appear to settle its extension uniquely:
the schema
(T) True (hpi) if and only if p,
(where ‘p’ is to be replaced by a sentence and ‘hpi’ by a structural-descriptive
name of that sentence). But in fact, this schema settles the extension uniquely
only as applied to "grounded" sentences; whether a given "ungrounded" sen-
tence is in the extension of ‘true’ will be either underdetermined or overdeter-
mined (determined in contrary ways). And this looks rather like what happens
in cases of vagueness and indeterminacy: our practices with a vague term like
‘bald’, or a term like ‘heaviness’ (which in the mouths of many is indeterminate
between standing for mass and standing for weight), don’t appear to uniquely
settle the reference or extension of the term. (There isn’t such a clear distinction
between underdetermination and overdetermination in these cases: e.g., it may
seem underdetermined whether Harry is bald in that Harry isn’t a paradigm
case of either baldness or non-baldness, but it may seem overdetermined in that
Sorites reasoning may be used to argue that he is bald and also to argue that
he is not bald. Similarly, a theorist who makes no distinction between mass and
weight hasn’t decided which one ‘heaviness’ stands for, so it may seem underde-
termined; but he may attribute to "heaviness" both features true only of mass
and features true only of weight, so it may seem overdetermined.)
∗ New York University. Email: [email protected]

1
 In this paper I will argue that an adequate treatment of each of the two
phenomena (the semantic paradoxes and vagueness/indeterminacy) requires a
nonclassical logic with certain features–features that are roughly the same for
each of the two phenomena. This suggests that there might be a common
logical framework, and I will propose a framework that seems adequate for
treating both. The core logic for the uniÞed framework is sketched in Section
5 (which is rather technical), though much of the discussion before and after
(e.g. the discussion of rejection in Section 3 and the discussion of defectiveness
in later sections) is highly relevant to the uniÞed treatment. Sections 1, 2, 7
and 8 are primarily concerned with the semantic paradoxes, and Sections 3, 4,
6 and 9 deal primarily with vagueness and indeterminacy (especially in the case
of 3 and 4), but there are close interconnections throughout.

1 The semantic paradoxes and attempts to re-

solve them in classical logic
In this section and the next I will discuss the naive theory of truth and some
of the semantic paradoxes which threaten to undermine it. As I’ve noted, the
naive theory of truth includes all instances of the schema (T) above. Perhaps
more centrally, it includes the principle that ‘True(hpi)’ and ‘p’ are always inter-
substitutable. (Of course I’m restricting to languages without quotation marks,
intentional contexts, and so forth; also without ambiguity, and where there are
no relevant shifts of context. Although it may be unnecessary, we can exclude
denotationless terms as well.) In classical logic, the schema implies the inter-
substitutivity and conversely; whether either direction of the implication holds
in a non-classical logic depends on the details of that logic, though I think that
there is reason to prefer logics that are "classical enough" for both directions of
the implication to hold.
In the context of a very minimal syntactic theory that allows for self-reference,
the naive theory of truth is inconsistent in classical logic: we can construct a sen-
tence Q0 ("the Liar") that is provably equivalent to ¬True(hQ0 i), so in classical
logic we can derive the negation of an instance of (T). It is totally unpromising
to blame the problem on the syntactic theory: among other reasons, there are
very similar paradoxes in the naive theory of satisfaction which don’t require
syntactic premises. The real choice is, do we restrict classical logic or restrict
the truth schema (and hence its classical equivalent, the intersubstitutivity of
‘True(hpi)’ with ‘p’).
The attractions of keeping classical logic sacrosanct are powerful, so let’s
look Þrst at the prospects of a satisfactory weakening of the naive theory of
truth within classical logic. By a satisfactory weakening, I mean one that serves
the purposes that the notion of truth is supposed to serve, e.g. as a device of
making and using generalizations that would be difficult or impossible to make
without it. I think it is pretty clear that any weakening of the naive theory
of truth will adversely affect the ability of ‘True’ to serve these purposes: the
intersubstitutivity of ‘True(hpi)’ with ‘p’ is very central to the purposes the
truth predicate serves. Still, if there were a sufficiently powerful but consistent
classical-logic substitute for the naive theory, we might be able to learn to live
with it.

2
 In fact, though, I think that restoring consistency requires massive revisions
in ordinary principles about truth, revisions that would be very hard to live
with. I won’t try to fully establish this here, but will make a few observations
that give some evidence for it. To begin with an obvious point, the problem in
classical logic isn’t simply that we can’t assert all instances of schema (T), it is
that there are instances (such as the one involving Q0 ) that we can disprove;
and in classical logic, that’s equivalent to proving the disjunction of

(1) p ∧ ¬True(hpi)

and

(2) True(hpi) ∧ ¬p

for certain speciÞc p. Now, it would seem manifestly unsatisfactory to have a

theory that proves an instance of (1): how can we assert p and then in the same
breath assert that what we’ve just asserted isn’t true? But it also seems mani-
festly unsatisfactory to have a theory that proves an instance of (2): once we’ve
asserted True(hpi), we’re surely licensed to conclude p, so going on to assert ¬p
just seems inconsistent. We seem to have, then, that it would be manifestly
absurd to have a theory that either proves an instance of (1) or proves an in-
stance of (2). Of course, a classical theory doesn’t have to do either: it must
prove the disjunction (given that it meets the minimal requirements on allowing
self-reference)1 , but it can remain silent on which disjunct to assert. But re-
maining silent doesn’t seem a satisfactory way to resolve a problem: if you have
committed yourself to a disjunction of thoroughly unsatisfactory alternatives,
it would seem you’re already in trouble, even if you refuse to settle on which of
these unsatisfactory alternatives to embrace.2
I will not further discuss the option of biting the bullet in favor of (2), or
the option of accepting the disjunction of (1) and (2) while remaining artfully
silent about which disjunct to accept.3 But I’ll say a bit more about the option
of biting the bullet in favor of (1).
A superÞcially appealing way to bite the bullet for option (1) is to say that
schema (T) should be weakened to the following:

(T ! ) If True(hpi) or True(h¬pi), then True(hpi) if and only if p.

(Proponents of this often introduce the term ‘expresses a proposition’, and say
that hpi’s expressing a proposition suffices for the consequent of (T ! ) to hold
and that the antecedent of (T ! ) suffices for hpi to express a proposition.) It is
easily seen that (T ! ) is equivalent to the simpler schema
1 As mentioned above, we wouldn’t need even these minimal requirements if we focused on

satisfaction rather than truth, and used the heterologicality paradox.

2 Note that the situation is far worse than for supervaluationist accounts of vagueness. Such

accounts allow commitment to disjunctions where we think it would be a mistake to commit

to either disjunct. But there the only problem with choosing one disjunct over the other is
that the choice seems quite arbitrary; the disjuncts are not thoroughly unacceptable, as they
seem to be in the case of the paradoxes.
3 My own dissatisfaction with the "artful silence" option is not entirely due to the general

consideration just raised, but also to the fact that a consistent view of this type must exclude
so many natural principles. See [11] and [18] for some important limitations on such theories.

3
 (T !! ) If True(hpi) then p.4

Obviously these equivalent schemas can’t be anything like complete theories of

truth: for they are compatible with nothing being true, or with no sentence that
begins with the letter ‘B’ being true, or innumerably many similar absurdities.
To get a satisfactory theory of truth that included them, one would have to add
a substantial body of partial converses of (T !! ); and one would presumably also
want principles such as

(TPMP) True(hpi) ∧ True(hp ⊃ qi) ⊃ True(hqi),

or better, the generalized form of this (that whenever a conditional and its
antecedent are true, so is the consequent: that is, that modus ponens is truth-
preserving).5 But whatever the details of the supplementation, theories based
on (T ! ) or its equivalent (T !! ) are prima facie unappealing because they re-
quire a great many instances of (1). Obviously the Liar sentence Q0 is one
example: since we can prove that Q0 ≡ ¬True(hQ0 i), (T !! ) yields both Q0
and ¬True(hQ0 i). But in addition, Montague [19] pointed out that (T !! ) plus
(TPMP) plus the very minimal assumption that all theorems of quantiÞcation
theory are true yields a proof of the untruth of some instances of (T !! ): that is,
there is a sentence M (a slight variant of Q0 ) such that we can prove

¬True(h If True(hMi) then M i).6

It seems highly unsatisfactory to put forward a theory of truth that includes

(T !! ), and use it to conclude that some instances of (T !! ) (including speciÞc
instances that you can identify) aren’t true.
4 (T ! )implies (T !! ): Suppose True(hpi); then by (T ! ), True(hpi) ≡ p, which with True(hpi)
yields p; so we have True(hpi) ⊃ p. (T !! ) implies (T ! ): This requires two instances of (T !! ),
both (i) True(hpi) ⊃ p and (ii) True(h¬pi) ⊃ ¬p. Suppose True(hpi); then by (i), p; so
True(hpi) ≡ p. Alternatively, suppose True(h¬pi); then by (ii), ¬p, and by (i), ¬True(hpi); so
again True(hpi) ≡ p. So True(hpi) ∨ True(h¬pi) ⊃ (True(hpi) ≡ p).
5 Equivalently (given a minimal assumption about how elimination double-negation leaves

truth unaffected),
True(hp ∨ qi) ∧ False(hpi) ⊃ True(hqi),
or the generalized form of that; where ‘False(hpi)’ means ‘True(h¬pi)’.
6 Let R be the conjunction of the axioms of Robinson arithmetic (which is adequate to con-

struct self-referential sentences). Standard techniques of self-reference allow the construction

of a sentence N that is provably equivalent, in Robinson arithmetic, to True(hR ⊃ ¬Ni); M will
be R ⊃ ¬N . Since this is provable in Robinson arithmetic, then R ⊃ [N ⊃ True(hR ⊃ ¬Ni)]
is a theorem of quantiÞcation theory, hence so is its quantiÞcational consequence
[True(hR ⊃ ¬N i) ⊃ (R ⊃ ¬N)] ⊃ (R ⊃ ¬N);
so the claim that that is True is part of the truth theory, and that together with (TPMP)
yields
True[hTrue(hR ⊃ ¬Ni) ⊃ (R ⊃ ¬N)i] ⊃ True(hR ⊃ ¬Ni).
So if our truth theory proves the negation of the consequent, it proves the negation of the
antecedent, which is the desired negation of the attribution of truth to an instance of (T !! ).
It remains only to show that a proof theory with (T !! ) and arithmetic does prove
¬True(hR ⊃ ¬Ni), but that’s easy: from (T !! ) we get True(hR ⊃ ¬N i) ⊃ (R ⊃ ¬N),
which given arithmetic yields True(hR ⊃ ¬Ni) ⊃ ¬N; but since N is provably equivalent to
True(hR ⊃ ¬N i), this yields ¬True(hR ⊃ ¬Ni).

4
 It is sometimes thought that one can improve this situation by postulat-
ing a hierarchy of ever more inclusive truth predicates Trueσ , and for each one
adopting (T !! σ ), i.e. the analog of (T !! ) but with ‘Trueσ ’ in place of ‘True’.
(The subscripts are notations for ordinals; the idea is that there will be truth
predicates for each member of an initial segment of the ordinals, with no largest
σ for which there is a truth predicate. There is no notion of truthσ for variable
σ.) For any ordinal σ for which we have such a predicate, we will be able to de-
rive ¬Trueσ (h If Trueσ (hMσ i) then Mσ i) for a certain sentence Mσ that contains
‘Trueσ ’; but we will also be able to assert Trueσ+1 (h If Trueσ (hMσ i) then Mσ i),
and this is sometimes thought to ameliorate the situation.
But even if it does, the cost is high. I can’t discuss this fully, but will con-
Þne myself to a single example. Suppose I tentatively put forward a "theory
of truth"–more accurately, a theory of the various truthσ s–that includes all
instances of (T !! σ ) for each of the truth predicates, together with general prin-
ciples such as (T P MPσ ) for each of the truth predicates, and various partial
consequences of each of the (T !! σ ). Someone then tells me that my theory has
an implausible consequence; I can’t quite follow all the details of his compli-
cated reasoning, but he’s a very competent logician and the general strategy he
describes for deducing the implausible consequence seems as if it should work,
so I come to think he’s probably right. Since the consequence still seems im-
plausible, it is natural to conjecture that my theory of truth is wrong–or at
least, to consider the possibility that it is wrong and discuss the consequences
of that. It is natural to do this even if I have no idea where it might be wrong.
But I can’t conjecture this, or discuss the consequences of it, since I have no
sufficiently inclusive truth predicate. (‘Wrong’ means ‘not true’.)7
And a more speciÞc conjecture, that my theory isn’t trueσ for some speciÞc
σ, won’t do the trick. For one thing, I already know for each of my truth
predicates trueσ that not all of the assertions of my theory are trueσ ; after all,
it was because I knew that certain instances of (T !! σ ) couldn’t be trueσ that I
was led to introduce the notion of truthσ+1 . Might I get around that problem
by Þnding a way of specifying for each sentence A of my theory of truth a σ A
such that A will be “trueσA if it is true at all” (if you’ll pardon the use of an
unsubscripted truth predicate)? I doubt that one can Þnd a way to specify
such a σ A for each A: the fact that many principles of a decent truth theory
contain quantiÞers that range over arbitrary sentences and hence sentences that
include arbitrarily high truthσ predicates gives serious reasons for doubt. But
even if one can do that–indeed, even if one can specify a function f mapping
each sentence A of the theory into the corresponding σ A –it wouldn’t fully get
around the problem. For it could well be that for each σ, I would be conÞdent
that all members of {A|f (A) = σ} are trueσ ; a doubt that there is a σ such
that not all members of {A|f (A) = σ} are trueσ does not entail a doubt for
any speciÞc σ. It is the more general doubt, that there is a σ such that not
all members of {A|f (A) = σ} are trueσ , that my story motivates; but that
more general doubt is unintelligible according to the hierarchical theory, for in
treating quantiÞcation over the ordinal subscripts as intelligible it violates the
principles of the hierarchy.
7 If the theory were Þnitely axiomatized I could avoid the use of a truth predicate, but it
isn’t: that’s guaranteed by the need of a separate instance of (T !!σ ) for arbitrarily high σ (or
rather, for arbitrarily high σ such that ‘trueσ ’ is deÞned).

