Davidson’s Semantics and Rationality as Constitutive Ideal: a defence against the objection from...

40
120434 Davidson’s Semantics and Rationality as Constitutive Ideal: a defence against the objection from Chalmers’ Puzzle. UNIVERSITY OF SUSSEX CANDIDATE NUMBER: 120434 HAHP MA PHILOSOPHY SEPTEMBER 2014 SUMMARY: The present paper aims to explicate a reading of Davidsonian semantics which avoids the probabilistic ‘Chalmers’ Puzzle’ objection from David Chalmers’ (2011) paper ‘Frege’s Puzzle and The Objects of Credence’. The said objection denounces theories of rational belief which are committed to what Chalmers terms referentialism and Bayesianism; I contend that despite Davidson’s clear referentialist tendencies in his semantics, and despite his grounding radical interpretation in a Bayesian decision theory, the objection is misguided since his notion of Bayesian rationality is constitutive, and need not be descriptive. In the process, I show how this reading both sidesteps the objection and provides an intersubjective escape route from the problem of speaker knowledge. 1

Transcript of Davidson’s Semantics and Rationality as Constitutive Ideal: a defence against the objection from...

120434

Davidson’s Semantics and Rationality as Constitutive Ideal: a

defence against the objection from Chalmers’ Puzzle.

UNIVERSITY OF SUSSEX

CANDIDATE NUMBER: 120434

HAHP

MA PHILOSOPHY

SEPTEMBER 2014 SUMMARY: The present paper aims to explicate a reading of Davidsonian semantics which avoids the

probabilistic ‘Chalmers’ Puzzle’ objection from David Chalmers’ (2011) paper ‘Frege’s Puzzle

and The Objects of Credence’. The said objection denounces theories of rational belief which

are committed to what Chalmers terms referentialism and Bayesianism; I contend that despite

Davidson’s clear referentialist tendencies in his semantics, and despite his grounding radical

interpretation in a Bayesian decision theory, the objection is misguided since his notion of

Bayesian rationality is constitutive, and need not be descriptive. In the process, I show how

this reading both sidesteps the objection and provides an intersubjective escape route from

the problem of speaker knowledge.

1

120434

CONTENTS

§0.0 ABSTRACT

§1.0 DAVIDSON’S SEMANTICS ­ §1.1 THE SHIFT TO MEANINGS AS TRUTH­CONDITIONS

­ §1.2 RADICAL INTERPRETATION AS EMPIRICAL APPLICATION

­ §1.3 CHARITY AS SOLUTION TO THE PROBLEM OF SPEAKER KNOWLEDGE

§2.0 REFERENTIALISM AND BAYESIANISM IN DAVIDSON

­ §2.1 REFERENTIALISM

­ §2.2 BAYESIANISM

­ §2.3 BAYESIANISM IN RADICAL INTERPRETATION

­ §2.4 CHALMERS’ PUZZLE

§3.0 RATIONALITY AS CONSTITUTIVE IDEAL ­ §3.1 CONSTITUTIVE, NORMATIVE, DESCRIPTIVE

­ §3.2 THE GAP BETWEEN MIND AND WORLD

­ §3.3 THE PROBABILITY AXIOMS AS CONSTITUTIVE, NORMATIVE

­ §3.4 CHALMERS’ PUZZLE REVISITED

­ BIBLIOGRAPHY

2

120434

§0.0 ABSTRACT This paper provides an argument to the conclusion that certain aspects of radical

interpretation can be suitably modified in order to meet an objection from Chalmers’

probabilistic version of Frege’s Puzzle (henceforth: Chalmers’ Puzzle). The difference from

Frege’s original lies in that where Frege’s Puzzle is typically leveraged against extensionalism

using examples from the Philosophy of Language, Chalmers’ Puzzle is pitched instead at the

level of a Bayesian decision theory. In section §1, I present an analysis of radical

interpretation as (i) a consequence of Davidson’s shift towards meanings as truth­conditions;

(ii) an empirical application of Davidsonian semantics; and (iii) wholly reliant on a principle of

charity in order to overcome the problem of speaker knowledge. Therefore, radical

interpretation is of central importance in a defence of Davidson’s extended project. In section

§2, I draw on Rescorla (2013) and Zilhão (2003) in order to demonstrate that Davidson’s

semantics are grounded in a version of referentialism, and his principle of charity is grounded

in a Bayesian decision theory. Since Chalmers’ Puzzle purportedly generates problems for

any view which relies on these two positions, I state a version of the puzzle and outline what I

consider the pertinent consequences for Davidsonian semantics. Section §3 draws further on

Rescorla (2013) and Zilhão (2003) in order to show that Chalmers’ Puzzle is easily overcome

once we note that Davidson’s principle of charity enshrines the constitutive ideal of rationality.

