More by D. Rothschild (All Souls, Oxon) on H. P. Grice

In his "Grice, Utterance Choice, and Rationality", D. Rothschild talks about Grice

1 The Status of Gricean Explanations

This paper is about pragmatic explanations of certain linguistic phenomena based on
Grice's theory of conversational implicature. According to Grice's theory, audiences
draw inferences about what the speakers are communicating by reference to the as-
sumption that speakers follow certain maxims governing cooperative conversation. Such
explanations of the inferences audiences make about speakers play a central role in manyparts of philosophy of language and linguistics.

Pragmatic explanations in terms of conversational implicatures are usually preferred
to other sorts of explanations. This practice is justi ed by one or both of these two
ideas: a) that Grice's theory of conversational implicature is a well-con rmed empirical account of aspects of language use, or b) that Grice's theory of conversational implicature actually follows from more basic assumptions about the speech-act situation (such as
the rationality of speakers and their goals). If either of a) and b) is right then Gricean
explanations, if they are su ciently predictive, should be sought where possible.

(Many thanks to Chris Barker, Eliza Block, Jessica Boyd, Richard Breheny, Arudra Burra, Nathan Klinedinst, Michael Rothschild, and Jacob Rosen for help with various bits of this. I'm particularly grateful to Michael Franke and other members of the Philosophy of Language and Beyond" reading
group at the ILLC for their general comments as well as for pointing out two signi cant mistakes in an earlier version. Many thanks, also, to an audience at the Centre for Metaphysics and Mind at the
University of Leeds)

This paper is not about a). As for b): surprisingly there's been rather little detailed discussion of the degree to which Grice's theory of conversational implicature follows
from plausible assumptions about communication. Certainly many people have thought
and said things to the e ect that Grice's maxim of quantity (\make your contribution
as informative as possible") simply follows from the fact that speakers are rational
agents with the goal of communicating as much information to the audience as possible.
Nonetheless, very few people have explained exactly what this claim amounts to or tried
to justify it with much rigor. Another way of seeing whether Gricean explanations are
stipulative or not is by working out whether the maxims are necessary at all in the
derivation of particular implicatures, or whether implicatures can be worked out by
way of more basic assumptions instead. Again there are not many su ciently precise
discussions of this question.1
This paper focuses on scalar implicatures which have received a lot of attention in
the pragmatics literature over the years. The plan is as follows: First, I isolate a means
of deriving certain scalar implicatures by way of a version of the maxim of quantity
(and a few other plausible assumptions). Then I ask the question whether the maxim
of quantity was really necessary in the derivation or whether the same derivation could
have been accomplished instead simply by reference to assumptions about the rationality
and the goals of the participants in speech-act situations. My conclusion is that in some
simple cases the fact that speakers apply the Gricean maxims is predicted by basic
assumptions of rationality and the desire to be informative, but that in more complex

(Notable exceptions occur in the game theory and pragmatics literature (e.g. Benz et al., 2005))

The basic problem, for our purposes, with much of the game theory literature is that the goal of the
enterprise tends to be to show how to derive Gricean implicatures using a game-theoretic approach,
not to assess the extent to which the implicatures are rationally compelling. Although I will use
game-theoretic representations here, this paper is not an attempt at grounding Gricean reasoning using
concepts borrowed from economics, rather it's an attempt at using those concept to assess the degree
to which Gricean reasoning is rationally compelling. The papers in the game theory literature perhaps
most closely related to my concerns are de Jager and van Rooij (2007) and van Rooij (to appear),
however for various reasons I'll touch on later I do not nd even that these papers do much to answer
the exact questions I'm posing.

cases rationality does not predict the application of the Gricean maxims.

1.1 Scalar Implicatures using the Maxims of Quantity and
Quality
Sentence (1) triggers a simple instance of a scalar implicature.

(1) John passed some of the students.

An utterance of (1) would naturally lead the audience to believe that the speaker did
not believe (2) (not that he believes it is false, just that he doesn't believe it is true):

(2) John passed all of the students.