5
 There is much more that could be said about these matters, but I hope I’ve
said enough to make it attractive to explore an option that weakens classical
logic.

2 Semantic paradoxes: a non-classical approach

The most famous non-classical resolution of the paradoxes, due to Kripke [14],
employs a logic K3 that can be read off the strong Kleene truth tables.
More exactly, suppose we assign each sentence A a semantic value ||A| | of
either 1, 0, or 12 , with the assignment governed by the following rules:
||A ∧ B| | is min{||A| |, ||B| |}
||A ∨ B| | is max{||A| |, ||B| |}
||¬A| | is 1 − ||A| |
||A ⊃ B| | is ||¬A ∨ B| |, hence max{1 − ||A| |, ||B| |}.
(Think of 1 as the "best" value, 0 as the "worst", and 12 as "intermediate". It
may seem more philosophically natural to avoid assigning the value 12 , and to
instead regard certain sentences as simply having no value assigned to them; but
obviously these two styles of formulation are intertranslatable, and the formu-
lation that uses the value 12 allows for a more compact presentation at several
points.) Assuming everything to have a name, as I will for simplicity,8 we also
set ||∀xA| | = min{||A(x/c)| |} and ||∃ xA| | = max{||A(x/c)| |}. Then Kripke
shows that if we start with the language of arithmetic or some other language
adequate to syntax, and any arithmetically standard model M for it that is
evaluated by these rules, then we can extend the model by designating a subset
of M as the extension of ‘True’ (leaving the ontology and the extension of the
other predicates alone) in such a way that for every sentence A, ||True(hAi)| |
will be the same as ||A| | (and where only objects that satisfy ‘Sentence’ satisfy
‘True’.) More generally, if sentences B and C are alike except that some occur-
rences of A in one of them are replaced by True(hAi) in the other, then ||B| |
will be the same as ||C| |.
One way to look at this is as showing that we can keep the intersubstitutivity
of True(hAi) with A in a revised logic K3 . In K3 we call an inference valid if
under every assignment to atomic sentences, if the premises have semantic value
1 then so does the conclusion. And we call a statement valid if it has semantic
value 1 under every assignment to atomic sentences. Note that instances of
the law of excluded middle (A ∨ ¬A) comes out invalid: they can have value
1
2 . (Indeed, no statement in this language is valid, though many of the familiar
classical rules are valid.) Kripke’s result shows that in the logic so obtained,9 we
can consistently assume that for every sentence A and every pair of sentences
B and C that are alike except that some occurrences of A in one of them are
replaced by True(hAi) in the other, the inference from B to C and from C to B
are valid. This is one of the two components of the naive theory of truth, and
is not consistently obtainable in classical logic.
8 Alternatively, we could extend the assignment of semantic values to pairs of formulas and

functions assigning objects to variables.

9 Indeed, even in a slightly expanded logic K + that includes disjunction elimination as a
3
meta-rule; this rule becomes relevant when one considers adding new validities involving new
vocabulary.

6
 It is not entirely clear that this use of nonclassical logic is what Kripke
is recommending in his discussion of the strong Kleene version of his theory
of truth: some of his remarks suggest it, but others suggest a classical-logic
theory later formalized by Feferman ([3], pp. 273-4). The classical-logic Kripke-
Feferman theory postulates truth-value gaps: it says of certain sentences, such
as the Liar, that they are neither true nor false. (‘False’ is taken to mean ‘has a
true negation’, so the claim is that neither they nor their negations are true.) As
a consistent classical theory, it gives up on the equivalence between ‘True(hpi)’
and ‘p’. (The Kripke-Feferman theory is one of the ones that commits itself
to disjunct (1) in the previous section.) The Kripke theory in its non-classical
version has no commitment to truth-value gaps: indeed, since the whole point
of the theory is to maintain the equivalence between ‘True(hpi)’ and ‘p’, the
assertion of ¬[True(hpi) ∨ ¬(True(h¬pi)] would be equivalent to the assertion of
¬[p ∨ ¬p]; that entails both p and ¬p in the logic, and in this logic as well as in
classical that entails everything. So it is very important in the Kripke theory
(on its non-classical reading) not to commit to truth-value gaps. It will give
certain sentences the value 12 , but that is not to be read as "neither true nor
false".
I think the non-classical reading of Kripke is the more interesting one, and
I will conÞne my discussion to it. I think there are two main problems with the
Kripke theory (on this reading).
Perhaps the more serious of the problems is that the logic is simply too
weak: as Feferman once remarked, "nothing like sustained ordinary reasoning
can be carried out in [the] logic" ([3], p. 264). One symptom of this is that not
even the law A ⊃ A is valid: since A ⊃ B is equivalent to ¬A ∨ B, this follows
from the invalidity of excluded middle. And note that the intersubstitutivity
of True(hAi) with A guarantees that True(hAi) ⊃ A and its converse are each
equivalent to A ⊃ A; since the latter isn’t part of the logic in the Kripke theory,
neither half of the biconditional True(hAi) ≡ A is validated in Kripke models.
So one consequence of the Þrst problem for the Kripke theory is that it does not
yield the full naive theory of truth.
The other problem for the Kripke theory, the "revenge problem", has been
more widely discussed, but I think much of that discussion has been vitiated by
a confusion between the non-classical version of Kripke’s theory and the classical
Kripke-Feferman theory: much of it has been based on falsely supposing that
the Kripke theory is committed to truth-value gaps. The only real revenge
problem for the non-classical Kripke theory has to do with the fact that the
"defectiveness" of sentences like Q0 is inexpressible in the theory, and there is a
worry that if we were to expand the theory to include a "defectiveness predicate"
the paradoxes would return. I will be proposing a theory that has much more
expressive power than the Kripke theory, and which avoids the revenge problem
by having the means to express the defectiveness of paradoxical sentences like
Q0 without this leading to inconsistency.
Returning to the Þrst of the two problems, a natural idea for how to avoid
it is to add a new conditional → to the Kleene logic, which does obey the law
A → A. There have been many proposals about how to do this; unfortunately,
most of them do not enable one to consistently maintain the intersubstitutivity
of True(hAi) with A (or even the truth schema True(hAi) ↔ A which that implies

7
given the law A → A). In fact, I know of only two workable proposals for how
to do this, both by myself; and one of them ([6]) is not very attractive. (There is
also a proposal in Brady [1], which is not an extension of Kleene logic but only
of a weaker logic FDE, which is a basic relevance logic.) These theories not only
contain A → A, they also contain a substitutivity rule that allows the inference
from A ↔ B to C ↔ D when C and D are alike except that one contains B in
some places where the other contains A; thus the logic is "classical enough" for
the two components of the classical theory of truth to be equivalent.
I will say a little bit about the more attractive of the two theories ([9]).
As with Kripke’s construction, we start out with a base language that doesn’t
include ‘True’, or the new ‘→’, and with a classical model for this base language
whose arithmetical part is standard. The semantics of the theory–which I’ll
call the Restricted Semantics, since I will generalize it in Section 5–is given by a
transÞnite sequence of Kripke-constructions. At each stage of the transÞnite se-
quence ("maxi-stage"), we begin with a certain assignment of values in {0, 12 , 1}
to sentences whose main connective is the ‘→’. Given such an assignment of
values to the conditionals, Kripke’s method of obtaining a minimal Þxed point
enables us (in a sequence of "mini-stages within the maxi-stage") to obtain a
value for every sentence of the language, in such a way as to respect the Kleene
valuation rules and the principle that True(hAi) always has the same value as A.
It remains only to say how the assignment of values to conditionals that starts
each maxi-stage is determined. At the 0th stage it’s simple: we just give each
conditional value 12 . At each successor stage, we let A → B have value 1 if the
value of A at the prior stage is less than or equal to the value of B; otherwise
we give it the value 0. At limit stages, we see if there is a point prior to the
limit such that after that point (and before the limit), the value of A is always
less than or equal to that of B; if so, A → B gets value 1 at the limit. Similarly,
if there is a point prior to the limit such that after that point (and before the
limit), the value of A is always greater than that of B, then A → B gets value 0
at the limit. And if neither condition obtains, A → B gets value 12 at the limit.
That completes the speciÞcation of how each maxi-stage begins; to repeat, it
serves as the input to a Kripke construction that yields values at that stage for
every sentence.10
In typical cases of sentences that are paradoxical on other theories, the val-
ues oscillate wildly from one (maxi-)stage to the next. But we can deÞne the
"ultimate value" of a sentence to be 1 if there is a stage past which it is always 1;
0 if there is a stage past which it is always 0; and otherwise 12 . It turns out that
there are ordinals ∆ ("acceptable ordinals") such that for any nonzero β, the
value of every sentence at stage ∆ · β is the same as its ultimate value. (This is
the "Fundamental Theorem" of [9].) Since the Kleene valuation rules are satis-
Þed at each stage, this shows (among other important things) that the ultimate
values obey the Kleene rules for connectives other than →. As remarked, this
construction validates naive truth theory, both in truth schema and intersub-
stitutivity form. (It validates it in a strong sense: it not only shows naive truth
theory to be consistent, it shows it to be "consistent with any arithmetically
1 0 Obviously there is a similarity to the revision theory of Gupta and Belnap [12]; but they

use a revision rule for the truth predicate instead of for the conditional, and get a classical
logic theory (one of the ones that refuses to commit between (1) and (2)).

8
standard starting model"–conservative, in one sense of that phrase. For a fuller
discussion see [9], note 27.)
One question that arises is the relation between the → and the ⊃. → is not
truth functional,11 but one can construct a table of the possible ultimate values
of A → B given the ultimate values of A and of B:
A→B B=1 B = 12 B=0
1
A=1 1 2 ,0 0
A = 12 1 1, 12 1
2 ,0
A=0 1 1 1
It is evident from this table that A → B is in some ways weaker and in some
ways stronger than A ⊃ B. However, from the assumption that excluded middle
holds for A and for B, we can derive (A ⊃ B) ↔ (A → B) (and (A ⊃ B) ≡
(A → B)). Moreover, from the assumption that excluded middle holds for each
atomic predicate in a set, we get full classical logic for all sentences built up
out of just those predicates. Thus the logic is a generalization of classical, and
reduces to classical when appropriate instances of excluded middle are assumed.
One way to look at the matter is that the logic without excluded middle is
the basic logic, but in domains like number theory or set theory or physics
where we want excluded middle, we can simply assume all the instances of it in
that domain as non-logical premises; this will make the logic of those domains
effectively classical. It is only for truth and related notions that we get into
obvious trouble from assuming excluded middle: there excluded middle gives
inconsistency, given the naive theory of truth.
I think this is a much more attractive resolution of the paradoxes than any
of the classical ones. One of its most attractive features has to do with a widely
held view that any resolution of the paradoxes simply breeds new paradoxes:
"revenge problems". I claim that there are no revenge problems in this logic.
More particularly, you can state in this logic the way in which certain sentences
of the logic are "defective"; because you can do so, and because there is a
consistency proof of naive truth theory in the logic, the notion (or notions) of
defectiveness cannot generate any new paradoxes. I will discuss this in Sections
7 and 8.
I will make one remark now, which is that like the non-classical version
of the Kripke theory, this is not a theory that posits truth-value gaps. In
particular, we can’t assert of the Liar sentence that it isn’t either true or false.
Nor can we assert that it is either true or false. Situations like this, where
we can’t assert either a claim or its negation, may seem superÞcially like the
situation that I complained about in the case of certain classical resolutions of
the paradox, where we are committed to a disjunction in which each disjunct
has bad consequences, but try to avoid those bad consequence by refusing to
decide which of the two disjuncts to assert. But in fact the nonclassical situation
isn’t like that at all. It is true that in the nonclassical examples we would have a
problem if we asserted A and we would have a problem if we asserted ¬A (where
A is a classically paradoxical sentence). But what made that so problematic in
the classical case was that there we were committed to the claim A ∨ ¬A. We’re
1 1 At least not in these values; but see [8] or [7] for an enriched set of semantic values in

which it is.

9
not committed to that in the non-classical case, so our refusal to commit to either
the classically paradoxical A or to its negation is not a defect in the account.
Similarly, we’re not committed to the claim that either A lacks truth value or
it doesn’t lack truth value, so the refusal to commit to A’s being "gappy" or to
its being "non-gappy" is no defect.

3 Vagueness and indeterminacy

Before discussing the revenge problem, let’s move away from the semantic para-
doxes to other quasi-paradoxes.
Many members of the right-to-life movement think that there is a precise
nanosecond in which a given life begins, though we may not know when it is.
Most of us think that this view is absurd, but Timothy Williamson [23] has in
effect offered an interesting argument that the right-to-lifers are correct on this
point. The initial argument goes as follows.
Select a precise moment about a year before Jerry Falwell’s birth, and call it
‘Time 0’. For any natural number N , let ‘Time N ’ mean ‘N nanoseconds after
Time 0’. By the law of excluded middle, we get each instance of the following
schema:

(3P) (Falwell’s life had begun by time N ) ∨¬(Falwell’s life had begun
by time N ).

From a Þnite number of these plus the fact that Falwell’s life hadn’t begun by
time 0 plus the fact that it had begun by time 1018 , plus the fact that for any
N and M with N < M, if Falwell’s life had begun by time N then it had begun
by time M , a minimal amount of arithmetic and logic yields that

(F) There is a unique N0 such that Falwell’s life had begun by time
N0 and not by time N0 − 1.