The paper concludes with the objection that Chalmers’ Puzzle is misguided since it only

applies to descriptive accounts of Bayesian decision theory. Since Davidson’s constitutive

ideal of rationality need not commit him to such a descriptive account, I conclude that

Chalmers’ Puzzle is misguided as criticism of Davidson’s extended project.

3

120434

Davidson’s Semantics and Rationality as Constitutive Ideal: a

defence against the objection from Chalmers’ Puzzle. §1.0 DAVIDSON’S SEMANTICS Donald Davidson’s Philosophy, though certainly thorough and expansive in the issues with

which it grapples, consistently returns to a common theme ­ rationality. However, in his

seminal paper 'Truth and Meaning' (1967)1, the topic appears initially to take something of a

back seat in favour of a focus on the ideal form of a semantics for a natural language. The

paper’s opening sentence identifies his vision of the task of semantics:

'a satisfactory theory of meaning must give an account of how the meanings of

sentences depend on the meanings of words' (1967, p.304).

This is an expression of the Fregean compositionality principle, which Davidson takes to

follow inevitably from an undisputed fact about natural languages, namely, that they may

contain potentially infinite numbers of sentences, yet can be mastered on the basis of

grasping a finite amount of words, syntax and logical grammar. The first part of this paper

outlines Davidson’s vision of what a semantics will need to look like if it is to meet the

constraints of the compositionality principle ­ that is, if it is to explain how the meaning of a

putative sentence is derivable from its constituent parts.

Historical accounts of semantics until Davidson tend to employ a framework whereby

sentences, such as 'The weather is warm,' express propositions, such as the proposition that

the weather is warm. In turn, these propositions express propositional content; this is what is

evaluable for meaning, truth or falsity. Frameworks such as these assign a meaning and a

truth­value to every proposition in the language they are a system for. However, this classic

system seems to fail when we incorporate the compositionality constraint. To use Davidson’s

own example in (1967, p.304), the sentence ‘Theaetetus flies’ expresses the proposition that

Theaetetus flies. We can then assign Theaetetus himself as the meaning of ‘Theaetetus’, and

the property of flying to ‘flies’. But how then do we generate the meaning of the whole

sentence using only these tools? It appears that to this end we must appeal to the relation

4

120434

between meanings of ‘Theaetetus’ and ‘flies’. The problem is that we then need to assign a

meaning to the relation ­ and quickly we are thrust into a regress. Hence the Theaetetus

example appears to fail on introduction of the compositionality constraint.

Extensionalism is the view that the meaning of a putative sentence, singular term or predicate

is wholly constituted ­ or exhausted ­ by its extension. The meaning of a singular term, then,

such as ‘Theaetetus’, is exhausted by the set of all things which that term picks out. The same

goes for predicates such as ‘flies’: their meaning is wholly exhausted by the set of all objects

to which the predicate refers. What makes this account 0­dimensional is that it recognises

reference ­ and reference only ­ in its giving an account of the semantics of a language.

Frege’s Puzzle ­ which originally appeared in his (1892)2 ­ is a series of situational examples

in which extensionalist treatment of word meaning is said to yield counterintuitive results. One

way to frame the problem for extensional accounts of word meaning is in terms of

propositional attitude reports, which are sentences containing attributions of beliefs (or certain

other cognitive states) to individuals. These are said to yield counterintuitive results since a

coextensivity principle follows directly from the extensionalist position. This coextensivity

principle maintains that all sentences, singular terms and properties which are alike in

truth­value must be substitutable salva veritate (preserving truth). This means that if

'Superman' and 'Clark Kent' refer to the same individual, then we must be able to switch them

at our leisure in containing sentences. Here is an example where this coextensivity principle

appears to hold true:

(1) Superman grew up on the planet Krypton

(1*) Clark Kent grew up on the planet Krypton

The coextensivity principle holds fast in this case since the sentences remain true despite our

switching the singular terms; since Superman and Clark Kent are the same individual, it is

true that both grew up in the same place. Given that the truth­value did not change, we say

the substitution occurred salva veritate (preserving truth). Frege’s Puzzle works by giving

examples of situations where substitution of coreferential terms does not happen salva

veritate. In this sense, Frege’s Puzzle is less a puzzle than a series of counterexamples to the

extensionalist claim about coextensivity. When we cannot switch coreferential terms salva

5

120434

veritate, we say there is an intensional context. Here is one such example on an intensional

context:

(2) Lois Lane believes that Superman is a superhero.