There are probably some assumptions required for the audience to really be entitled to
draw this conclusion. For instance, that the truth of (2) rather than (1) is relevant, that
the speaker is trying to be helpful and wants the audience to know what he does. But if
these assumptions are in place, it is natural for the audience to assume that a speaker
who utters (1) does not believe (2). (It is called a scalar implicature because it involves
the relation between terms like \some" and \all" that are often said to be part of a scale
of strength (Horn, 1972).)
The Gricean story about this inference goes as follows: The speaker said (1). He
could have said (2) instead. (2) entails (1). If he had said (2) it would have been more
informative. The maxim of quantity (limited by constraints on quality ((\say only what
you believe to be true")) and relevance) says that we should utter the most informative,
relevant statement that we believe to be true. So if the speaker had believed (2) he
would have said it. Therefore he does not believe (2).2
2Or, anyway, it goes something like that. . . . Often in the course of this paper, I'll assume without
mentioning that various propositions are relevant to both the and the speaker and audience as well as
3
A well-known problem with this sort of argument is that this general inference scheme
is too strong.3 For instance, it is equally true the speaker could also have said:
(3) John passed some but not all of the students.
However, an utterance of (1) most de nitely does not lead to the inference, generally,
that the speaker does not know that (3) is true|in fact an utterance of (1) often leads
to the inference that the speaker does believe that (3) is true.
Some have tried to respond to this by appealing to the maxim of manner, which
enjoins the speaker to \be brief", however, it is not clear how exactly this works. Cer-
tainly there is no way, based on these simple formulations of the maxims alone, to derive
the implicature associated with (2) but not the one associated with (3). Maybe there
is some way of spelling out the way the maxims interact that will give the right results,
but someone will actually have to do this to provide a Gricean story of the inference
from an utterance of (1) to the fact that the speaker does not believe (2). Such a story
would amount to an algorithm for generating quantity implicatures. I don't know of
any such algorithm, nor am I convinced that such an algorithm is possible (at least
possible to formulate in such a way that it is even plausibly grounded in the various
Gricean maxims). Instead, in the next section, I'll discuss a way of side-stepping this
issue entirely.4
1.2 Scalar Implicatures as Lexical Choice
Speaking a language is a bit like riding a bike. When you ride a bike, certain decisions
you make present themselves as choices. In particular, if you are paying attention, you
the fact that the speaker actually wants to convey as much information as possible to the hearer, since
it gets a bit exhausting to continually reiterate all these assumptions.
3The problem is discussed by Fox (2006), it is attributed to class notes of Irene Heim and Kai von
Fintel and Kroch (1972).
4I am indebted to Eliza Block for many discussions of this issue.
4
choose whether to shift up or shift down the gears of the bike. Other equally important
parts of bike riding, such as how much you lean into turn, do not present themselves as
choices even though what you do is under conscious control. In the activity of speaking
there may also be certain choices you make: For instance whether you say \walked" or
\ran", in a sentence of the form \John to the post o ce" is perhaps sometimes a
salient choice for you. Likewise, whether you should say \some" or \all" in a sentence
of the form \John passed of the students" may also be a choice that presents itself to
you. Perhaps, however, whether to say \some" or \some but not all" in that sentence
is not a choice that generally presents itself to you.
In other words, when you expresses a thought in speech, there are a myriad of
di erent ways you could do so, but only some of the decisions come in the form of
psychologically prominent choices. Now, one could suppose (and to do so is an empirical
supposition) that some choices salient to the speakers are guided by the Gricean maxims.
In particular, we could suppose that when one says a sentence such as (1) it is a salient
choice to say \some of the students" rather than \all of the students", and that this
choice is guided by the Gricean maxims. It may be that Gricean maxims applied globally
are simply too vague to yield any conclusions about what speakers mean when they
speak, but that because of the way language use is structured, we often apply them to
local, salient choices.
Following Horn (1972), I will assume that at least in some instances the choice of
which of a closed class of scalar items to use is itself subject to the Gricean maxims
(independently of all the other choices made in uttering a sentence). In this case, with
some plausible assumptions, we can derive the implicatures we want. When a speaker
utters (1) he is choosing to utter \some" rather than \all", hence he is choosing to utter
(1) rather than (2), and this limited choice is governed by Grice's maxims. The maxim
of quantity enjoins him to say as much relevant information as possible. Let us assume
5
that this amounts to a directive to utter the sentence with the strongest truth-conditions
that you believe (all else equal). So it follows that if the speaker had believed (2) he
would have said it. It follows that since he said (1), he does not believe (2) (i.e. he lacks
belief in the truth (2), he may also have believed it is false). Thus the audience can
infer that a speaker who says (1) does not believe (2). (This inference is what we're
calling the implicature|though Grice probably wouldn't have called such an unadorned
inference an implicature.)
2 Does the Maxim of Quantity Follow from Other
Assumptions?
I am not sure whether or not this manner of deriving scalar implicatures is correct or
not. It is quite similar to the best contemporary accounts (Spector, 2006; Schulz and van
Rooij, 2006; Sauerland, 2004). However, it is not my intention to argue that this way of
doing things is right. Rather I want to ask this question: supposing Gricean reasoning
does operate on certain lexical choices, then do the Gricean maxims (in particular the
maxim of quantity), as applied in these particular situations, follow from more basic
assumptions about the communicative situation? To be concrete consider the above
explanation of the scalar implicature from an utterance (1) to the fact that the speaker
does not believe (2). There we assumed that the speaker was obeying the maxim of
quantity in the sense that he was constrained to utter the most informative sentence that
he could (consistent with his beliefs). Is this assumption, in this particular situation,
eliminable? Could we instead give an explanation that only made reference to some
more basic facts, such as the goals and rationality of the speaker and audience?
The answer to this question might seem very obvious. Here is a way of trying to