But then it seems that there is a fact of the matter as to which nanosecond his
life began, viz. that between time N0 − 1 and time N0 (inclusive of the latter
bound but not the former). That is the initial argument. And the most obvious
way around it is to question the use of excluded middle.
There have, of course, been attempts to get around the right-to-lifer’s con-
clusion without giving up classical logic: e.g. by introducing a notion of deter-
minate truth and determinate falsehood such that sentences of form ‘Falwell’s
life began in the interval (N − 1, N ]’ are neither determinately true nor deter-
minately false. But Williamson has given extensions of the initial argument
that close off most of these attempts: the basic strategy is to argue that even if
such sentences are conceded to be neither determinately true nor determinately
false, in whatever sense of determinateness one favors, it’s hard to see why this
should give any sense of non-factuality to the question of when his life began,
given the commitment to (F). Even if I concede that there’s no "determinate"
truth here, in whatever sense I may give that phrase, why can’t I wonder what
the unique N0 is, or wonder whether it is even or odd? Why can’t I be very
worried about the possibility that the unique N0 occurred before I performed a
certain act, or very much hope that N0 is odd? And even if I take the question

10
of whether it is odd to be beyond the scope of human knowledge, why can’t I
imagine an omniscient god who (by hypothesis of his omniscience) knows the
answer; or a Martian who, though not knowing everything, knows this? And
so forth. But if I do wonder these things or have worries or hopes like this or
concede the possibility of beings with such knowledge, all pretense that I am
regarding the question as non-factual seems hollow. In the past I’ve tried to
Þnd a way around this kind of argument, in part by a nonstandard theory of
propositional attitudes within classical logic, but I’ve come to see this task as
pretty hopeless. It now seems to me that rejecting some of the instances (3P) of
excluded middle is the only viable option (short of giving in to the right-to-lifers
on this issue).
But will the no-excluded-middle option work any better? Let’s Þrst get
clear on an issue (which could have been raised in connection with the semantic
paradoxes too) of what it is to "reject" certain instances of excluded middle. We
don’t reject all of them, only some; what exactly is this difference in attitude
we have between those that we reject and those that we don’t?
First of all, "reject" can’t mean "deny", that is, "assert the negation of".
Suppose we deny an instance of (3P), that is, assert

(3N) ¬[(Falwell’s life had begun by time N ) ∨¬(Falwell’s life had

begun by time N )].

The expression in brackets is a disjunction, and surely on any reasonable logic

a disjunction is weaker than either of its disjuncts. So denying the disjunction
has got to entail denying each disjunct, and so asserting (3N) clearly commits
us to asserting both of the following:

(4a) ¬(Falwell’s life had begun by time N )

(4b) ¬¬(Falwell’s life had begun by time N ).

But (4b) is the negation of (4a), so (3N) has led to a classical contradiction.
And as noted before, the conjunction of a sentence with its negation is also
a contradiction in the Kleene logic K3 described previously, in the sense that
there too it implies everything.
Now, that isn’t the end of the matter: instead of using K3 we could follow
Graham Priest [20] and opt for a "paraconsistent logic" on which classical con-
tradictions don’t entail everything, and therefore aren’t so bad as in classical
logic. I wouldn’t dismiss that view out of hand. But there are problems with
using it in the present context. For one thing, since the paraconsistentist accepts
(4a), and (3P) is a disjunction with (4a) as one disjunct, the paraconsistentist
will accept (3P) as well as (3N): (3P) follows from (4a) on any reasonable logic,
including all the standard paraconsistent logics. But then we can argue from
(3P) to Williamson’s conclusion that there is a unique nanosecond in which
Falwell’s life began, in precisely the same way as before, so the conclusion has
not been blocked. The conclusion has been denied–-from (3N) we can conclude
that there is not a unique nanosecond during which his life began12 –-but it has
also been asserted. This classical inconsistency is not in itself a problem, it is
1 2 Indeed, we can conclude both (i) that there are multiple nanoseconds during which his

life began, rather than one, and (ii) that there is no nanosecond during which his life began.

11
just a further instance of paraconsistentist doctrine that classical inconsistency
is no defect; but it is disappointing that we are left in a position of thinking that
the right-to-lifers are no less correct to assert that there is a fact of the matter
as to the nanosecond in which Falwell was born than we are to deny that there
is a fact of the matter.
So rejection must be interpreted in some other way than as denial. A com-
mon claim is that to reject A is to regard it as not true. The problem with
this is that on the most straightforward reading of ‘true’–-and the one I took
great pains to maintain in the earlier sections on the semantic paradoxes–-the
claim that A is true is equivalent to A itself; so asserting that A is not true is
equivalent to asserting ¬A, and this account of rejection reduces to the previous
one.
Perhaps rejection is just non-acceptance? No, that’s far too weak. Compare
my attitude toward

(5) Falwell’s life began in an even-numbered nanosecond

with my attitude toward

(6) Attila’s maternal grandmother weighed less than 125 pounds on

the day she died.

(5) seems intuitively "non-factual", and I reject both it and its negation in the
strongest terms. That is not at all the case with (6): I have no reason to doubt
that this question is perfectly factual. I don’t accept (6) or its negation, for lack
of evidence; but I don’t reject them either, for given the "factuality" of (6) and
its negation I could only reject one by accepting the other. Rejection is more
than mere non-acceptance.13
The same point arises for the acceptance and rejection of instances of ex-
cluded middle. The point would be easier to illustrate here with examples that
have less contextual variation than does ‘life’, and where the higher order in-
determinacy is less prevalent; but let’s stick to the life case anyway. Suppose
I am certain that on my concept of life, if it is determinate that a person’s
conception occurred during a certain minute then it is indeterminate whether
their life began during that minute, but determinate that their life didn’t begin
before that minute. Then if I knew enough about Falwell to be sure that his
conception occurred during some particular precisely delimited minute, and N
were the nanosecond marking the end of that minute, then I would reject the
corresponding instance of (3P). If however I have no very clear idea how old
Falwell is, so that for all I know nanosecond N might be before his conception
or after his birth, I will be uncertain about the corresponding instance of (3P):
1 3 Rejecting A is also not to be identiÞed with believing it impossible that one could have

enough evidence to accept A. Why not? That depends on the notion of possibility in question.
(a) On any interestingly strong notion of possibility, belief in the impossibility of such evidence
does not suffice for rejection: there are intuitively factual ‘yes or no’ questions (e.g. about the
precise goings-on in the interior of the Sun or in a black hole or beyond the event horizon) for
which there is no possible evidence, but because I take them to be factual I could only reject
one answer by accepting the other. (b) On a very weak notion of possibility (e.g. bare logical
possibility), we have the opposite problem: even for claims that seem "non-factual", like (5),
there is a bare logical possibility that there is such a thing as "living force" and that someone
will invent a "living force detector" that could be used to ascertain whether the claim is true.

12
I will neither accept it nor reject it. (The same point can arise even if I do
have detailed knowledge of the times of his conception and birth and the var-
ious intermediate stages: suppose that I’m undecided whether there is a God
who injects vital ßuid into each human body at some precise time, but think
that if there is no such God then N would correspond to a borderline case of
Falwell’s life having begun.) So for instances of excluded middle too, we have
that rejection is stronger than mere non-acceptance.14
Should the failure of all these attempts to explain the notion of rejection
required by the opponent of excluded middle lead us to suppose that there is no
way to make sense of the no-excluded-middle position? No, for in fact there is an
alternative way to explain the concept of rejection (and it doesn’t require a prior
notion of indeterminacy). The key is to recognize that the refusal to accept all
instances of excluded middle forces a revision in our other epistemic attitudes. A
standard idealization of the epistemic attitudes of an adherent of classical logic
is the Bayesian one, which (in its crudest form at least) involves attributing to
each rational agent a degree of belief function that obeys the laws of classical
probability; these laws entail that theorems of classical logic get degree of belief
1. Obviously this is inappropriate if rational agents needn’t accept all instances
of excluded middle. But allowing degrees of belief less than 1 to some instances
of excluded middle forces other violations of classical probability theory. In
particular, if we keep the laws
P (A ∨ B) + P (A ∧ B) = P (A) + P (B)
and
P (A ∧ ¬A) = 0,
then we must accept
P (A ∨ ¬A) = P (A) + P (¬A).
In that case, assigning degree of belief less than 1 to instances of excluded middle
requires that we weaken the law
(7) P (A) + P (¬A) = 1
to
(7w ) P (A) + P (¬A) ≤ 1.

The relevance of this to acceptance and rejection is that accepting A seems

intimately related to having a high degree of belief in it; say, a degree of belief
at or over a certain threshold T > 12 .15 So let us think of rejection as the dual
1 4 The point arises as well in connection with potentially "ungrounded" sentences that may

not be actually "ungrounded". If sentence A is of form "No sentence written in location D is

true", and I know that exactly one sentence is written in location D but am unsure whether
it is ‘1+1=3’ or A itself, then I am not in a position to accept or reject either the sentence A
or the sentence A ∨ ¬A.
1 5 We can take T to be 1, but only if we are very generous about attributing degree of

belief 1. If (as I prefer) we take T to be less than 1, some would argue that the lottery
paradox prevents a strict identiÞcation of acceptance with degree of belief over the threshold;
I doubt that it does, but to avoid having to argue the matter I have avoided any claim of
strict identiÞcation.

13
notion: it is related in the same way to having a low degree of belief, one at
or lower than the co-threshold 1 − T . In the context of classical probability
theory where (7) is assumed, this just amounts to acceptance of the negation.
But with (7) replaced by (7w ), rejection in this sense is weaker than acceptance
of the negation. (It is still stronger than failure to accept: sentences believed
to degrees between 1 − T and T will be neither accepted nor rejected). I take
it that in a case where a sentence A is clearly indeterminate (e.g. case (5), for
anyone certain that there is no such thing as "vital ßuid"), the degree of belief
in A and in ¬A should both be 0.
Some may feel it more natural to say that the degree of belief in "Falwell’s
life began in an even-numbered nanosecond" should be not the single point 0,
but the closed interval [0,1]. That view is easy to accommodate: represent
the degree of belief in A not by the point P (A), but by the closed interval
R(A) =df [P (A), 1 − P (¬A)].16 This is merely a matter of terminology: the
functions P and R are interdeÞnable, and it is a matter of taste which one is
taken to represent "degrees of belief". (In terms of R, acceptance and rejection
of A go by the lower bound of R(A).)
The value of introducing probabilistic notions is that they give us a natural
way to represent the gradations in attitudes that people can have about the
"factuality" of certain questions–-at least, they do when higher order inde-
terminacy is not at issue. To regard the question of whether A is the case as
"certainly factual" is for the following equivalent conditions on one’s degree of
belief to obtain:
P (A) + P (¬A) = 1;
P (A ∨ ¬A) = 1;
R(A) is point-valued;
R(A ∨ ¬A) = {1}.
To regard it as "certainly nonfactual" is for the following equivalent conditions
to hold:
P (A) + P (¬A) = 0;
P (A ∨ ¬A) = 0;
R(A) = [0, 1];
R(A ∨ ¬A) = [0, 1].
In general, the degree to which one believes A determinate is represented (in
the P -formulation) by P (A) + P (¬A); i.e. P (A ∨ ¬A); i.e. 1 − w, where w
is the breadth of the interval R(A); i.e. the lower bound of R(A ∨ ¬A). In
the P -formulation, belief revision on empirical evidence goes just as on the
classical theory, by conditionalizing; this allows the "degree of certainty of the
determinacy of A" to go up or down with evidence (as long as it isn’t 1 or 0 to
start with).
The idea can be used not only for examples like the Falwell example, but
for potentially paradoxical sentences as well. Consider a sentence S that says
that no sentence written in a certain location is true, and suppose that we know
that exactly one sentence is written in that location; our degree of belief that
the sentence in that location is ‘2 + 2 = 4’ is p, our degree of belief that it is
‘2 + 2 = 5’ is q, and our degree of belief that the sentence written there is S
1 6 Note that R(A ∨ ¬A) will always be an interval with upper bound 1; its lower bound will

be 1 − w, where w is the width of the interval R(A).

14
itself is 1 − p − q, which I’ll call r. I submit that our degree of belief in S should
be q, our degree of belief in ¬S should be p, and our degree of belief in S ∨ ¬S
should be p + q, i.e. 1 − r. The key point in motivating this assignment is that
relative to the assumption that the sentence written there is S, then S and ¬S
each imply the contradiction S ∧ ¬S, and so S ∨ ¬S implies this contradiction as
well; given this, it seems clear that if we were certain that the sentence written
there were S, then we should have degree of belief 0 in S, in ¬S and in S ∨ ¬S.
As we increase our degree of certainty that the sentence written there is S, our
tendency to reject the three sentences S, ¬S and S ∨¬S should become stronger.
Of course, the idea that we can attribute to an agent a determinate P -
function (or R-function) is a considerable idealization. Even in the case of
classical P -functions, where we don’t allow P (A) + P (¬A) to be less than 1,
the issue of whether a person’s degree of belief is greater than say 0.7 often
seems indeterminate. How are we to make sense of the indeterminacy here? It
should be no surprise that on my view, we make sense of this by giving up the
presupposition of excluded middle for certain claims of form "X’s probability
function PX is such that PX (A) > 0.7".17 (It isn’t that we need to develop the
theory of probability itself in a non-classical language; where excluded middle
is to be questioned, rather, is in the attribution of a given perfectly classical
probability function to a given agent X. If you like, this gives failures of
excluded middle for "claims about PX ", though not for claims about individual
probability functions speciÞed independently of the agent X.) We can take the
same position for non-classical P -functions or R-functions too. I’m inclined
to think that there is a strong connection between this indeterminacy in the
degree of belief function and "higher order indeterminacy": in cases where X
attributes higher order indeterminacy to A, some assertions about the value of
PX (A ∨ ¬A) (or RX (A ∨ ¬A)) will be ones for which excluded middle can’t be
assumed. In any case, the indeterminacy in attributions of probability doesn’t
essentially change the picture offered in the preceding paragraph: to whatever
extent that we can say that X’s degree of belief function attributes value 1
to A ∨ ¬A, to precisely that extent we can say that X regards A as certainly
factual.
So far I have not said anything about introducing a notion of determinacy
into the language. I have argued that even without doing so, we can represent
a "dispute about the factuality of A" as a disagreement in attitude: a dis-
agreement about what sort of degrees of belief to adopt. An "advocate of the
factuality of A" will have a cognitive state in which P (A ∨ ¬A) is high (i.e. in
which R(A) is close to point-valued). An "opponent of the factuality of A" will
have a cognitive state in which P (A ∨ ¬A) is low (i.e. R(A) occupies most of
the unit interval). I think it important to see that we can do all this without
bringing the notion of determinacy into the language: it makes clear that there
is more substance to a dispute about factuality than a mere debate about how
a term like ‘factual’ or ‘determinate’ is to be used.
1 7 A common suggestion ([17]) is that we should represent the epistemic state of an agent X

not by a single probability function but by a non-empty set ΣX of them. That is in some ways
a step in the right direction, but it too involves unwanted precision; and while that could be
somewhat ameliorated by going to nonempty sets of nonempty sets of probability functions,
or iterating this even further, I think that ultimately there is no satisfactory resolution short
of recognizing that excluded middle fails for some attributions.