(2*) Lois Lane believes that Clark Kent is a superhero.

If the coextensivity principle were true, then switching ‘Superman’ with ‘Clark Kent’ would not

yield a change in truth­value. However, although (2) is clearly true, the falsity of (2*) is

essential to the workings of the story, since Lois Lane knows that Superman is a superhero,

but is unaware that Clark Kent is a superhero. And it is precisely this switching of coreferential

terms which delivers the change in truth­value. The conclusion we must take is that

extensional theories of meaning, at least in their current form, must be abandoned. According

to Davidson, 'this is the natural point at which to turn for help to the distinction between

meaning and reference' (1967, p.306). In our terminology, this is the point at which we are

compelled to introduce a first dimension into our formal semantic vocabulary. What naturally

follows Frege’s Puzzle is a conception of meaning as an entity distinct from reference.

Davidson ­ whose vision of the central task of semantics is the satisfaction of the

compositionality constraint ­ is compelled to answer: how does a 1­dimensional account take

us closer to capturing meaning?

The new dimension facilitated by 1­dimensional semantics engages meaning and reference

as two separate entities. Where previously a 0­dimensional account of meaning nominated

the extension of a sentence, singular term or predicate as exhausting meaning, a

1­dimensional account seeks meaning for every sentence, singular term or predicate in the

language. Meanings are thus granted a status as distinct from extension. For Davidson, at

this point, these meanings need to take the form of a meaning­giving sentence entailed by a

theorem for the language we are trying to understand. This takes the form ‘s means m’, where

s is a description of the sentence in the language we seek to understand (henceforth: object

language), and m is a meaning­giving sentence in any form such that interpreters might

understand it (henceforth: metalanguage). What Davidson needs is a theory which takes in all

possible sentences in an object language and yields a meaning­giving theorem in the above

format. Suppose again that 'Theaetetus flies' is one such possible sentence in the object

language. The theory we have in mind is one which generates:

6

120434

(3) ‘Theaetetus flies’ means that Theaetetus flies.

The problem with this kind of meaning­giving theorem is that it fails to be informative (in the

meaning­giving sense) since it is not possible to define the 'means that' except in terms of the

related concept of synonymy. And this is a circular enterprise. Further, whereas a central task

of semantics is to generate a meaning­giving sentence corresponding to every sentence in

the language, there is no clear way to test these meaning­giving sentences for semantic

accuracy: it is not possible to tell, for such examples, whether or not one has missed the

mark. Davidson’s worry that 'we are enmeshed in the intensional springs' (1967, p.309) is

countered by ­ perhaps paradoxically ­ stripping meaning itself of the status as a

meaning­giving entity.

In swift abandonment of the current theme, Davidson abstains both from honing a

1­dimensional account or indeed developing a 2­dimensional narrative for meaning. Instead,

the move is towards truth­conditions of a statement ­ in particular, the circumstances under

which those sentences are true and false ­ as taking on the role of the meaning­giving entity.

Two closely related questions arise: (i) how can truth­conditions fully capture the meaning of a

sentence?; and (ii) what form must a meaning­giving entity take, if it is to yield meaning in

terms of truth­conditions?

§1.1 THE SHIFT TO MEANING AS TRUTH­CONDITIONS

Thus far, we have seen that 0­dimensional theories of meaning cannot deal with the problems

beset upon them by Frege’s Puzzle. We have also seen that the obvious alternative route is

similarly vacuous. This involves trying to generate a theory which yields a meaning­giving

entity of the form ‘s means m’ for every sentence in the object language. Inevitably this fails

since without a plausible, non­circular definition of ‘means that’, we are left in much the same

position as we started. Davidson’s positive proposal involves eliminating ‘meaning’ from his

technical repertoire, at least at the level of formal semantics. Instead, he opts to try to

squeeze a viable semantics for a natural language using only truth­conditions. His inspiration

comes directly from the work of Alfred Tarski who, writing in the 1930s, developed a

truth­conditional formal semantics for logical languages, as well as a useful procedure ­

7

120434

Convention T ­ for testing their adequacy (1956)3. On Tarski’s view, an adequate semantic

theory ­ for a formalised logical language ­ must entail T­sentences which take the following

form:

(4) s is T in L if and only if p

In order to see how Davidson extends this semantics’ domain of application to include natural

languages, there are several crucial differences between T­sentences and the 1­dimensional

meaning­giving theorems which we must note. The overtly intensional ‘means that’ is

discarded in place of a material biconditional ‘if and only if’, meaning ‘m’ is replaced by

metalanguage sentence ‘p’, and ‘in L’ is introduced, which identifies the object language (i.e

the sentence we are trying to interpret). In any possible semantic scenario in which one of

these T­sentences is generated, it is fairly easy to read off features of the situation in order

that we can fill in most of the gaps. For example, suppose someone utters 'water is wet' in

English, and a corresponding T­sentence is generated for a monolingual Englishman such as

myself. On the basis of this information, we can fill out at least the following:

(5) 'Water is wet' is T in English if and only if water is wet.

'Water is wet' is the English sentence under semantic examination. However, where ‘m’

previously figured is now occupied by ‘p’, which has been switched for the sentence 'Water is

wet' in the metalanguage. This is because ‘p’ ­ on Tarski’s theory ­ is supposed to be a

statement of the truth­conditions of the sentence, given in the metalanguage. And the most

obvious candidate for the truth­conditions of a sentence s is just its metalinguistic translation.

The guiding intuition is that if we replace ‘p’ with a trivial, though correct, statement of the

truth­conditions of ‘s’, then the only undefined predicate left in the T­sentence must be

tantamount to truth. In this sense, the definition of truth in a semantics where T­sentences are

prevalent must be reducible to an explicit definition of the predicate ‘is T’.

Not only does Tarski’s work serve to provide the explicit form of a theory of truth for a formal

language, it also provides us with an explicit means by which to test it ­ Convention T. To use

Michael Morris’ (2011)4 expression of the principle (which he calls ‘Tarski’s Test’), this test

amounts to the following:

8

120434

(TT) If you always get a truth from the schematic formula (T) [T­sentences] when you

replace the letter ‘s’ with the name of a sentence, and the letter ‘p’ with a sentence

which gives the meaning of that sentence, then ‘T’, in effect, means true.

In other words, despite the apparent triviality of (5), as long as a T­sentence provides us with

a truth, then there is nothing preventing our giving a definition of truth in terms of an explicit

definition of the predicate ‘is T’. The intuition is that insofar as we can always replace ‘p’ with a

sentence in the metalanguage ­ one with identical truth­conditions ­ all and only sentences

involving the predicate ‘is T’ will be true sentences ­ something which would constitute an

explicit definition.

What Davidson purports to show is two­fold. One aspect of the demonstration is that

T­sentences provide a consistent and non­paradoxical definition of truth; the other is that it is

at least conceivable that such T­sentences might plausibly replace meanings as

meaning­giving theorems in his semantics for natural languages. This marks a change from

the 0­dimensional and 1­dimensional theories where previously meanings as entities in the

form ‘s means m’ were the sole candidates for meaning­giving theorems.

One problem with the view that truth­conditions can take over the role of meaning­giving

entities is that as yet Davidson has failed to show how a semantics drawing on T­sentences

can play any useful explanatory role in our everyday concept of truth and meaning. More

specifically, whilst Davidson does appear to have shown that his theory of truth is coherent

within a certain domain of application, he has not yet given us any reason to suppose that it

has any useful empirical application beyond a formal semantic framework. Radical

interpretation is perhaps best conceived of as Davidson giving his semantics such an

empirical application ­ albeit one within a highly idealised thought experiment. The importance

of this empirical application is two­fold. First, it allows us to see that truth­conditions can yield

meaning­giving theorems in certain cases; secondly, it establishes a presumption in favour of

an extensional theory of truth that sheds light on meaning as the basis for semantics. Both

have consequences which will later become clear.

§1.2 RADICAL INTERPRETATION AS EMPIRICAL APPLICATION

9

120434

According to LePore and Ludwig (2005)5, whereas the Davidsonian semantics we have

outlined constitutes Davidson’s 'initial project', radical interpretation is best conceived of as

'extended'. However, the two projects are complementary. The first gives a nod to pursuing

semantics in terms of a theory of truth; the second attempts to give a substantive theory of

communication by means of an idealised thought experiment in which agents overcome

language barriers by means of T­sentence construction alone. The two are complementary at

least in the following sense: everything there is to be known about truth and meaning is

knowable in principle using the perspective of the radical interpreter. Hence radical

interpretation is best seen not just an empirical application, but also as an empirical test.