reproduce the Gricean reasoning about an utterance of (1) without explicitly using the
6
maxim of quantity: Both the speakers and the audience are rational agents. Moreover,
in the cooperative conversational situation the speaker has relevant information that the
audience lacks. It is one of the speaker's goals to convey as much relevant information
as possible to the audience. Now if we assume that the truth of (2) rather than (1) is
relevant it would seem obvious that a speaker who has the choice of saying (1) or (2)
should say (2) (all else equal).
As it happens, in this particular case this conclusion, and indeed this sort of line of
reasoning, is correct. However, there is a clear respect in which this sort of reasoning
is de cient in general. The problem is this: A rational agent wants to convey as much
information as possible. However, we cannot assume that he wants to say the sentence
whose literal meaning is the strongest. For the rational agent should not care whether
he conveys his information by direct assertion or by implicature. So the fact that the
speaker wants to convey as much information as possible does not immediately yield the
conclusion that the speaker will always say the truth-conditionally strongest sentence
in a choice situation. In other words, if we formulate the maxim of quantity as a
constraint on which sentence a speaker should utter (i.e. the most informative one the
speaker believes) then it does not automatically follow from the simple goal of trying to
convey as much information as possible.5
To derive the maxim of quantity (as it applies in this instance) we need to take
for granted the truth-conditional conventions of the language (i.e.the semantics of the
language and the convention of assertion), and show that a speaker cannot convey (by
any means) more relevant information that he believes by saying (1) rather than (2).
This will be a bit laborious, but it is worth the e ort since we need to see why these
5Note that Grice's own formulation is ambiguous: \make your contribution as informative as is
required for the purposes of the conversation". Does your contribution mean what is said by your
utterance or what is said and what is implicated? When I refer to the maxim of quantity I mean
\make your truth-conditional (pre-implicature) contribution as strong as possible". If you formulate it
the other way, then you have to phrase the dialectic of this paper di erently, but nothing important
changes.
7
explanations succeed in simple cases, in order that we can later see why they fail in
more complex cases.
In order to make our explanation work we need to assume that there are not extra
implicatures (not based on quantity-type reasoning) arising from (1) but not (2) which
would give the speaker an independent reason for saying (1) rather than (2). Let us
divide the possible states of mind of an utterer of (1) or (2) into two types: those who
believe (1) but not (2) and those who believe both (call them type 1 and type 2, respec-
tively). As a way of limiting the possible implicatures (and thus making our problem
solvable) we will assume that the only possible information being directly conveyed is
about whether the speaker is of type 1 or type 2 (by \directly" I don't mean conveyed
by assertion rather than implicature, I mean conveyed rst rather than as the logical
consequence of something else that is conveyed).6 I will also assume that the audience
assigns a positive subjective probability to the proposition that the speaker is of type
1 and to the proposition that the speaker is of type 2 (i.e. he has not ruled out either
option in advance of the speech act).
Now that we have laboriously spelled out the situation, it is easy to see why the
Gricean explanation really does come for free. The conventions of the language and
assertion (i.e. say only what you believe to be true) allow a speaker of type 2 to say
either (1) or (2), but they allow a speaker of type only 1 to say (1). A rational audience
when he hears an utterance of (1) will have to put some degree of credence in the
proposition that the speaker is of type 1 (since we must assume he has not ruled that
possibility out to begin with). A rational audience who hears an utterance of (2) will
also have to believe with full credence that the speaker is of type 2 (since a speaker of
6This is not an unreasonable assumption: lots of other information can be conveyed by utterances of
(1) and (2), but it typically follows from the type of the speaker plus other assumptions. For instance,
that (1) is true follows from the speaker being of type 1 and his beliefs being true. That (2) is false
follows from the speaker being of type 1 and him not being ignorant about the truth of (1) or (2). See
Spector (2006), Schulz and van Rooij (2006), and Sauerland (2004) for extensive discussion of these
latter inferences based on speaker expertise.
8
type 1 cannot say (2), given the semantics of the language and the norms governing
assertion). It follows that any rational speaker of type 2 who wants to convey as much
information to the audience as possible will need to utter (2) rather than (1) given the
choice.
2.1 Game-Theoretic Interlude
The situation described above can be modeled more explicitly using the resources of
game theory. Game theory has the theoretical apparatus to describe decisions that
involve more than one agent: conversational situations are very naturally thought of as
an interaction between two players, the speaker and the hearer.7 Moreover, game theory
has equilibrium concepts which give a precise characterization of pairs of strategies (ways
of acting in the game for both players) that have certain desirable features (such as being
mutually rationalizable in certain respects).
It is best to think of speech games as a special kind of two-player games called
signaling games. A signaling game has essentially three stages: rst the sender gets sent
a message (by nature) revealing his type, then the sender performs an action (he sends a
message), and then the receiver performs an action. The crucial feature of the signaling
game is the informational asymmetry. The sender but not the receiver knows what type
he is, so the hearer when he responds to the sender only knows what action is performed
by the sender not (directly) what type the sender is. A second feature, special to our
game, is that it is a cheap-talk game. This means the payo to both players depends
just on the type of the sender and the action of the receiver, but not at all on the action
of the sender (i.e. the message the sender sent). (Some terminology: each move that
is made by any player|including nature|puts us is in a new node, and corresponding
to every set of moves made up to some point in the game there is separate node.) A
7It is the way Lewis (1969) thought of them.
9
strategy for either player is a complete speci cation of the action the player performs at
each node of the game (although in any given game, some nodes will not be reached).
(More terminology: probabilistically speci ed strategies are called mixed, deterministic
ones are called pure.)
All this is a bit abstract, so let us move to the speci cs of the game we are interested