15
 Still, what we have so far falls short of what we might desire, in that so far
we have no means to literally assert the nonfactuality of the question of whether
A: having a low degree of belief in A ∨ ¬A is a way of rejecting the factuality
of A, but not of denying it. It would be very awkward if we couldn’t do better
than this: debates about the factuality of questions would be crippled were we
unable to treat the claim of determinacy or factuality as itself propositional.
What we need, then, is an operator G, such that GA means intuitively that A
is a determinate (or factual) claim, i.e. that the question of whether A is the
case is a determinate (or factual) question. Actually it’s simpler to take as basic
an operator D, where DA means that it is determinately the case that A. The
claim GA (that it is determinate whether A) is the claim that DA ∨ D¬A. The
point of the operator is that though ¬(A∨¬A) is a contradiction, ¬(DA∨D¬A)
is not to be contradictory.
"Determinately operators" are more familiar in the context of attempts to
treat vagueness and indeterminacy within classical logic, and their use there in
representing nonfactuality is subject to a persuasive criticism. The criticism
is that whatever meaning one gives to the D operator, it is hard to see how
¬(DA ∨ D¬A) can represent the nonfactuality of the question of whether A: for
any claims to nonfactuality are undermined by the acceptance of A ∨ ¬A. But
when we have given up the acceptance of A ∨ ¬A, the criticism doesn’t apply.
People can have degrees of belief about determinateness, so their degree of
belief function should extend to the language containing D. If we had only Þrst
order indeterminacy to worry about, and could stick to the idealization of a
determinate degree of belief function PX for our agent X, some constraints on
how PX extends to the D-language would be obvious: since PX (A) + PX (¬A)
represents the degree to which the agent regards A factual, which is PX (DA ∨
D¬A), which in turn is PX (DA)+PX (D¬A), we must suppose that PX (DA) =
PX (A) for any A.18 That is, we must regard the lower bound of RX (DA) as the
same as the lower bound of RX (A). Indeed, with higher order indeterminacy
excluded, RX (DA) should just be a point: PX (¬DA) will be simply 1 − PX (A).
So the fact that DA is strictly stronger than A comes out in that PX (¬DA)
can be greater than PX (¬A) but not less than it, i.e. in that the upper bound
of RX (DA) can be lower than that of RX (A) but not greater than it. It is
however important to allow for higher order indeterminacy, and there may be
some question how best to do so. A proper representation of higher order
indeterminacy presumably should allow excluded middle to fail for sentences of
form DA, so we want to allow that PX (DA)+PX (¬DA) falls short of 1, i.e. that
RX (DA) not be point-valued. I’m inclined to think that we ought to keep the
demand that the lower bound of RX (DA) is always the same as that of RX (A);
this would leave the upper bound unÞxed.19 (As noted before, the situation
1 8 More generally, we could argue that for any A and B, P (DA ∧ B) = P (A ∧ B).
1 9 There are intuitions that go contrary to this: sometimes we seem prepared to assert A
but not to assert "It is determinately the case that A". I’m somewhat inclined to think that
this is so only in examples where A contains terms that are context-dependent as well as
indeterminate, and that it is so because we give to ‘determinately A’ a meaning like ‘under
all reasonable contextual alterations of the use of these terms, A would come out true’; and
this seems to me a use of ‘determinately’ different from the one primarily relevant to the
theory of indeterminacy. But I confess to a lack of complete certainty on these points; for
instance, another possibility would be to allow that in some contexts the upper bound of
R(A) plays a role in governing the assertion of A. (I’d like to thank Richard Dietz, Stephen

16
is complicated by the fact that there is indeterminacy in the attribution of P -
functions or R-functions to the agent, so we can’t assume excluded middle for
all claims about PX and RX . I don’t believe that this requires modiÞcation of
what I’ve said, but the matter deserves more thought than I have been able to
give it.)
I think that what I’ve said clariÞes important aspects of the conceptual
role of the determinately operator. Indeed, until recently I thought that not
a whole lot more could be said to clarify that operator. I now realize that
that former opinion was far too pessimistic: in fact, I’m tempted to say that
the determinately operator can be deÞned in terms of a more basic operator,
a conditional ‘→’. As we’ll see in the next section, such a conditional plays
a very central role in the theory of vagueness; in Section 5 I will then make a
case that the conditional required is "essentially the same as" the one used in
connection with the semantic paradoxes. Starting in Section 6, I will explain
how the conditional can be used to deÞne a very good candidate for the deter-
minately operator. The deÞnition of ’determinately’ in terms of the conditional
would not make the probabilistic laws governing the determinately operator ir-
relevant: for they would indirectly constrain our degrees of belief in sentences
involving the conditional. The conditional, however, can be given a very rich
set of deductive relationships; it is these deductive relationships on it, together
with the probabilistic constraints on the determinately operator deÞned from
it, together with that deÞnition, which would jointly clarify the conditional and
the determinately operator together.
The picture just sketched may eventually need to be complicated slightly:
while we can certainly deÞne a very good candidate for the intuitive determi-
nately operator D in terms of the conditional, the deÞned operator D may not
match the intuitive operator in every respect. But even if this turns out to be
so, the intuitive operator and the deÞned operator will share enough properties
for D to provide a very good model of D; in particular, the intuitive laws gov-
erning D, including the laws of its interaction with the conditional, will be very
close to the laws provable for D. In short: I’m tempted by the view that we
have a strict deÞnition of D in terms of ‘→’, but even if not, what we do have
will give rich structural connections between the two that, when combined with
the probabilistic account above, do a great deal to settle the meaning of D.

4 The conditional again

What kind of logic do we want to use for vagueness? For reasons mentioned
already, it should not include the law of excluded middle. I will now assume
that the logic should be like the Kleene logic K3 (or K3+ : see note 9) as regards
the basic connectives ¬, ∧, and ∨. This assumption is both plausible and widely
accepted, and seems to have the best hope of providing a uniÞcation of the logic
of vagueness with the appropriate logic for the semantic paradoxes.
Schiffer and Timothy Williamson for discussions of the complications of any extension of talk
of probability to a language with a determinately operator, when higher order vagueness is
allowed. In particular, Dietz pointed out a substantial problem in an earlier attempt of mine,
in a classical-logic context.)

17
 The standard semantics for the logic K3 (or indeed, K3+ ) is the strong Kleene
valuation tables, mentioned previously, on which each sentence receives one and
only one of the values 1, 0 and 12 . These should not be read ‘true’, ‘false’ and
‘neither true nor false’: then assigning value 12 would be postulating a truth
value gap, which we cannot do for reasons given in Section 2. A somewhat
better reading would be ‘determinately true’, ‘determinately false’, and ‘nei-
ther determinately true nor determinately false’. That isn’t quite right either:
among other things, since it is assumed that each sentence has one of the three
values, this reading would commit us to the claim that excluded middle holds for
attributions of determinate truth and determinate falsehood,20 thereby ruling
out higher order indeterminacy. (This argument assumes that the semantics is
given in a classical metalanguage; an alternative is to keep to the readings ‘de-
terminately true’, ‘determinately false’ and ‘neither of those’, but refrain from
assuming of every sentence that it has one of the three values. More on this
later.) In Sections 5 and 8 I will say a bit more about these issues of how to
interpret the semantics. But for now, the readings ‘determinately true’, ‘deter-
minately false’, and ‘neither determinately true nor determinately false’ will be
close enough to serve our purposes.
In the non-classical treatment of the semantic paradoxes in Section 2, we
saw that there was a strong need to supplement the Kleene logic with a new
conditional: a conditional A → B not deÞned in the classical way, as ¬A ∨ B.
For whatever the merits of that deÞnition in a context where we have excluded
middle, it fails miserably when excluded middle is abandoned: it doesn’t even
obey such elementary laws as A → A, and because of this it is very hard to
reason with.
Indeed, the failure of the connective ¬A ∨ B (which I’ll call the Kleene
conditional, and abbreviate A ⊃ B) as a deÞnition of A → B is perhaps even
more striking in connection with vagueness than it is in connection with the
semantic paradoxes. For consider the following claims:

A: There are nearly 1000 red balls in Urn 1

B: There are nearly 1000 black balls in Urn 1.

In fact, let us suppose, there are exactly 947 red balls and exactly 953 black
balls in Urn 1. Let us imagine a context where both 947 and 953 seem to be
in the borderline region of ‘nearly 1000’, so that we’d be inclined to regard A
and B as each having value 12 . But since the number of black balls is closer
to 1000 than is the number of red balls, we’d surely want to give "If there are
nearly 1000 red balls in urn 1 then there are nearly 1000 black balls in urn 1"
semantic value 1. But if we use the connective ‘⊃’ to deÞne ‘if...then’, we don’t
get this result: ‘A ⊃ B’ will come out with value 12 . So we very much need
another conditional.
The point is not a new one: nearly all advocates of non-classical logics of
vagueness have proposed using a conditional other than ‘⊃’. The most popular
expansion of Kleene-logic is Łukasiewicz continuum-valued logic, aka "fuzzy
2 0 The commitment to one of the three values is DT ∨DF ∨(¬DT ∧¬DF ), which is equivalent

to (DT ∨ DF ∨ ¬DT ) ∧ (DT ∨ DF ∨ ¬DF ); since DF entails ¬DT and DT entails ¬DF ,
this entails (DT ∨ ¬DT ) ∧ (DF ∨ ¬DF ).

18
logic". For the semantics of this conditional we need to further partition the
values other than 0 and 1: instead of just 12 , we allow arbitrary real numbers
between 0 and 1. The value of A ∧ B is then the minimum of the values of
A and B, while the value of A ∨ B is the maximum; and the value of ¬A is 1
minus the value of A. Clearly, as far as these connectives go the valuation rules
are a generalization of the Kleene rules: the Kleene rules result by restricting
attention to the three values 0, 12 , and 1. Moreover, taking 1 as the sole
designated value, the class of valid inferences in these connectives is unaffected.
The point of the new semantics (as far as logic as opposed to pragmatics goes)
is that it enables us to give rules for a new connective →; the valuation rule is
that the value of A → B is 1 if the value of A is less than or equal to that of B,
and otherwise is 1 minus the extent by which the value of A exceeds that of B.
The Łukasiewicz semantics works reasonably well for dealing with vagueness,
but I have never seen a compelling argument for it. And it does have some
intuitive defects: for instance, the linear ordering of values is unintuitive, and
leads to the claim that sentences such as
It is either the case that if Tim is thin then John is old, or that
if John is old then Tim is thin
are logical truths. Moreover, using the Łukasiewicz logic would doom all hope
of a combined treatment of vagueness and the semantic paradoxes which pre-
serves the naive theory of truth, for the naive theory cannot be preserved in the
Łukasiewicz logic ([22], [13]). But despite these doubts about the Łukasiewicz
logic, we do need a new conditional. (Whether the one proposed in connec-
tion with the semantic paradoxes would do for this purpose is something I will
consider in the next section.)
The issue of the conditional is also relevant to an inßuential argument of
Kit Fine’s [10]. Fine argued that classical logic accounts of vagueness are far
superior to nonclassical logic accounts based on the Kleene semantics, in that the
Kleene-based accounts cannot handle "penumbral connections" between distinct
vague terms. Fine’s point is this. Suppose that the claim that an object b
is red and the claim that b is small both get value 12 . It is unsurprising that
the conjunction and disjunction of the two claims should get value 12 ; and that
is what the Kleene tables say. But how about the claim that b is red and the
claim that b is pink, when b is a borderline case of each? Fine thinks that their
conjunction ought to get value 0, and that if b is clearly in the red-to-pink region
their disjunction should get value 1 even though the disjuncts get the value 12 .
I don’t Þnd these intuitions as compelling as Fine does, but there is a more
neutral way to put his point: we ought to be able to say, somehow, that ‘red’
and ‘pink’ are contraries; and we ought to be able to say, somehow, that an
object is red-to-pink, using only the terms ‘red’, ‘pink’ and logical devices. But
the obvious proposals for how to do these things won’t work. E.g. we can’t
say that something is red-to-pink by saying that it is red or pink, since that
gets value 12 when the object is on the border between red and pink; and no
other logical function in Kleene logic will do any better. Put this way, Fine’s
objection against the use of the unadorned Kleene logic is compelling.
But once we add a new conditional to the logic, it is much less obvious that
there is any problem with penumbral connections. For now we can easily ex-

19
plain the idea that ‘red’ and ‘pink’ are contraries: it consists in the fact that
¤∀x[x is red → x is not pink ],21 where ¤ indicates some kind of conceptual
necessity; the claim is that what follows the ¤ is guaranteed by a conceptual con-
straint on the simultaneous values that are allowed for ‘red(x)’ and ‘pink(x)’.22
And we can explain the idea of an object being in the red-to-pink region: "red-
to-pink(x)" just means "x is not red → x is pink".23 We can, if we like, even
introduce a "pseudo-disjunction" t and "pseudo-conjunction" u:
A t B iff (¬A) → B
A u B iff ¬(A → ¬B).
So to call something red-to-pink(x) is to say that it is red t pink, and the
contrariness of the two consists in the fact that nothing can be red u pink. The
logic of u and t will be slightly odd (just what it is will depend on the logic
of →, obviously),24 but perhaps it is close enough to the ordinary ∧ and ∨ to
explain whatever intuitive force there is in Fine’s own way of presenting the
penumbral connection problem.
As we’ll see in Section 6, an appropriate conditional can also be used in
clarifying the notion of determinateness, and the related notion of defectiveness.
But just what sort of conditional is appropriate?