On Davidson’s formulation of the radical interpretation situation, a field linguist is immersed

into a foreign community, endowed with no prior understanding of the customs, community or

­ most importantly ­ the language. What options are available to her if she wishes to

communicate? Now it is important to understand that this question is framed at a particular

level. The answer Davidson is looking for will not come from actual empirical observations of

people in analogue situations. This is because Davison’s construction of the situation is

designed in order to encourage a certain method of interpretation ­ utilising T­sentences as

the primary tool.

(4) s is T in L if and only if p

Davidson thinks that only two kinds of evidence are required in order that a field linguist might

begin to interpret a speaker. One kind of evidence is the utterances to which they assent ­

which enters into the T­sentence as ‘s’. The other is the set of environmental conditions which

appear to prompt these ‘hold true’ assertions, which take the place of p. Hence we can

construe the radical interpretation situation as prompting the field linguist with a specific task ­

to replace ‘s’ with each respective ‘hold true’ assertion they are faced with, and to replace ‘p’

with the corresponding environmental prompt. For example, imagine that it starts to rain in the

radical interpretation situation. In seeming response to this environmental stimuli, a native

speaker utters the words 'Llueve'. This is all the information that we need in order to start

pencilling in the modest beginnings of a truth­theory:

10

120434

(6) 'Llueve' is ‘held true’ in L if and only if it is raining.

This modest hypothesis is plausibly confirmed each time the onset of rain is met with native

utterances of 'Llueve'. But just how far does this constitute Davidsonian semantics fully

absorbed? The field linguist’s best efforts above may have constructed a T­sentence, but it is

less clear that the T­sentence so built has all features necessary to play a role in the formal

semantics. This is because the T­sentence above differs from the template (below) in one key

respect.

(4) s is T in L if and only if p

The difference is encapsulated in what we may term the problem of speaker knowledge. The

root of the problem is that the predicate ‘is T’ figures in the Tarskian semantic formula,

whereas the predicate ‘is ‘held true’’ figures in the field linguist’s empirical construction. Where

the Tarskian semantic formula commands that the sentence be such that ‘is T’ is coextensive

with ‘is true’, the field linguist can accommodate this in her empirical construction only insofar

as she can secure coextensivity with ‘is true’ and ‘is held true’. Since ­ in the radical

interpretation situation ­ a speaker does not currently have any independent justification for

holding that the native ‘held true’ is coextensive with ‘is true’ on the formalised theory,

Davidson cannot establish the essential coextensivity claim.

Despite this apparent shortcoming, it is important to note that it would be misdirected to

protest that no actual field linguists use this method ­ nor face this problem ­ in constructing

their own linguistic hypotheses. The claim at question is not that field linguists actually

interpret speakers using this method ­ only that one could. By showing that it is at least

possible in principle for someone in a sparse epistemic position to build T­sentences ­ even

with empirical evidence so limited ­ Davidson goes some way to establishing its importance.

The guiding intuition here is that if a field linguist can construct a Davidsonian semantics on

an empirical base so scarce, then this must provide support for his claim that 'All

understanding of the speech of another involves radical interpretation’ (1973, p315)6.

§1.3 CHARITY AS A SOLUTION TO THE PROBLEM OF SPEAKER KNOWLEDGE

11

120434

At this paper’s outset, I claimed that rationality is a topic that persistently pervades Donald

Davidson’s Philosophy. The next section aims to show that Davidson has no hope of

achieving either of his ‘initial’ or ‘extended’ projects unless he satisfactorily addresses the

problem of speaker knowledge ­ that a speaker’s conviction does not necessarily equal actual

truth­conditions. This is because Davidson requires the principle of charity, and a certain

holistic constraint, in order to address this problem ­ and this brings with it a set of difficult

consequences.

The commitments of the principle of charity vary across formulations. The original expression,

found in Quine’s 'Word and Object' (1960, p.58­9)7, urges that in translation we should avoid

attributing denials of basic logical truths to individuals. This is because of a certain kind of

rational principle according to which 'your interlocutor’s silliness is less likely than your bad

interpretation' (1960, p.59). In other words, if you are consistently interpreting a speaker to

have beliefs and attitudes which differ wildly from your own, then it is more likely to be your

own error in interpretation than their error of judgement which is the source of the trouble.

Davidson’s version is wider in scope than Quine’s, extending its domain from language to

intentionality more broadly. In particular, where Quine’s charity focuses on translation into

one’s own language, Davidson’s emphasises interpretation of both the meaning of utterances

and the content of belief. On Davidson’s view, radical interpretation involves both since

'neither language nor thinking can be fully explained in terms of the other... the two are,

indeed, linked, in the sense that each requires the other in order to be understood'. (1975,

p.156)8 In bringing linguistic meaning and intentional understanding together under one and

the same interpretative project, Davidson so extends the principle of charity to cover meaning

as interdependent with belief.