in, which is actually a very special variant on the basic form of the signaling game (in
some sense it is in fact not a signaling game at all, though it is closely related to one). We
will assume a couple of other things. In our game nature assigns the sender (henceforth,
the speaker) either type 1 or type 2 (which are the same as in the previous section). The
receiver (henceforth, the audience) has antecedent beliefs about the probability that the
the speaker is of type 1 or type 2 and the speaker knows these and neither is zero. There
are two actions available to speakers of type 2: utter (1) or (2), and one action available
to speakers of type 1: utter (1). A special feature of our game is that the receiver does
not perform an action, per se, but rather the payo to both players simply depends on
the degree of credence the audience puts in what type the speaker is. We won't need
to get too speci c about the payo function, but we'll assume that it is ordered by
the degree of credence the audience has in the true proposition about the type of the
speaker. In other words, both players get the same payo and it depends just on how
much credence the receiver puts in the truth about the sender (the more, the better).
Figure 1 diagrams the options for speakers of di erent type. Obviously, there's no choice
of action for the audience, just changing belief states about the speaker.
The standard solution concept for signalling games (which this is a variation on) is
that of a perfect Bayesian equilibrium.8 Given the somewhat di erent character of the
game I am considering from the standard signaling game, I need to substantially adapt
8Gibbons (1992) is a quick and clear introduction to signaling games and perfect Bayesian equilibria;
thicker books on game theory go into more gory detail.
10
type action
t1 /some
t2
uu: uuu uuu uuu / all
Figure 1: Simple \some"/\all" game
the equilibrium concept.9
Variant of Perfect Bayesian Equilibrium (PBE) A speaker strategy s is part of a
perfect Bayesian equilibrium of the speaker strategy and audience's belief dynamics
if the following holds:
1. The speaker knows what type he is prior to his action and the audience has
credences at all points in the game as to what type the speaker is.
2. The audience updates his beliefs as the game progresses according to Bayes'
rule and his belief that the speaker is playing s (insofar as the latter is possible
to maintain).
3. The speaker's action after he has been assigned a type is rational in light of
his type and the fact that the audience will update his beliefs according to
rule 2.10
Let me make some comments. First of all, condition 3 ensures the Nash-style to the
equilibrium: the speaker has no incentive to switch his strategy. Given that the payo s
9I hope, however, it will be recognizable as pretty much the only way to adapt the perfect Bayesian
equilibrium concept to this sort of game. Michael Franke has stressed to me the di erence between
this type of game and the standard signaling games, and hence the di erence between this type of
equilibrium strategy and the normal prefect Bayesian equilibrium concept.
10Note that given a speakers strategy, how the audience updates his belief simply follows from his
antecedent belief in the probability of the di erent speaker types and s, so there is no real \choice" of
audience action. Nonetheless, there is a real sense in which the audience's belief updating and speaker
choice are in equilibrium since the former must be rational in light of the latter and the payo s.
11
are based on the correctness of the audience's belief, the updating of beliefs based on
Bayes' rule will also be the best \response" to the speaker's strategy. Second, note that
there are no tricky assumptions of common knowledge or anything: to check if s is a
PBE all you need to do is calculate what the audience's credences will be after each
message (using Bayes's rule) and then check that s is rational for each type of speaker
in light of those credences (and the payo s associated with them).
More generally, note what this solution concept is getting at: If a strategy is a PBE
(in this variant) then it is a speaker strategy such that if the rational audience believes
the speaker is playing the strategy, then the speaker has no incentive to deviate from it.
This is, in a non-technical sense, makes the strategy rationalizable.
It is easily provable that only one sender's strategy will yield a PBE: the strategy
where the sender of type 2 utters (2) and the sender of type 1 utters (1).
Proof. Suppose a speaker's strategy s is such that there is a positive probability that the
speaker of type 2 utters (1). If the audience knows the speaker's strategy is s, by Bayes's
rule, the audience must give nonzero credence to the proposition that the speaker is of
type 1 when the speaker utters (1), given that the audience had some credence before
that the speaker might be of type 1. In this case the speaker is better o altering s so
that he always utters (2) when he is of type 2 since the audience will always respond to
an utterance of (2) by assigning full credence to the proposition that the speaker is of
type 2.11 So s is not a PBE.
In our game, then, there is only one PBE. Does that mean that it is rationally
compelling that the speaker use that strategy when making a choice between (1) and
(2). I take it that might be a defeasible assumption that the unique PBE is really the
11Note that I assume that when the audience receives a message that isn't expected on s he updates
his belief in a way compatible with Bayes' rule and the structure of the game. Otherwise if s was the
pure strategy of always uttering (1) then it would not be determinate how the audience would react.
12
only rational thing to do. In this case we don't need to rely on the equilibrium concept
to see why people should act as they do. For the reasons explained it's plainly irrational
for speakers of type 2 to do anything but utter (1), and that is su cient to show that
they will act according to the Gricean strategy.12
3 Slightly More Complex: 3-Point Scales
We just considered a choice between two utterances (deriving from a lexical choice
between \some" and \all"). Now consider a choice between three sentences (deriving
from a lexical choice between \some", \most", and \all"):
(4) John passed some of the students.
(5) John passed most of the students.
(6) John passed all of the students.
Corresponding to these three sentences are three types of speaker: type 1 believes only
(4), type 2 believes only (4) and (5), type 3 believes all three statements. Informally
there are pretty good arguments why speakers of type 3 should say (6): However the
audience interprets an utterance of (4) or (5), it seems likely that the audience will put
some credence into the proposition that someone who says (4) or (5) is not of type 3.
Thus it would seem that a speaker of type 3 should utter (6), since that is guaranteed
to make the audience have the right belief about him. If speakers of type 3 always utter
(6) then the rest of the choice looks like the previous game and so we should expect
speakers of type 2 to utter (5). This anyway is the quick, informal version. But as we