5 Generalizing the previous conditional

Before deciding what the semantics of a conditional appropriate to vagueness
and indeterminacy should be, I need to say something about the role I expect
the semantics to play. There are two approaches to giving a semantic account
of a language with vague terms.
2 1 If → were not contraposable, we’d have to add ¤∀x[x is pink → x is not red ]; but both

the Łukasiewicz conditional and the conditional introduced in Section 2 are contraposable (as
will be the more general conditional introduced in Section 5).
2 2 In the case of the Łukasiewicz semantics, the appropriate constraint would be that the

values assigned to ‘red(o)’ and to ‘pink(o)’ never add to more than 1. In the case of the
Restricted Semantics I suggested for the paradoxes, the appropriate constraint would have to
be that the values assigned to ‘red(o)’ and to ‘pink(o)’ at any sufficiently large stage never
add to more than 1. The more general semantics to be introduced in the next section will
drop the applicability of stages to sentences like ‘red(o)’ and ‘pink(o)’, but will contain a more
general sort of variation of extension; but the constraint will be analogous, that the values
assigned to ‘red(o)’ and to ‘pink(o)’ at any "world" (at least, any "world" near "the actual
one") never add to more than 1.
2 3 In Łukasiewicz semantics, an object will thus satisfy ’red-to-pink’ iff the values assigned

to ‘red(o)’ and to ‘pink(o)’ add to exactly 1; in the semantics suggested for the paradoxes or
its generalization in the next section, the condition is that the values assigned to ‘red(o)’ and
to ‘pink(o)’ add to exactly 1 at any sufficiently large stage, or at any world near the actual
one.
2 4 It may be useful to display the appropriate "possible value tables" for t and u that we

get if we use the semantics outlined in Section 2:

t B=1 B = 12 B=0
A=1 1 1 1
A = 12 1 1, 12 1
2
,0
1
A=0 1 2
,0 0

u B=1 B = 12 B=0
A=1 1 1, 12 0
A = 12 1, 12 1
2
,0 0
A=0 0 0 0

20
 One is to use a perfectly precise metalanguage. This approach has one great
advantage and one disadvantage. The great advantage is that we are all used
to reasoning classically, and there are no real controversies about how to do
it; the semantics will thus be easy to use and have a clear and unambiguous
content. The disadvantage is that there is no hope of precisely capturing the
content of the sentences of a vague language in a precise metalanguage: the
precise metalanguage inevitably draws artiÞcial lines. (This is why higher
order vagueness looks like such a problem when we think about it in terms of a
precise metalanguage.)
The only way one could conceivably give a semantic theory for a language
with vague terms that accurately reßects the content of its sentences is to drop
the assumption of excluded middle in the metalanguage. The metalanguage
could still be like the metalanguage used before, in assigning semantic values
in the set {0, 12 , 1} to sentences; but we could not assume the law of excluded
middle for attributions of such values.
A semantic theory that avoids excluded middle in the metalanguage may be
what we should ultimately aspire to, but it does throw away the great advantage
of a semantics in a classical metalanguage, and I think that for the here and now
it is more useful to give our semantics in a classical metalanguage. A result of
doing this is that we cannot expect our semantics to be totally faithful to the
language it seeks to represent.25
What then should our standards on a semantics in a classical metalanguage
be? There is a very modest conception of semantics, which Dummett ([2])
calls "semantics as a purely algebraic tool", where the only point of a semantics
is to yield an extensionally adequate notion of logical consequence. If that
were the limits on our ambition for a semantics, then it is not obvious that
the Restricted Semantics offered in Section 2 in connection with the semantic
paradoxes couldn’t be simply carried over to the case of vagueness and indeter-
minacy. In describing the construction there, I stipulated that we were to start
out with a perfectly precise base language: one where each atomic predicate
gets a 2-valued extension that it keeps at each stage of the construction. I then
proposed adding the term ‘True’, and I took its semantics to have two features
that might be taken as typical of indeterminacy: Þrst, ‘True’ gets only a "3-
valued extension" ("positive extension", "negative extension" and "remainder")
at each stage, and second, this 3-valued extension varies from stage to stage.
(Its varying from stage to stage is connected with the higher order indetermi-
nacy of ‘True’.) And so a natural thought is that we might simply allow that
predicates like ‘red’ and ‘bald’ could be treated like ‘True’: we could simply
assign to ‘Red’ or ‘Bald’, at each stage α, a 3-valued extension, with this exten-
sion allowed to vary from one stage to the next. There is, actually, a technical
constraint we’d want to impose on the way that the 3-valued extension can vary,
in order to ensure that we could still add ‘True’ to the language in accord with
2 5 For independent reasons, the clariÞcation of the meaning of logical connectives can’t

proceed wholly by means of a formal semantics: any formal semantics will itself use logical
connectives, often the very ones being "explained", and often in ways that would make the
"explanations" grossly circular (e.g. "‘not A’ is true if and only if A is not true", "‘A and B’
is true if and only if A is true and B is true", etc.). In my view, part of what clariÞes all
these notions, the conditional of this section included, is its connection to degrees of belief:
see the discussion at the end of Section 3.

21
the naive truth theory. But the required constraint is fairly evident from the
proof of the technical result mentioned in Section 2.26 With this constraint
imposed, we would validate the naive theory of truth even when there are vague
predicates in the language; and do so in such a way that the same logic of →
that works when we add ‘True’ to a precise language works for vague language
generally.
This is all Þne if we accept the view that the only role of the semantics is to
serve as an "algebraic tool" for getting a consequence relation, but I think it is
reasonable to demand more. True, it is inevitable that a classical semantics for
a nonclassical language won’t reßect the semantics totally faithfully, but we’d
like it to represent it as faithfully as possible, and the semantics just suggested
seems to me to fall short of reasonable expectations. In particular, it isn’t
at all clear what the assignment of 3-valued extensions to stages "means" in
the case of vague predicates. It does seem natural (at least when one models
the semantics in a classical metalanguage, which is what’s under consideration
here) to think of models that have varying 3-valued extensions. What is not
so natural, though, is the well-ordering of these 3-valued extensions into stages.
We certainly don’t have the clear rules constraining the transÞnite sequence of 3-
valued extensions of ordinary vague predicates that we have in the case of ‘True’.
We might be tempted to just require that the 3-valued extension stays the same
from stage to stage; but if we do that, the model will have no representation at
all of higher order vagueness, and that doesn’t seem satisfactory.
What I think is needed is a more substantial generalization of the Restricted
Semantics used for the paradoxes; I’ll call it the Generalized Semantics. A
natural idea for a semantics of vagueness in a classical metalanguage is to employ
an inÞnite set W of classical partial worlds, or just worlds for short, with one of
them @ singled out as privileged. (I prefer to think of them not as alternative
possibilities, but as alternative methods for assigning semantic values to actual
and possible sentences, given the way the world actually is in precise respects;
@ represents "the actual assignment".) We need to equip W with a certain
structure; I propose that to each w ∈ W we assign a (possibly empty) directed
family F- w of nonempty subsets of W , which I call w-neighborhoods. To say
that -Fw is directed means

(*) (∀w ∈ W )(∀U1 , U2 ∈ -Fw )(∃U3 ∈ -Fw )(U3 ⊆ U1 ∩ U2 ).

I will add a few further conditions on the -Fw as we proceed. Think of each
w-neighborhood as containing the "worlds" that meet a certain standard of
closeness to w. I’m allowing that the standards of closeness for different w-
neighborhoods may be incomparable, so that two w-neighborhoods for the same
w can have members other than w in common without one being a subset of the
2 6 A sufficient constraint on the valuation of the atomic predicates other than ‘True’, for a

consistent addition of ‘True’ to be possible, is that for each n-ary predicate p and each n-tuple
of objects o1 ,...,on , the sequence of members of {1, 12 , 0} assigned to p(o1 , ..., on ) satisÞes the
following two conditions: (i) it eventually cycles, in the sense that there are α and β, the latter
non-zero, such that for any two stages γ 1 , γ 2 ≥ α, the values of p(o1 , ..., on ) at these stages
is the same if γ 1 and γ 2 differ by a right-multiple of β; (ii) unless the "cycle length" (the
smallest such β) is 1, then the value of p(o1 , ..., on ) is 12 at all sufficiently high right multiples
of β. As long as this regularity condition is assumed for the base language, the Fundamental
Theorem of [9] goes through with no real change. (And it tells us that the regularity condition
holds for ‘True’ as well.)

22
other. (If you don’t want to allow for incomparability, you could replace (*)
with the simpler requirement that the members of F - w are linearly ordered by
⊆ ; this would simplify the resulting theory a bit without drastically changing
its character.)
It would be natural from the motivation to assume that for any w, w is a
member of each U in F - w . But for reasons that might only be relevant to the
semantic paradoxes, I want to allow for the existence of abnormal worlds that
do not meet this condition; all I assume in general is that @ is normal. I should
also say that in my intended applications, {@} ∈ / -F@ (indeed, it should probably
be the case that for all w, {w} ∈ / -Fw ); so each member of -F@ includes a point
other than @. The validities that follow do not depend on this, but if {@}
were a member of F - @ we would get new validities that would rule out higher
order indeterminacy and prevent the application of the theory to the semantic
paradoxes.
A model for a vague language L will consist of a domain V and for each
w ∈ W an assignment to each n-place atomic predicate of a "3-valued extension"
in V n . (That is, to each (n+1)-tuple whose Þrst element is an n-place predicate
and whose other elements are in V , we assign at each w ∈ W a member of
{0, 12 , 1}.) I will later add a constraint on the assignment of these 3-valued
extensions to atomic predicates. The valuation rules for the Kleene connectives
and quantiÞers at a given world are just the usual Kleene rules; no reference to
other worlds is required. For ‘→’ I propose the following:
|A → B|w is 1, if (∃U ∈ -Fw )(∀u ∈ U )(|A|u ≤ |B|u );
0, if (∃U ∈ -Fw )(∀u ∈ U )(|A|u > |B|u );
1
2 otherwise.
(The stipulation that the members of -Fw are all nonempty is needed to keep
the value of |A → B|w unique.) If w is abnormal, F - w has members that don’t
contain w; we could without alteration restrict the quantiÞcation in the 1 and
0 clauses to such members, given the directedness condition.
I call an inference universally valid if in any model and any world w in
it, whenever all the premises have value 1 at w, so does the conclusion. I
call an inference strongly valid if in any model, whenever all the premises have
value 1 at all normal worlds, so does the conclusion. And I call an inference
valid if in any model, whenever all the premises have value 1 at @, so does the
conclusion. Validity, strong validity and universal validity for a sentence are
just validity, strong validity and universal validity for the 0-premise argument
whose conclusion is that sentence. Validity is the notion that will be of direct
interest, but information about the other two helps in proving results about
validity; besides, it illuminates the semantics to see what is universally valid,
what is merely strongly valid, and what is just valid.
The above stipulations effectively generalize what was done before in the
case of the semantic paradoxes, though this may not be completely obvious.
There, the space W was, in effect, a closed initial segment [0, ∆] of the ordinals,
where ∆ was one of the "acceptable ordinals" proved to exist in the Fundamental
Theorem of [9];27 @ was ∆. For any ordinal α in W , the F - α used in specifying
the semantic values of conditionals at α had as its members all (nonempty)
2 7 To quell a possible worry, I remark that it is possible to choose the acceptable ordinal "in

advance". (Certainly the initial ordinal for any cardinal greater than that of the power set of

23
intervals of form [β, α). So no member of W was normal, not even ∆; however,
I used a slightly different deÞnition of validity, so the notion of validity didn’t
trivialize. How then can I say that the account here "effectively generalizes" the
one there? Because the Fundamental Theorem shows that if we redeÞne -F∆ to
include intervals of form [β, ∆] (making ∆ normal), while leaving the other F -α
as before, then the values of conditionals is unchanged; this allows us to redeÞne
validity in accord with the general account above, and the redeÞned one and
the original are equivalent in extension. In short: the Restricted Semantics
sketched in Section 3 could be rewritten so as to be a special case of the one
just outlined.
Returning to the general case, let’s get some results about which sentences
and inferences are valid, strongly valid, and universally valid. Even without
imposing any additional constraints, we can easily prove the following:
1. The universal validities include every inference that is valid in the Kleene
logic K3 . In addition, the following inferences involving the conditional are
universally valid:
`A→A
` ¬¬A → A
` A → A ∨ B and ` B → A ∨ B
` A ∧ B → A and ` A ∧ B → B
` A ∧ (B ∨ C) → (A ∧ B) ∨ (A ∧ C)
` (A → ¬B) → (B → ¬A)
` (A → ¬A) ↔ ¬(> → A)
` ¬(A → B) → (B → A)
(A → B) ∧ (A → C) ` A → (B ∧ C)
(A → C) ∧ (B → C) ` (A ∨ B) → C
` ∀xA → At (with the usual restrictions on legitimate substitution)
` ∀x(A ∨ Bx) → A ∨ ∀xBx, when x is not free in A

2. In addition, the following are strongly valid:

A, A → B ` B
` ¬(A → B) → (A ∨ ¬B)
3. If A ` C and B ` C are both valid, so is A ∨ B ` C; and analogously
for strong validity and universal validity. (This means that we validate K3+ , not
just K3 .)
The proofs of all of 1-3 are completely straightforward; indeed, the only ones
that even require the directedness condition are the two universal validities with
conjunctions of conditionals as premises.
We can do better if we add two additional constraints on the families -Fw .
First let’s deÞne U1 is w-interior to U2 to mean: U1 , U2 ∈ -Fw ∧ U1 ⊆ U2 ∧ (∀x ∈
U1 )(∃X ∈ -Fx )(X ⊆ U2 ). The Þrst constraint is
(i) For any normal world w, every member of F - w has w-interior subsets.
The second constraint is a generalization of directedness. If { is a cardinal
number, call a family F - of subsets of W {-directed if for any collection S of
subsets of -F that has cardinality no greater than {, there is a U ∈ -F such
the domain V will be acceptable; I suspect that there are acceptable ordinals of much lower
cardinality.)

24
that U is contained in each member of S. (Directedness as deÞned before is
2-directedness; and is equivalent to n-directedness, for any Þnite n ≥ 2.) Then
letting ||V || be the cardinality of the domain V , we stipulate
(ii) For any world w, F - w is ||V ||-directed.
We can now prove the following additional strong validities:
A → B ` (C → A) → (C → B)
A → B ` (B → C) → (A → C)
∀x(Ax → Bx) ` ∀xAx → ∀xBx
∀x(¬Ax → Ax) ` ¬∀xAx → ∀xAx.28
Finally, let us now add the previously-promised constraint on the assignment
of 3-valued extensions to atomic predicates. Actually I’ll impose the constraint
originally only for what I’ll call "standard" predicates, by which I’ll mean, all
but special predicates like ‘True’. There’s no need to be very precise here,
because the idea is that for any "nonstandard" predicates we introduce, we’ll
prove that the constraint holds for them too, though this will not be part of the
initial stipulation governing their 3-valued extensions. So let p be a standard
n-place predicate; we assume

(**) for each choice of n objects o1 , ..., on in V , if hp, o1 , ..., on i

is assigned an integral value (0 or 1) at @, then there is an @-
neighborhood throughout which hp, o1 , ..., on i is assigned that same
value.
We can now prove a general lemma: that for any sentence A all of whose
predicates are standard,
If A takes on an integral value at @ then there is a @-neighborhood
throughout which A takes on that same value.
The proof is by induction on complexity; the only noteworthy cases are the
conditional and the quantiÞer. For the conditional, the noteworthy fact is that
the proof for A → B goes not by the induction hypothesis, but rather by (i). (If
|A → B|@ = 1 then there is an @-neighborhood U throughout which |A| ≤ |B|;
letting U 0 be @-interior to U , it is clear that |A → B|p = 1 for any p in U 0 .
Similarly for 0.) For the quantiÞer case, we again use (ii).

Given this Lemma, we get two more validities (for sentences all of whose
atomic predicates are standard):
2 8 Proofs: (I) A → B ` (C → A) → (C → B): Suppose that w is normal and that

|A → B|w = 1. Then there is a U in -F w throughout which |A| ≤ |B|. By (i), there’s a U 0

that is w-interior to U. Suppose now that |(C → A) → (C → B)|w < 1. Then there’s a
p ∈ U 0 such that |C → A|p > |C → B|p . So either |C → A|p = 1 or |C → B|p = 0. But both
lead to contradiction. For instance, in the Þrst case, there’s a p-region X throughout which
|C| ≤ |A|; by directedness, there’s a p-region Y ⊆ X ∩ U 0 ; so throughout Y , |C| ≤ |A| ≤ |B|,
so |C → B|p = 1, giving the desired contradiction. The second case is analogous.
(II) A → B ` (B → C) → (A → C): Analogous to (I). (Indeed, (II) follows from (I), given
the laws already established.)
(III) ∀x(Ax → Bx) ` ∀xAx → ∀xBx: Suppose that w is normal and that |∀x(Ax →
Bx)|w = 1. Then for all t, |At → Bt|w = 1. So for each t, there is a w-region Ut such that
|At| ≤ |Bt| throughout Ut . By (ii), there is a w-region U such that for each t, |At| ≤ |Bt|
throughout U, and hence such that |∀xAx| ≤ |∀xBx| throughout U. (Recall that every
element of V is assumed to have a name.) So |∀xAx → ∀xBx|w = 1.
(IV) ∀x(¬Ax → Ax) ` ¬∀xAx → ∀xAx: Analogous to (III).

25
 B`A→B
A, ¬B ` ¬(A → B)
The proof of the Þrst is totally trivial (given the Lemma), and the proof of the
second almost so (it uses the directedness condition).
The validities here established for the Generalized Semantics include almost
all the ones that I derived in [9] for the Restricted Semantics (plus two additional
ones that I neglected to consider there); the only one derived there that we don’t
have here is the relatively minor one ¬[(C → A) → (C → B)] ` ¬(A → B).
Actually we would get that too if we made the additional supposition that there
is an @-neighborhood in which all worlds are normal; that assumption seems
fairly plausible in applications of the logic to ordinary cases of vagueness, though
it does not hold in the Restricted Semantics (where the validity in question
holds for a different reason). It seems then that the logic validated by the
Generalized Semantics is the main core of what’s validated by the Restricted
Semantics, and I propose it as the uniÞed logic for vagueness and the paradoxes.
(There may be ways to expand it slightly, e.g. to get the above-mentioned law
¬[(C → A) → (C → B)] ` ¬(A → B).)29
One thing that needs to be shown, if this is really to be a successful uniÞca-
tion of the logic of vagueness with the logic of the paradoxes, is that the naive
theory of truth is not merely "consistent with any classical starting model (that
is adequate to syntax)", as proved in [9], but "consistent with any generalized
model for the logic of vagueness (that is adequate to syntax)". More precisely,
let L be any language without ‘True’, but which may contain vague terms and
the ‘→’; we also suppose that L contains arithmetic, so that syntax can be
done within it. Let M be any generalized model of the sort just described
for a language without ‘True’ (where M is built from a set W of worlds with
distinguished world @, an assignment to each of an -Fw , a domain V , and an
assignment of 3-valued extensions in V n to each n-place predicate that is sub-
ject to (**)); the only restriction is that the extension of arithmetical predicates
be the same at each world and take on only the values 0 and 1, and that the
resulting 2-valued model be a standard model of arithmetic. The claim then is
that given any such generalized model M, we can extend it to a model M ∗ on
the same structure that satisÞes naive truth theory; by ‘extend it’ I mean that
each predicate other than ‘True’ has the same extension in M ∗ as in M. This
can in fact be shown, by a straightforward generalization of the proof given in
[9]; and (**) turns out to hold for ‘True’ as well, so that even the Þnal two
2 9 One natural expansion—which wouldn’t yield that law, but might yield others—is suggested

by the fact that in the Restricted Semantics there is a "uniformness" to the structure imposed
on the worlds; that is, there is a way to compare the size of α1 -neighborhoods to those of
α2 -neighborhoods even when α1 6= α2 , namely the ordinal that must be added to the lower
bound of the interval to get the upper bound. The idea seems natural in the vagueness case
too. So in the general case, we might want to add to the structure an equivalence relation ≈
on the set < of pairs hw, Ui for which U ∈ -F w , with hw1 , U1 i ≈ hw2 , U2 i having the intuitive
meaning that U1 is a w1 -neighborhood and U2 is a w2 -neighborhood and U1 is like U2 in size,
shape and orientation from its "base point" (w1 or w2 as the case may be). Axioms, besides
those of being an equivalence relation and hw1 , U1 i ≈ hw2 , U2 i ⊃ U1 is a w1 -neighborhood
and U2 is a w2 -neighborhood, should certainly include (a) and probably (b):
(a) hw, U1 i ≈ hw, U2 i ⊃ U1 = U2
(b) (∀w1 , w2 )[-F w1 6= ∅ ∧ -F w2 6= ∅ ⊃ (∃U1 , U2 )[hw1 , U1 i ≈ hw2 , U2 i∧
(∀X ⊆ U1 )(∃Y ⊆ U2 )(hw1 , Xi ≈ hw2 , Y i) ∧ (∀Y ⊆ U2 )(∃X ⊆ U1 )(hw1 , Xi ≈ hw2 , Y i)].
Just what additional strength this would bring to the logic I’m not sure.

26
validities established above hold in the full language.30 Without going through
the details, let me just say that predicates other than ‘True’ are assigned the
same 3-valued extension at every stage of the construction: there is no longer
a need to vary them from stage to stage to get reasonable results about higher
order vagueness, since that is already handled by the multiple "worlds" in the
initial stage. We thus avoid the artiÞciality that (early in this section) we saw
would be required were we to use the Restricted Semantics for vagueness.

6 Defectiveness and determinateness

Whether or not one accepts the uniÞcation just proposed between vagueness/indeterminacy
and the semantic paradoxes, we’ve seen that in each case there are strong
grounds for giving up the law of excluded middle: for certain sentences A, we
should not assert that A ∨ ¬A. But we should not assert that ¬(A ∨ ¬A) either:
that lands us in contradiction. How then do we assert the "defectiveness" of A?
Not by simply asserting that we are not in a position to assert A ∨ ¬A: that’s
too weak, since our inability to assert

Attila’s maternal grandfather was bald on the day he died ∨¬(Attila’s

maternal grandfather was bald on the day he died)

is due not to conviction that its disjuncts are defective but to ignorance as to
whether they are defective. (He might have been clearly bald or clearly non-
bald.) Nor can we assert defectiveness by asserting that no possible evidence
could lead anyone to assert A ∨ ¬A; examples about the interiors of black holes
in indeterministic universes, or events outside of the event horizon of any re-
gions that can support conscious beings, can easily be constructed to show this.
I’ve suggested that our treating A as certainly defective consists (to a Þrst ap-
proximation anyway) in our believing A ∨ ¬A to degree 0 (or to degree [0,1],
if we use R-degrees instead of P -degrees; that’s the same as R-believing A to
degree [0,1]). Similarly, our treating A as probably defective consists (to a
Þrst approximation) in our believing A ∨ ¬A to a low degree. So perhaps we
could assert the "defectiveness" of A by saying that we ought to believe A ∨ ¬A
to degree less than some amount (the amount Þxed by the conÞdence of our
assertion of defectiveness).31 But it’s natural to think that we should be able
3 0 An alternative way to think of the procedure, which illustrates more clearly the sense in

which the proposed account is a uniÞcation, is as going from the initial M to a new model
M ∗∗ with the same domain V , but a new set W ∗ of worlds. On this account, W ∗ is W ×
[0, ∆], where ∆ is a large ordinal (a sufficiently large power of ω), and the new @∗ is h@, ∆i.
For hw, αi in W ∗ , F - hw,αi consists of precisely the subsets of form U × [β, α) where in the
original model U ∈ -F w and where β < α, except in the case where α is ∆, in which case the
subsets have form U × [β, α]. Predicates other than ‘True’ get the same value at any hw, αi
that they got in the original model at w; the value of ‘True’ at any hw, αi is determined by a
Kripke construction of the minimal Þxed point given the values of the ‘True’-free sentences.
The model M ∗ (in the formulation in the main text) is simply the submodel of M ∗∗ consisting
of worlds of form hw, ∆i.
The modest additional assumption is that for each w ∈ W , F - w is ||∆||-directed, not merely
||V ||-directed. This may not really be an additional assumption, as I suspect that we can
take ∆ to have cardinality ||V ||; but at the moment I don’t know how to prove that.
3 1 It’s worth noting that we don’t presently have a satisfactory account of the laws of "prob-

ability" (rational degree of belief) for the logic: we do have a satisfactory probability theory

27
to assert the defectiveness in a more direct way, without bringing ourselves or
other believers into the story. Can we do so?
One possibility is simply to introduce an undeÞned operator DA into the lan-
guage, give some laws (e.g. probabilistic laws) for how it works, and deÞne a de-
fectiveness predicate BAD(x) by ¬G(True(x)), i.e. ¬D(True(x))∧¬D¬True(x).
But one worry is that we don’t really understand such an operator. In addition,
in the semantic paradox case it raises a serious issue of "liar’s revenge": there
is reason to worry that it might give rise to new paradoxes, e.g. for sentences
that assert of themselves that they are not determinately true.
It turns out that in the logic suggested in Section 2 and the generalization
of it suggested in Section 5, there is a good case to be made that the only de-
fectiveness predicates we really need are already deÞnable in the language, using
the new conditional. Obviously that would remove the worry about the intel-
ligibility of the operator, if one grants that the conditional itself is intelligible.
Moreover, it would mean that no revenge problem can arise: the construction
sketched in Section 2 (and later generalized) shows that we can validate naive
truth theory for that language (using a certain non-classical logic), and that
means that we can validate it even for sentences about defectiveness and truth
together, so that there is no revenge problem.
A Þrst stab at deÞning a determinately operator D might be this:

D? A =df > → A (or equivalently, ¬A → ⊥)

where > is some logical truth (say B → B, for some arbitrarily chosen B), and
⊥ is some absurdity. In the "worlds" semantics of Section 5, the value of this
at a given world w is 1 if there is a neighborhhod of w throughout which A
has value 1; 0 if there is a neighborhhod of w throughout which A has value
less than 1 (it needn’t be all 0 or all 12 , it could be a mixture of the two); and
1
2 if there is neither of these sorts of neighborhood. The only clear problem
with this as a deÞnition of determinately is that if the world w is abnormal, the
neighborhood needn’t include w, so D? A could have a higher value than A at
w. But the situation can be remedied by a familiar trick: deÞne D as

DA =df (> → A) ∧ A (or equivalently, (¬A → ⊥) ∧ A).

(I use a different typeface than for D so as not to prejudge whether D adequately

deÞnes the notion we need.)
Inferentially, the proposed deÞnition of D obeys the obvious laws. Because
the semantics validates the rule
(8) B ` A → B,
we get the validity of the D-introduction rule

A ` DA.
for the Kleene logic, but it is not obvious how to extend it to the logic that includes the →.
(In what follows I will suggest a possible deÞnition of D in terms of →; that together with my
earlier remark on the desired probability for sentences of form DA would give a constraint on
how the probability theory is extended to the language with →, but would fall far short of
settling it.)

28
Modus ponens guarantees the validity of D? A ` A, but not of ` D? A → A
(since @ might have abnormal worlds in all its neighborhoods); but now that
we’ve shifted from D? to D, we get the universal validity of
` DA → A
(strong D-elimination).
The deÞnition is also quite natural in connection with the Restricted Seman-
tics used for the paradoxes in Section 2: e.g. if A has value less than 1 at a
stage, DA has value 0 at the next stage.
Indeed, the merits of this deÞnition of determinateness come out especially
clearly in its consequences for the semantic paradoxes, in particular as regards
the "revenge problem". It is to this that I now turn.