12It might be useful to compare the discussion of van Rooij (to appear) to this section, he considers
a related game and comes up with a similar conclusion, using some di erent equilibrium concepts more
appropriate to his particular game. Our approaches di er more substantially for the remainder of the
paper.
13
type action
t1 /some
t2 /
uu: uuu uuu uuu most
t3 /
u: uu uu uu uu u
C
all
Figure 2: \some"/\most"/\all" game
shall see formulation this reasoning in a rigorous and compelling manner is not that
easy.
Let us look at the utterance choice between (4){(6) as a game again. The possibilities
of utterance choice are shown in Figure 2. Again the utterances are limited by the
semantics of the language and the convention that assertion is restricted by belief. Again
we assume that the audience forms a belief about the speaker's type after the speaker
acts, and we assume the payo s (the same for both players) are ordered by the audience's
degree of credence in the truth about the speaker. (As before, we're treating the maxim
of quality as a convention governing assertion.) It is easy to show that there is more
than one perfect Bayesian equilibrium for this game. I will describe two pure speaker's
strategies that make di erent equilibria (Figures 3 and 4).
Gricean Speakers of type 1 say (4), speakers of type 2 say (5), and speakers of type 3
say (6).
Partial Pooling Speaker of type 1 say (4), speakers of type 2 say (4), and speakers of
type 3 say (5).
That the Gricean strategy is part of a perfect Bayesian equilibrium is obvious. But why
is the Partial Pooling strategy? To see this we need to think about what deviations
14
type action
t1 /some
t2 / most
t3 / all
Figure 3: The Gricean strategy for\some"/\most"/\all" game
type action
t1 /some
t2
uu: uuu uuu uuu most
t3
u: uu uu uu uu u
all
Figure 4: Partial Pooling strategy for\some"/\most"/\all" game
15
the speaker could make from the pooling strategy and demonstrate that they will not
improve his payo s assuming the audience believes the speaker is playing the pooling
strategy. Speakers of type 1 have no choice. Speakers of type 2, on the pooling strategy,
say (4) but they could instead say (5). However, saying (5) will reduce their payo s
since the audience will then believe they are of type 3. Speakers of type 3 already get
the audience to believe they are of type 3 with certainty, on the pooling strategy, so
they have no incentive to change their strategy (of course they could just as well say (6)
so it is not a strict perfect Bayesian equilibrium, as the Gricean strategy is).
3.1 Why the Gricean Strategy is Compelling
Even though there are two perfect Bayesian equilibria, one of them may still be rationally
compelling to speakers and audiences in this particular situation. We can rst note that
the Gricean equilibrium has certain advantages over the Partial Pooling equilibrium.
Importantly, players are better o (or at least no worse o ) playing the Gricean strategy,
no matter what the type of the speaker. Also, the Gricean strategy is a strict equilibrium,
but the Partial Pooling strategy is not. (In other words, deviating from the Gricean
strategy always makes the players worse o when the other player plays according to the
Gricean strategy). However, it is not clear that either of these facts makes the Gricean
strategy rationally compelling in any way.
I think, however, that in the speech-act situation we are considering the Gricean
strategy is rationally compelling. In particular, I will argue that given that speakers
have not, in any sense, agreed ahead of time to play according to any particular strategy
and this fact is common knowledge, then speakers will have to play the Gricean strategy.
We already know that speakers of type 1 must play the Gricean strategy (since they
have no choice), so all we need to show is that speakers of types 2 and 3 must play the
Gricean strategy.
16
First I will argue that it is common knowledge that speakers of type 3 will play the
Gricean strategy (and hence that they will actually play it). The argument goes as
follows: we assume that it is common knowledge that speakers have not agreed ahead
of time to play any particular strategy. We will assume, thus, that a speaker of type 3
will prefer a move m to a move m0 if m is guaranteed to yield a payo u given just the
structure of the game and the rationality of the players, while m0 never yields a payo
higher than u and is not guaranteed to yield u on those same assumptions alone. Now
in this case uttering (6) guarantees the top payo for a speaker of type 3 no matter what
the audience believes about the speaker's strategy, but no other utterance guarantees it
in this way. Thus we can reason that given the common knowledge that the players have
not agreed to play the game in any particular way, it follows that speakers of type 3 will
utter (6) (and thus play the Gricean strategy). When something follows from common
knowledge, it is itself be part of common knowledge, so this is also common knowledge.
Now, I will argue that if it is common knowledge that speakers of type 3 play
the Gricean strategy, the only rational strategy for speakers of type 2 is to utter (5).
This actually follows quite quickly. If a speaker of type 2 utters (5) then, given this
assumption, he will induce a belief of probability 1 in his being of type 2 in the audience,
and thus get the highest payo . He cannot do so if he utters (4), so he must utter (5).
We can see, then, that common knowledge of lack of coordination makes one strategy
compelling. The speaking situation we are considering is clearly one in which there is
such lack of coordination (after all, the semantic conventions are meant to exhaust the
pre-arranged aspects of the game), so the Gricean strategy really is rationally compelling
in these cases. In the next section, I will generalize this style of reasoning to a wider
class of games. Readers uninterested in the details should skip this section.13
13van Rooij (to appear) also considers three-point scales. However, he decides for the Gricean strategy
in part by simply stipulating that strategies such as the Partial Pooling one are not allowed.
17
3.2 A Generalization
In this section, I formalize an idea of how people should play games when it is common
knowledge that there is no prior coordination on how to play (in other words, it is
commonly known that there are no pre-established conventions of play). I then prove
that this notion will show that speaker's must choose Gricean strategies in the types of
games we have considered above.
I will not restrict myself to considering the particular scalar-choice signaling games
we looked at above but instead consider a wider class of games. The formalization is
cumbersome, but I think it respects our intuitions of how people should act in these
situations. First, some de nitions:
Conventionlessly Dominates A move m conventionlessly dominates a move m0 if for
some n, m is conventionlessly(n) preferred to m0. A move m is conventionlessly
dominant if it is not conventionlessly dominated by any other move. A strategy s
conventionlessly dominates s0 if s contains a move m and s0 contains a move m0 s.t.
m conventionlessly dominates m0 and s0 doesn't have a move that conventionlessly