7 Revenge (1)
A simple illustration of how the determinately operator D works is afforded
by the Liar sentence Q0 , which asserts its own untruth. It’s clear that on the
semantics of Section 2, it must have value 12 at every stage; so the value of DQ0
and of D¬Q0 are both 0 at every stage after stage 0, and so BAD(hQ0 i) has
ultimate value 1. (‘BAD’ is deÞned from D in the same way that ‘BAD’ is
deÞned from D–see the previous section.)
Once we have a determinately operator D, it’s natural to consider a "weak-
ened" liar sentence Q1 , that says of itself that it is not determinately true.32
But the construction outlined in Section 2 shows that this must be consistently
evaluable in the language. Indeed, it isn’t hard to see that it’s value is 12 at all
even ordinals and 1 at all odd ordinals. DQ1 gets value 12 at all even ordinals and
0 at all odd ordinals; ¬DQ1 thus has the same value as Q1 , as desired. ¬D¬Q1
gets ultimate value 1, as we might expect: so we can assert that Q1 is not de-
terminately untrue. As for the claim that Q1 is determinately true, its ultimate
value is 12 , so we can’t assert DQ1 ∨ ¬DQ1 (and indeed, can reject it). So ex-
cluded middle can’t be assumed (and indeed, can be rejected) even for claims of
determinateness: that is, we have a kind of second order indeterminacy. But we
can assert that Q1 isn’t determinately determinately true. So we can assert that
Q1 is BAD2 , where BAD2 (x) means that ¬DD(True(x)) ∧ ¬DD¬True(x). If
we rename the earlier predicate BAD as BAD1 , then badness1 entails badness2 ,
but not conversely.
We can now consider a "still weaker" liar sentence, that says of itself that
it is not determinately determinately true, and so forth. Indeed, we can iterate
the determinately operator a fair way through the transÞnite: for as long as our
system of ordinal notations lasts.33 For each σ for which there is a notation,
we can then consider the σ-Liar Qσ , which says of itself that it is not Dσ -
true (Dσ being the σ-fold iteration of D). We will never be able to assert
Dσ Qσ ∨ ¬Dσ Qσ , whatever the σ, showing that indeterminacy of arbitrarily
3 2 ‘Weakened’ is in quotes since though Q attributes a weaker property than Q does, it
1 0
attributes it to a different sentence.
3 3 At limit stages we need to form inÞnite conjunctions; but we can do that, for limits simple

enough to have a notation in some reasonable system, by using the truth predicate.

29
high levels34 is allowed; but we can assert that Qσ is not determinately untrue,
and not determinatelyσ+1 true. (There will be a notation for σ + 1 whenever
there is one for σ.) DeÞning the predicates BADσ in analogy to the above, we
get that each is more inclusive than the previous, and for each Liar sentence,
there will be a σ for which the Liar sentence can be asserted to be BADσ .
This hierarchy of defectiveness predicates has something of the ßavor of the
hierarchy of truth predicates that we have in the classical case. But it is different
in important ways. For one thing, it affects a much more peripheral notion: the
relevant notion of defectiveness (or paradoxicalness) plays only a marginal role
in the lives of most of us, whereas the notion of truth is ubiquitous. I don’t
think it is that hard to learn to live with the idea that we do not have a uniÞed
notion of defectiveness! that captures all "levels of defectiveness". A second
difference is that we can reasonably hope that whatever the σ, our overall theory
has the virtue of non-defectivenessσ ; this is in sharp contrast to the hierarchical
classical truth theories, where for each σ there are sentences of our overall theory
that we assert while asserting not to be trueσ . So I don’t think that settling
for the hierarchy of defectiveness predicates would be debilitating. (Indeed,
it is especially un-debilitating if we allow schematic reasoning about levels of
determinateness: see [4], [15], and pages 141-3 of [5] for discussions of schematic
reasoning in other contexts.)35
So far I’ve been talking about determinately operators and defectiveness
predicates that are actually deÞnable in the language. But it might be thought
that these deÞnitions do not adequately capture the intuitive notions of deter-
minateness and defectiveness—or to put it in a way that avoids the supposition
that there are unique such notions, it might be supposed that there are intuitive
notions of determinateness and defectiveness that are not adequately captured
by these deÞnitions. (I don’t myself know of any clear reason to suppose that
there are intuitive notions that the deÞnitions fail to capture, but I see no reason
to be dogmatic about the issue.) So we might want to add a new primitive
determinately operator, governed by axioms, from which a corresponding defec-
tiveness predicate could be deÞned. By using the axioms that we know are true
of the deÞned operator D as a guide to the choice of axioms for the primitive
operator D, one can have a reasonable hope of achieving a consistent theory of
determinateness slightly different from that given by D or any of its iterations.
The most natural idea would be to model such a D quite closely on D. In
that case, we would expect that iterations of the operator are non-trivial: more
particularly, that for each value of σ, the Dσ -Liar can’t be asserted not to be
Dσ -true but can be asserted not to be Dσ+1 -true.
I don’t say that it would be hopeless to add to the language a "maximum
strength" determinately operator D! that iterates only trivially, in that D! D!
is equivalent to (intersubstitutable with) D! ; that would give rise to a uniÞed
defectiveness predicate BAD! (x). But I have doubts as to whether we really
understand this. In any case, if we do add such an operator, we must be careful
3 4 ‘Arbitrarily high levels’ means ‘for as large ordinals as there are notations’; for higher

ordinals than that, talk of levels of indeterminacy makes no clear sense.

3 5 Allowing this is effectively the same as allowing universal quantiÞcation over the σ, but

not allowing such universal quantiÞcations to be embeddable in other contexts like negations.
But whereas a restriction on embedding seems grossly ad hoc, the use of schematic reasoning
has a natural rationale.

30
that the combination of laws we postulate for it does not breed paradox, and
the paradox-free ways of introducing such a D! are not altogether attractive.
Option 1: Suppose we want both the inferential rules D! A |= A and A |=
D! A (perhaps strengthening the Þrst to |= D! A → A); if so, and we also
keep the rule of disjunction elimination as a general logical law, then we must
reject excluded middle for a sentence Q! that asserts that it is not D! -true.36
But if being forced to reject excluded middle shows a kind of indeterminacy,
then any determinateness operator or defectiveness predicate subject to the
rules D! A |= A and A |= D! A thus embodies a kind of indeterminacy (in a
background logic like K3+ that allows for reasoning by disjunction-elimination).
And this kind of indeterminacy (if that is the right way to think of it)37 isn’t
expressible by ¬D! : to assert ¬D! True(hQ! i) would be tantamount to asserting
Q! and thus lead to paradox. Thus D! wouldn’t seem to be the most powerful
determinateness operator after all, in which case the introduction of the notion
would be self-defeating. Option 2: That problem could be avoided by giving up
the rule A |= D! A; we can then keep |= D! A → A, disjunction elimination, and
excluded middle for sentences beginning with D! , for D! True(x) becomes very
much like the unadorned truth predicate in classical theories with (T !! ). The
problem with classical theories containing (T !! ) was that in them we have to
assert certain claims and then deny that thet are true, which seems decidely odd.
Under Option 2 we don’t have precisely that problem (indeed, whenever we can
assert A, we can assert that it is true), but we have something uncomfortably
close to it: we assert of speciÞc sentences A of our own theory that they are
not D! -true, which raises the obvious question "If you think those speciÞc A
aren’t D! -true, why did you assert them?" For this reason, I think that Option
2 should be avoided.

Option 1 may not be beyond defense. Defending it would require refusing

to accept that the "D! -liar" Q! is either defective or non-defective; indeed, it
would require having degree of belief 0 in this disjunction. So we would believe
the equivalent disjunction Q! ∨ ¬Q! to degree 0, and would not be able to
explain our doing so in terms of a belief that its disjuncts were defective. (Or
in any sense "quasi-defective": for the point of D! has been stipulated to be that
there is no further such notion.) The rationale for this would be that to assert
Q! ∨ ¬Q! requires more than that Q! be non-defective, it requires that it be
assertable that it is non-defective; so the inability to assert the non-defectiveness
of Q! is enough to explain our inability to assert Q! ∨¬Q! without commitment
as to whether Q! actually is non-defective. As I say, this may be defensible;
but it is better, I think, to regard the operator D! as not fully intelligible and
3 6 The assumption of D! Q! leads directly to paradox, as does the assumption of ¬D! Q! ; so

by disjunction-elimination, so does the assumption of D! Q! ∨ ¬D! Q! ; but that assumption

is an instance of excluded middle.
3 7 It may not be, for the italicized supposition is in fact dubious. The reason: because of

higher order indeterminacy, excluded middle must fail for claims of defectiveness, so the degree
of belief that a claim is defective and the degree of belief that it is non-defective need not
sum to 1. I’ve equated the degree of belief in A’s non-defectiveness with the degree of belief
in A ∨ ¬A; that means that I can’t in general equate the degree of belief in A’s defectiveness
with 1 minus the degree of belief in A ∨ ¬A, but can only say that the degree of belief in A’s
defectiveness is no greater than 1 minus the degree of belief in A ∨ ¬A. So though we may
reject Q∗ ∨ ¬Q∗ , indeed believe it to degree 0, it does not follow that the degree of belief in
the defectiveness of Q∗ need be high.

31
to simply make do with a hierarchy of determinateness operators.

8 Revenge (2)
But don’t we already have a uniÞed defectiveness predicate, viz., ‘has ultimate
semantic value 12 ’ ? And a predicate corresponding to a uniÞed determinateness
operator, viz. ‘has ultimate semantic value 1’ ? These predicates needn’t even
be added to the language: they are already there, at least if (i) the base language
from which we started the construction in Section 2 included the language of set
theory and (ii) the model M0 for the base language from which the construction
started is deÞnable in the base language;38 for then the set-theoretic construction
of Section 2 can be turned into an explicit deÞnition of these predicates in terms
of the vocabulary of the base language. And since we have assumed excluded
middle for the base language, the deÞnitions show that excluded middle must
hold for attributions of semantic value based on these deÞnitions. Can’t we
then re-institute a paradox, based on sentences that attribute to themselves a
semantic value of less than 1?
No we can’t, and the construction of Section 2 shows that we can’t: our
construction yields the consistency of the claim

True(h||A| | 6= 1i) ↔ ||A| | 6= 1,

and of all other instances of (T) in the language. (Indeed, it shows that we can
add all such claims to the base theory, without disrupting any given model of
that theory.) But it may seem puzzling how paradox has been avoided. Why
isn’t the sentence that attributes to itself a value less than 1 paradoxical?
The answer to this has more to do with the limitations of explicit deÞnition
than it does with the notions of truth and determinacy. Let L be a language
that includes the language of classical set theory. L may contain terms for which
excluded middle is not valid, but let us assume that the set-theoretic portion
of L is classical: excluded middle (and other classical axioms and rules) all
hold for it. Tarski’s undeÞnability theorem shows that if we insist on explicitly
deÞning a predicate, say ‘is a sentence of L that has ultimate semantic value 1’,
in classical set theory, then the deÞned predicate can’t correspond to any normal
notion of truth: it gives the intuitively wrong results even for the classical part
of L. More fully, there will be a sentence A of classical set theory, not containing
‘True’, such that we can prove that either A ∧ ¬(||A| | = 1) or (||A| | = 1) ∧ ¬A.
(With minimal additional assumptions, ones that are met by our deÞnition of
semantic value, we can indeed specify a sentence B such that one disjunct fails
when A is taken to be B and the other fails when A is taken to be ¬B.)39
Thus the notion of ultimate semantic value we’ve deÞned in classical set theory
doesn’t fully correspond to the intuitive notion of truth even for sentences that
are in no way indeterminate or paradoxical. (The reason is that in order to
give a deÞnition of semantic value we have to pretend that the quantiÞers of
3 8 Condition (ii) will be met for any starting model that is at all natural (provided that

conbdition (i) is met).

3 9 The required assumptions are that every sentence of classical set theory gets one of the

values 0 and 1, and that ||¬A| | is 1 when ||A| | is 0 and 0 when ||A| | is 1.

32
the language range only over the members of a given set, viz. the domain of
the starting model, rather than over absolutely everything. What we’ve deÞned
should really be called ‘ultimate semantic value relative to the particular starting
model M0 ’.) That’s why in constructing a theory of truth we need to add ‘True’
as an undeÞned predicate.
Since ‘having ultimate value 1 relative to M0 ’ doesn’t quite correspond to
being true, even for sentences that are in no way indeterminate or paradoxical,
it also doesn’t quite correspond to being determinatly true; for as applied to
sentences of this sort, truth and determinate truth coincide. And that is why
sentences that assert that their own ultimate value is less than 1 (relative to
M0 ) are not genuinely paradoxical: no one who clearly thought through the
limitations of Tarski’s Theorem should have expected ‘having semantic value 1
relative to M0 ’ to precisely correspond to any intuitive notion of determinate
truth. (In particular, it is clear from Tarski’s Theorem that one of the rules

(||A||M0 = 1) |= A

and

A |= (||A| |M0 = 1)

must fail even for sentences of set theory that don’t contain notions like truth
or determinateness–indeed, both rules must fail–so the notions of truth and
determinateness are in no way to blame. Without such rules, the argument for
inconsistency collapses.)
I’ve argued that we can’t within classical set theory deÞne any notion of
determinate truth that fully corresponds to the intuitive notion (if there is a
unique intuitive notion). This is unsurprising; for even if there is a unique
intuitive notion, it doesn’t obey excluded middle and so couldn’t possibly be
deÞned within a classical language. In any case, any explicit deÞnition of
‘has semantic value 1’ in a classical metalanguage is bound to make this notion
demonstrably fail to correspond to any reasonable notion of determinateness,
for the reasons just reviewed. But that raises the question: what is the point
of explicitly deÞning semantic value within classical set theory? I addressed
this question in the context of vagueness at the beginning of Section 5, but it is
worth going through a parallel discussion here.
One very important function of explicitly deÞning ‘has semantic value 1
(relative to starting model M)’ within classical set theory is that doing so enables
us to increase our conÞdence that we have the principles of the non-classical
logic right. For we have an immense amount of experience with classical set
theory, enough to make us very conÞdent of its ω-consistency, and hence of the
correctness of its claims of consistency. The semantics I’ve provided for truth
theory, despite its distortions, gives a proof within classical set theory of the
consistency of naive truth theory in a nonclassical logic, so we know that that
logic is indeed consistent. Indeed, the particular form of the proof shows that
the resulting theory is far more than just consistent, it is "consistent with any
arithmetically standard starting model"; "conservative", in one sense of that
phrase. That implies that it is ω-consistent; it also implies that inconsistency
can’t be generated by combining it with a consistent ‘True’-free theory, e.g.

33
a consistent theory about the history of the Paris Commune. The explicit
deÞnition of semantic value (relative to a starting model) thus serves a very
important logical purpose, even if (as we’ve seen must be the case) there is no
way to turn it into an absolute deÞnition of semantic value that completely
corresponds to any intuitive notion that relates to the full universe.
I don’t mean to suggest that the only function of the semantics is to give a
consistency proof for naive truth theory in the logic. After all, it is certainly not
the sole demand on a logic that it make naive truth theory consistent; the logic
should also be intuitive, and there is no doubt that Þnding a notion of semantic
value in terms of which the valid inferences can be characterized can play a
substantial role in making the logic intuitive. If one starts the construction
from a model M0 that is sufficiently natural (e.g. it is just like the real universe
except for not containing sets of inaccessible rank), then the deÞned notion
of having semantic value 1 relative to M0 is very close to what we’d want of
an intuitive notion of determinate truth (e.g. in the parenthetical example, it
gives intuitive results as applied to any sentence all of whose quantiÞers are
restricted to exclude sets of inaccessible rank); so inferences that preserve value
1 will be valid in a fairly natural sense. In my view, this fact does a great
deal to help make the logic intuitive. It is hard to see how we could do much
better in deÞning determinate truth for this logic in a classical metalanguage,
given that the notion of determinate truth (if there is a unique such notion) is
non-classical.
This does raise an interesting question: mightn’t we develop a non-classical
set theory or property theory without excluded middle, and within it deÞne a
semantic value relation which is appropriate to the full universe (rather than
just to the domain of a classical starting model), but which is otherwise closely
analogous to the deÞnition of the model-relative notion of semantic value given
earlier? And mightn’t this both validate the same logic, and be such that
the deÞned notion ‘has real semantic value 1’ completely corresponds to "the
intuitive notion of determinate truth"? The Þrst part of this can deÞnitely be
done: naive property theory (property theory with unrestricted comprehension)
is consistent in the same extension of Kleene logic used for the semantic para-
doxes: see [7]. (As I brießy discuss there, there is some difficulty in extending
this to naive set theory (i.e. adding an extensionality principle), though I don’t
completely rule out that this can be done.) But as for the program of using
such a theory (whether the non-extensional one or the contemplated extensional
one) for a semantics closely modeled on the one given here,40 I’m skeptical: the
notion of having real semantic value 1 would seem to have to be a uniÞed de-
terminateness predicate much like the D! True(x) of the previous section; and
while this might not raise difficulties about validating the full logic advocated
in this paper as regards sentences not containing that predicate, it would cause
difficulties about validating the application of some of the rules to sentences that
do contain that predicate.41 (A similar reservation arises for the idea of simply
4 0 As
opposed, e.g., to a "homophonic semantics" based on claims like
True(hA → Bi) ↔ [True(hAi) → True(hBi)]
or
True(hDAi) ↔ D(True(hAi));
and also as opposed to a conceptual role semantics of some sort.
4 1 An alternative to modelling the notion of having semantic value 1 on a uniÞed determi-

34
taking a notion of having real semantic value 1 as an undeÞned primitive and
using it in a semantics; the semantics might be OK for the language without
the addition, but not for the expanded language.)
How worrisome is this? Not very. As I already observed in note 25 and
as countless others have observed before me, a semantic theory expressed in a
given logic can explain or justify that logic only in a very minimal sense: the
"explanation" or "justiÞcation" it gives of its logical principles will not only
employ those very logical principles, it will usually do so in a grossly circular
fashion. Logic must stand on its own; the role of a formal semantics for it
(insofar as it goes beyond being a "merely algebraic tool" for consistency proofs),
especially a non-homophonic one (see note 40), is as a heuristic. We who use a
language L governed by the logic I’m recommending get a pretty good though
not perfect heuristic for that logic without expanding the language beyond L,
and indeed by using merely the classical sub-part of L. If we insist upon a
heuristic that avoids the problem of the classical one then we probably do need
to expand L a bit; perhaps this would be done by adding a relation of "real
semantic value" that is not in L but is governed by the same logic as L. If the
use of the logic of L had to depend on the semantics that might be worrisome,
but it doesn’t. Indeed, we don’t even need to regard the heuristic as completely
intelligible: cf. frictionless planes.
Unlike the notions of truth and determinateness, the notion of semantic value
is a technical notion of formal semantics (sometimes a technical notion for giving
consistency proofs, sometimes a technical notion for heuristic "explanations" of
logical principles, sometimes both). The notion of truth, on the other hand,
is certainly not a mere technical notion: as is well known ([21], Ch 1; [16]),
the role of a truth predicate is to serve as a device for conjoining and disjoining
sufficiently regular sets of sentences not otherwise easily conjoined and disjoined,
and this role (which has great importance in everyday life) has little or nothing
to do with formal semantics. To serve this role, what is needed of a notion of
truth is that it adhere as closely as possible to the naive theory of truth, and
the truth theory in this paper adheres to that completely. Settling for only a
model-relative notion of truth would be a huge defeat, in a way that settling for
only a model-relative notion of semantic value is not.
The role of the notion of determinateness is also not especially technical:
we would like to be able to assert the defectiveness of certain sentences in the
language, such as Liar sentences and certain sentences that crucially employ
vague terms, and the determinateness predicate allows us to do so. (It isn’t
something we need to add on to the language, we’ve seen that we get it for free
once we have a conditional suitable for reasoning in absence of excluded middle
and for expressing the naive theory of truth.) We’ve also seen that a single
notion of determinateness can be iterated to express the defectiveness even of
sentences (e.g. extended Liar sentences) that involve the notion of determinate-
ness. Unfortunately we can’t get a single uniÞed notion of defectiveness, but
must rest content with an increasing hierarchy; but I’ve argued that to be less
nateness predicate would be to suppose it hierarchical: i.e., suppose that though the inference
from ‘hAi has semantic value 1’ to ‘hhAi has semantic value 1i has semantic value 1’ is valid,
still the second is stronger in that the conditional with the Þrst as antecedent and the sec-
ond as consequent is not valid. But this too would cause difficulties about validating the
application of some of the rules to sentences that contain the predicate ‘has semantic value’.

35
problematic than a hierarchy of truth predicates would be. There’s no need
to go beyond the language L except perhaps for introducing a heuristic for the
logic we employ; and we don’t even need to do so for that, if we are satisÞed by
the heuristic given by the classical semantics.

9 Conclusion
The logic I’ve proposed for dealing with the paradoxes seems quite satisfactory,
but there could well be others that are equally satisfactory, or better. (I think
the good ones are all likely to be "similar in spirit", in that they will be obtained
from the Kleene logic K3+ by adding a new conditional, in a way consistent with
the naive theory of truth.) I’d like to believe that the good ones can all be
uniÞed with an adequate logic of vagueness, by a uniÞed semantics something
like the one suggested in Section 5 for the logic of paradoxes proposed here. For
we’ve seen some rather general considerations that suggest the naturalness of
uniÞcation:

• The semantic paradoxes seem to arise from the fact that the standard
means for explaining ‘True’ (namely, the truth schema) fails to uniquely
determine the application of the term to certain sentences (the "ungrounded"
ones); and this seems to be just the sort of thing that gives rise to other
sorts of vagueness and indeterminacy.
• For both the semantic paradoxes and for vagueness, it seems important
to give up certain instances of the law of excluded middle. In the case of
the paradoxes, this is required if we are to consistently maintain the naive
theory of truth; in the case of vagueness, we must do so to resist the view
that for every meaningful question there is a fact of the matter as to its
answer.
• For both cases, giving up excluded middle doesn’t involve denying it, but
rather, rejecting it, in a sense that requires explanation. (I believe that
the account of rejection offered in Section 3, in terms of degrees of belief,
is adequate to both cases. In the case of the semantic paradoxes that
don’t turn on empirical premises, there is little need to invoke degrees of
belief other than 0 and 1; which is why the discussion of degrees of belief
ended up playing little role in my discussion of the semantic paradoxes.)
• In both cases we need a reasonable conditional, not deÞnable within Kleene
logic. This is needed to allow for natural reasoning. It is also needed
for more particular purposes: in the case of the semantic paradoxes, it is
needed if we are to maintain the standard truth schema; in the case of
vagueness, it is needed to allow the representation of penumbral connec-
tions between vague concepts, and also needed for expressing such obvious
truths as that if 947 is nearly 1000 then so is 953.
• In both cases we need to be able to express not only a notion of truth
(obeying the naive truth schema) but also a notion of determinate truth;
so we need a determinately operator D. In both cases it is fairly natural
to try to deÞne it from the conditional. And in both cases it is natural to
expect arbitrarily high orders of indeterminacy, at least in the following

36
 sense: no matter how far we iterate D–even into the transÞnite, using
the truth predicate–we never get excluded middle back. In other words,
for any σ, we will never have Dσ A ∨ ¬Dσ A as a general law.

Do we get arbitrarily high orders of indeterminacy in the stronger sense,

that for each σ there are sentences A and possible circumstances under which
we can assert ¬Dσ+1 A but cannot assert ¬Dσ A? As we’ve seen, we get that
in the case of the semantic paradoxes, so it is consistent with the generalized
semantics. I doubt, though, that we want it in the case of vagueness or ordinary
indeterminacy, and I doubt that any choices of the generalized models of Section
5 that are natural to employ in these cases will yield the result. It seems to
me that the best way to represent higher order vagueness in a classical model is
simply to make it the case that no matter how far you reiterate the determinately
operator D, you never bring back excluded middle.
As I noted early in Section 5, a model for vagueness in a classical metalan-
guage is inevitably somewhat distorting. The most obvious distortion is that
it assigns each sentence exactly one of the semantic values 1, 12 , or 0. It is
thus natural to think that when a sentence A gets value 12 , then even when the
semantics does not justify the assertion of ¬DA, then the semantics is really
treating the sentence as indeterminate; in which case the semantics is drawing a
sharp line between those sentences it treats determinate and those it doesn’t. If
so, then the semantics does, in a sense, rule out higher order vagueness, despite
the fact that excluded middle can fail for sentences of form DA (and indeed, for
sentences of form Dσ A for arbitrarily high σ).
But this problem (which is an analog for vagueness of the version of the
revenge problem discussed in the previous section) seems simply to be the in-
evitable result of using a classical metalanguage to do the semantics. As noted
both in Sections 5 and 8, the only way we could hope to get a formal semantics
that portrays a nonclassical language without distortion is to give that semantics
in a nonclassical metalanguage. Whether even the use of a non-classical meta-
language would enable us to develop a semantics (of a basically truth-theoretic
sort) that is both more informative than a merely homophonic semantics and in
no way distorting is a matter on which I take no stand. (I have expressed some
skepticism as to whether such a semantics is needed: it certainly isn’t needed
for an understanding of the language.) In any case, it would be hard to give
the semantics in this way prior to getting a grasp on what the logic to be used
in the semantic theory (and in the object language) ought to be; that’s why in
this paper, in exploring what the logic ought to be, I have restricted myself to
the use of a classical metalanguage.42
4 2 Thanks to Steve Yablo for very useful comments at the conference, and to Graham Priest

for a helpful critique of my discussion of revenge problems that led to a substantial elaboration.
The line on vagueness that I take here grew out of a sequence of unsatisfactory attempts to
elaborate on the classical logic account offered in chapter 10 of [5]. I’m indebted to more
people than I can name for their comments and criticisms of those earlier attempts, but in
addition to those mentioned in note 19, I’d like to single out Joshua Schechter for detailed
and helpful comments at several stages along the way.

37
References
[1] Ross T. Brady. The non-triviality of dialectical set theory. In Graham
Priest, Richard Routley, and Jean Norman, editors, Paraconsistent Logic:
Essays on the Inconsistent, pages 437—470. Philosophia Verlag, 1989.
[2] Michael Dummett. The justiÞcation of deduction. In Michael Dummett,
editor, Truth and Other Enigmas. Harvard University, 1978.
[3] Solomon Feferman. Toward useful type-free theories, I. Journal of Symbolic
Logic, 49:75—111, 1984.
[4] Solomon Feferman. Reßecting on incompleteness. Journal of Symbolic
Logic, 56:1—49, 1991.
[5] Hartry Field. Truth and the Absence of Fact. Oxford University Press,
Oxford, 2001.
[6] Hartry Field. Saving the truth schema from paradox. Journal of Philo-
sophical Logic, 31:1—27, 2002.
[7] Hartry Field. The consistency of the naive(?) theory of properties. In
Godehard Link, editor, One Hundred Years of Russell’s Paradox - Logical
Philosophy Then and Now. Walter de Gruyter, 2003.
[8] Hartry Field. Is the liar sentence both true and false? In JC Beall and
Brad Armour-Garb, editors, Deßationism and Paradox. Oxford University
Press, 2003.
[9] Hartry Field. A revenge-immune solution to the semantic paradoxes. Jour-
nal of Philosophical Logic, 32, 2003.
[10] Kit Fine. Vagueness, truth and logic. Synthese, 30:265—300, 1975.
[11] Harvey Friedman and Michael Sheard. An axiomatic approach to self-
referential truth. Annals of Pure and Applied Logic, 33:1—21, 1987.
[12] Anil Gupta and Nuel Belnap. The Revision Theory of Truth. MIT Press,
Cambridge, MA, 1993.
[13] Petr Hajek, Jeff Paris, and John Shepherdson. The liar paradox and fuzzy
logic. The Journal of Symbolic Logic, 65:339—346, 2000.
[14] Saul Kripke. Outline of a theory of truth. Journal of Philosophy, 72:690—
716, 1975.
[15] Shaughan Lavine. Understanding the InÞnite. Harvard University Press,
Cambridge, MA, 1994.
[16] Stephen Leeds. Theories of reference and truth. Erkenntnis, 13:111—129,
1978.
[17] Isaac Levi. On indeterminate probabilities. Journal of Philosophy, 71:391—
418, 1974.

38
[18] Vann McGee. How truthlike can a predicate be? a negative result. Journal
of Philosophical Logic, 14:399—410, 1985.
[19] Richard Montague. Syntactic treatments of modality, with corollaries on
reßexion principles and Þnite axiomatizability. Acta Philosophica Fennica,
16:153—167, 1963.
[20] Graham Priest. In Contradiction. Martinus Nijhoff, Dordrecht, 1987.
[21] W. V. O. Quine. Philosophy of Logic. Prentice-Hall, Englewood Cliffs,
1970.
[22] Greg Restall. Arithmetic and truth in Łukasiewicz’s inÞnitely valued logic.
Logique et Analyse, 139—140:303—312, 1992.
[23] Timothy Williamson. Vagueness. Routledge, London, 1994.

39