dominates a move in s.
Conventionlessly(0) Dominates A move m conventionlessly(0) dominates a move
m0 i there exists a payo u that satis es these conditions.
1. It follows from the common knowledge of the structure of the game and
common knowledge that both players are rational and that m yields a payo
of at least u (for each player).
2. It does not follow from the common knowledge of the structure of the game
and common knowledge that both players are rational and that m0 will give
u.
18
3. m0 never yields a payo higher than u
Conventionlessly(n) Dominates For n > 0, a move m is conventionlessly(n) domi-
nates a move m0 if 1 or 2 obtains:
1. m conventionlessly(n 􀀀 1) dominates m0.
2. There exists a payo u that satis es these three conditions:
(a) It follows from common knowledge of the structure of the game, common
knowledge that both players are rational, and common knowledge that
players will choose conventionlessly(n􀀀1) dominant moves that m yields
a payo of at least u (for each player)
(b) It does not follow from common knowledge of the structure of the game,
common knowledge that both players are rational, and common knowl-
edge that players will choose conventionlessly(n 􀀀 1) dominant moves
that m0 will yield u.
(c) m0 never gives a payo higher than u.
I think that in the case of common knowledge of lack of coordination on any particular
strategy, players should not play conventionlessly-dominated strategies. This is because
players can each assume that each knows that no one will play a conventionlessly(0)
dominated strategy since, if there is lack of coordination, such a strategy is less \safe"
than the strategy that dominates it. Moreover each player knows they each knows this,
and thus can reason similarly that no one will play a conventionlessly(1) strategy, and
so on. . . . So this is not just an arbitrary equilibrium concept, but one that should
manifestly govern the actions of people in this sort of situation.
Now let us consider certain two-player signaling games with n-types of speaker that
have the feature that for a speaker of a given type, the highest payo for both players
19
can be achieved if the audience has completely accurate beliefs about the speaker's
type after the speaker sends his message (this includes not only my kind of game, but
also many more standard cooperative signaling games). We need to make two further
assumptions about the structure of the game: For each type of speaker, i, there is a
unique message k available to him such that no speaker of type < i can also use k. And
one piece of terminology: call the Gricean strategy in this game the strategy on which
each a speaker of type k sends the unique message l s.t. no player of type i < k can also
send l. Call games of this type well-structured communication games. (It should be clear
that the two-point and three-point scale games, considered above, are examples of well-
structured communication games, in addition to certain other games with non-linear
structures.)
I will now give a proof by induction that the Gricean strategy conventionlessly dom-
inates all others in well-structured communication .
Fact 1. In a well-structured communication game with n players, the Gricean strategy
is conventionlessly dominant.
Proof. First the base case:
Fact 2. For speakers of type n the conventionlessly(0) dominant move is the Gricean
one.
Proof. The Gricean move is uniquely available to speakers of type n, so audiences who
receive this message will have to assume that the speaker is of type n with probability
1. For any other move, the audience could rationally assign a probability of less than
1 to the proposition that the speaker is of a di erent type. So the Gricean strategy is
conventionlessly(0) dominant for players of type n.
Now the induction step:
20
Fact 3. If for all players of type > n 􀀀 i the Gricean strategy is conventionlessly(i 􀀀 1)
dominant over all others, then for players of type n 􀀀 i the Gricean strategy will be
conventionlessly(i) dominant over all others.
Proof. We just need to prove that given common knowledge that all players will play
conventionlessly(i 􀀀 1) dominated strategy, then it follows that the Gricean strategy is
the only strategy that yields the correct belief in the audience given just this common
knowledge assumption along with the common knowledge of the structure of the game
and the player's rationality. Given the common knowledge assumption the only sender
who can play the Gricean move for type n 􀀀 i is a player of type n 􀀀 i, so the audience
will have to believe with credence 1 that the player is of type n􀀀i if the speaker makes
the Gricean move for his type. For all other available moves it is clear that this does
not follow, since it is consistent with the assumptions we have made that other senders
of lower types may also use those messages.
4 A Di erent Kind of Case
Our previous cases have all been very simple. It was a bit surprising how hard it is to
show that Gricean speaker strategies are rationally compelling in such very simple cases.
However, what is perhaps even more surprising, is that Gricean speaker strategies are
probably not rationally compelling in slightly more complex cases.
The case I will consider is one much discussed in the recent literature on implicatures,
though in a slightly di erent context. This is the case where there are two scalar terms
in one sentence. The recent literature on these sorts of cases has raised the question of
whether the Gricean programme can handle such cases at all.14
14Chierchia (2004) is largely responsible for the recent interest in this topic. He argues that scalar
21
Let me consider one simple case and show that the sort of Gricean reasoning I
outlined at the beginning of this paper can plausibly be extended to cover this particular
case. Consider an utterance of (7):
(7) Some of the students answered some of the questions.
Let us assume that there was a salient choice between \some" and \all' for both in-
stances of \some" in this sentence. Let us also assume that these choices were made
simultaneously. In this case the speaker was choosing between four sentences. The one
he uttered, (7), and these three others:
(8) All of the students answered some of questions.
(9) Some of the students answered all of the questions.
(10) All of the students answered all of the questions.
Now we assume that a speaker utters the most informative relevant sentence that he
can (the maxim of quantity). Thus (assuming the relevance of all these utterances) if
the speaker uttered (7) it is because he did not believe (8){(10). It's debatable if this
result matches what we nd empirically (which depends so much on focal intonation),
but for these purposes let's assume it does.
Can we derive these implicatures on grounds independent of the maxim of quantity?
It might seem, from previous discussion that would should be able to, but a little thought
can show us that a rather thorny issues arises.
First we need to describe the situation more precisely. Let's assume again that the
speaker is trying to maximize the information he conveys by whatever means to the
implicatures need to be embedded compositionally into the semantics. Gricean responses to Chierchia's
problems are given in Sauerland (2004), Schulz and van Rooij (2006), and Spector (2006).
22
audience. Let's also assume that the only means to convey information (besides direct
assertion) is by some implicature about the extent of the speaker's beliefs based on
what he uttered. Note rst that there are actually ve relevant states of belief15 for the
speaker with respect to these four alternatives sentences (I'll label them with letters so
it's easier to keep track of):
Type ss Believes only (7)
Type as Believes only (7) and (8)
Type sa Believes only (7) and (9)
Type as=sa Believes only (7), (8) and (9)
Type aa Believes (7), (8), (9), and (10)
Here is a version of the intuitive problem: Suppose we assume that a speaker of type
as says (8), a speaker of type sa says (9), and a speaker of type aa says (10) . These
assumptions are, after all, things we can derive from the maxim of quantity. If we assume
these facts, we cannot derive as a further piece of Gricean behavior, that a speaker of
type as=sa should utter either (8) or (9). The reason we cannot derive this is that all we
are assuming is that a speaker will want to convey as much as possible by his utterance.
However, given our assumptions, an utterance of (7) will convey that the speaker is not
of type sa, as, or aa whereas an utterance of type (8) will convey that the speaker is
not of type ss, sa or aa and an utterance of (9) will convey that the speaker is not of
type ss, as or aa. None of these utterances is strictly more informative than the any
other, so, given our assumptions, there is no way to derive the result that a speaker of
type as=sa should say (8) or (9) rather than (7). The problem is this: once we assume
15Again I'm excluding negative beliefs, i.e. beliefs in the falsity of these sentences, since I assume that
inferences about these come indirectly as consequences about inferences about positive belief states (or
their lack) and speaker expertise.
23
type action
ss /some=some
as /
iii4 iiii iiii iiii iiii ii all=some
sa /
ss9 sss sss sss sss sss sss sss sss s some=all
sa=as
iii4 iiii iiii iiii iiii
ss9 ss ss ss ss ss ss ss ss ss ss ss s
}> }} }} }} }} }} }} }} }} }} }} }} }} }} }} }} }
aa /
ss9 sss sss sss sss sss sss sss sss s
}> }} }} }} }} }} }} }} }} }} }} }} }} }} }} }} }}
B all=all
Figure 5: Speaker options in the simultaneous-choice game
the speaker follows the maxim of quantity in certain cases, new implicatures arise which
prevent us from showing that he needs to say the strongest statement he believes in
other cases.
This is just to show that the whole Gricean strategy is not even compelling when we
assume that speakers are following almost all of it. Whether it is compelling without
any assumptions is harder to see. At the least, this indicates that we need a more formal
approach to assessing what strategy the rational speaker (and hence the audience) should
choose.
So let us again look at the signaling game associated with this utterance choice, which
is in Figure 5. We will assume in this game, again, that the payo s are proportional to
the credence the audience puts in the speaker being of the type that he actually is.
Here is a version of the Gricean strategy for this game (see Figure 6):
Gricean Type aa utters (7), type as utters (8), type sa utters (9), type as=sa utters
(8) or (9) (say with 50/50 probability), and type aa utters (10).
24
type action
ss /some=some
as / all=some
sa / some=all
sa=as
iii4 iiii iiii iiii iiii
ss9 ss ss ss ss ss ss ss ss ss ss ss s
aa / all=all
Figure 6: Gricean strategy in simultaneous-choice game
This strategy is obviously a PBE. As with the three-scale game it is provable that
any strategy on which players of type aa do not utter (10) will be untenable if it is
believed by the speaker that the audience is not completely certain as to whether the
players are playing a non-Gricean strategy. However, there are, in this case, non-Gricean
PBEs on which it is common knowledge that speakers of type aa utter (10). Here is one
such strategy, which is in some ways more natural than any Gricean strategy since it is
pure as well asymmetric16 (see Figure 7):
Pure Type aa utters (7), type as utters (8), type sa utters (9), type as=sa utters (7),
and type aa utters (10).
These two strategies are both, I think, entirely reasonable. Neither is Pareto-dominant
over the other. Nor does either equilibrium have the feature that we saw in the three-
stage game where one has parts which are rationally compelled in certain cases of uncer-
16By symmetric, I mean that the symmetries in the structure of the game are mirrored in the
strategies. The Gricean strategy is symmetric, but mixed.
25
type action
ss /some=some
as / all=some
sa / some=all
sa=as
}> }} }} }} }} }} }} }} }} }} }} }} }} }} }} }} }
aa / all=all
Figure 7: Pure strategy in simultaneous-choice game
tainty (the game is obviously not a well-structured communication game in the sense I
de ned in Section 3.2). Obviously, the Gricean strategy has certain properties and (per-
haps) advantages that the non-Gricean strategy does not have. Nonetheless, I think it
would be a hopeless task to argue that one of the two strategies is rationally compelling
for speakers.
As the informal reasoning I gave above suggests, the case of choosing two scalar terms
simultaneously is very complex. The complications presumably arises from the fact that
there are more relevant speaker information states than there are utterances, thus there
is more than one possible way to use one utterance to cover di erent information types
while still using all the utterances.17 My conclusion is that, in these cases, Gricean
behavior according to the maxim of quantity cannot be independently grounded in a
desire to be as informative as possible. Making the strongest possible statement is
only one of many strategies for being maximally informative. Unfortunately, without
17There is a brief discussion in van Rooij (to appear) of cases where there are more speaker types
than utterances.
26
coordination, none of these strategies is guaranteed to work. If the Gricean account is
right, then, we need to have maxims in place as conventions to solve this coordination
problem.
Note that if we add more alternatives to our choice situation we will no longer derive
these same results. However, given the symmetry problem it is clear that we need to
sharply restrict which alternatives are available to a given speaker. My point here is
only to show that on the most standard assumptions about what the alternatives are
in this case, the Gricean strategy does not follow from more basic assumptions about
rationality.
5 Conclusion
This is a very limited study of the degree to which Gricean derivations can be derived
without stipulating maxims as primitive conventions. Many simple cases do seem to
not require the maxims to be stipulated. In more complex cases we either need to
stipulate maxims or make assumptions about the structure of alternatives that may be
quite hard to motivate. My aim here is not to advocate a particular way of thinking
about implicatures, but rather to show that we need to be wary about taking Gricean
explanations as defaults across all cases.


References

Benz, Anton, Jager, Gerhard, and van Rooij, Robert (eds.) (2005). Game Theory and
Pragmatics. Palgrave MacMillan.

Chierchia, Gennaro (2004). Scalar implicatures, polarity phenomenon, and the syntax/pragmatic interface. In Andrea Belleti (ed.), Structures and Beyond. Oxford
University Press.

Fox, Danny (2006). Free choice and the theory of scalar implicatures. Unpublished
manuscript.

Gibbons, Robert (1992). Game Theory for Applied Economists. Princeton University
Press.

Grice, H. P. 1941. Personal identity. Mind.
-- 1948. Meaning
-- 1949. Disposition and intention.
-- 1961. The causal theory of perception.
-- 1966. Logic and conversation: the Oxford lectures.
-- 1967. Logic and Conversation: the William James Lectures
-- 1989. Studies in the way of words.
-- 1991. The conception of value
-- 2001. Aspects of reason.

Horn, Larry (1972). The Semantics of the Logical Operators in English. Ph.D. thesis,
UCLA.
---- A brief history of negation.

de Jager, Tikitu and van Rooij, Robert (2007). Explaining quantity implicatures. In
Proceedings of the 11th conference on theoretical aspects of rationality and knowledge.

Kroch, Anthony (1972). Lexical and inferred meanings from some time adverbs. Quar-
terly Progress Report of the Research Laboratory of Electronics, 104.

Lewis, David (1969). Convention. Harvard University Press.
van Rooij, Robert (to appear). Games and quantity implicatures. Journal of Economic
Methodology.

Sauerland, Uli (2004). Scalar implicatures in complex sentences. Linguistics and Phi-
losophy, 27:367{391.

Schulz, Katrin and van Rooij, Robert (2006). Pragmatic meaning and non-monotonic
reasoning. Linguistics and Philosophy, 29(2):205{250.

Spector, Benjamin (2006). Aspects de la pragmatique des op erateurs logiques. Ph.D.
thesis, Universite Paris 7, Denis Diderot