This interdependence of meaning and belief has come to be known as the holistic constraint

on interpretation. Translation of words and interpreting beliefs are drawn together under the

same project ­ interpretation. According to Malpas (1992)9, this holism is developed in

Davidson’s work such that meaning and information are indistinct. Crucially, the interpreter is

then able to interpret the speaker ­ rather than just her words or beliefs ­ based solely on her

utterance and environment; holding these two factors fixed whilst she solves for an

interpretation.

12

120434

The principle of charity is what allows us to hold fixed utterance and environment as thus

described. In this sense, the it serves dual purpose ­ which take the form of assumptions. The

principle of charity is justification for an interpreter to assume two charitable sub­principles in

the behaviour of the speaker ­ coherence and correspondence. According to coherence, an

interpreter must assume a priori that a speaker’s beliefs and intentional attitudes are by and

large consistent both with one another, and with the normative constraints of rationality. One

way to think about the coherence principle is as an empirical signal to an interpreter: each

time an interpretation appears to attribute an assertion consistent with a logical truth, an

interpreter should continue on their present course;. and each time a speaker appears to deny

a basic logical truth, the interpretation is likely worth revising. Hence coherence is a

methodological maxim ­ which concerns the assumption of logical consistency ­ essential to

the pertinent interpretative enterprise, and hence to interpretation more broadly.

Another way to think about the coherence principle is as the first step in a solution to the

problem of speaker knowledge. I introduced this earlier as the supposed gap between

T­sentences as an interpretative tool in radical interpretation, and the T­sentences required as

meaning­giving theorems in a formal semantics. Earlier I concluded that this is reducible to

the issue making coextensive ‘is true’ (in the former case) and ‘is ‘held true’’ (in the latter).

The principle of coherence is useful here because it permits the assumption that a speaker is

at least rational in the sense that they cannot deny basic logical truths. Not only does this

enable the interpretative project to get off the ground, it also gives us the modest beginnings

of a reason to believe that what is held true by the native speakers might constitute

truth­conditions of an utterance. We move closer to achieving the coextensivity required when

we turn to combining the coherence principle with that of correspondence.

On the charitable principle of correspondence, interpreters are granted a similar

methodological maxim of assumption, but this time related to the way in which the speaker

perceives environmental conditions. Davidson eschews talk of ‘conceptual schemes’ (very

briefly, individuals’ unique ­ or perhaps shared ­ systems for categorising sense data, or way

of seeing the world), but this is not what the principle of correspondence pertains to. Rather,

the thought is that in a putative example from radical interpretation, such as (Quinean)

narrative in which a rabbit scurries by, the interpreter must assume that the speaker sees the

13

120434

world in such a way that they respond as would the interpreter to environmental stimuli. This

allows us to affirm that the environmental conditions prompt respective ‘hold true’ utterances.

Again, another way to think about the principle of correspondence is as warranting an

interpretative assumption ­ this time that speakers conform to an interpreter’s own standards

of response to environmental stimuli. A consequence of this is that ­ together with the

coherence principle ­ we would expect to be able to attribute to speakers a repertoire of true

empirical beliefs. And this gives us yet another reason to think that as far as interpreters are

concerned ­ speakers’ responses to environmental stimuli are as good a guide to

truth­conditions as our own. With a warranted a priori assumption that speakers’ logical and

environmental beliefs are sound, and by and large true, an interpreter has all the warrant

required in order to claim coextensivity with respect to the predicate ‘is true’ and ‘is held true’.

And thus ­ insofar as the principle of charity is itself warranted ­ so the problem of speaker

knowledge appears to have been solved. And so charity cements itself as essential to

Davidson’s 'extended' project.

§2.0 REFERENTIALISM AND BAYESIANISM IN DAVIDSON The first section of this paper devoted itself to demonstrating the motivations behind the

principle of charity, and its subsequent employment in overcoming the problem of speaker

knowledge, which previously threatened Davidson’s project. The coming section endeavours

to show that certain vulnerabilities are both implicit in Davidson’s Tarskian semantics, and

ingrained in his account of radical interpretation. These threaten to undermine the ‘extended

project’ insofar as they demonstrate a commitment to two positions ­ Bayesianism and

referentialism. Before detailing how commitment to these positions is open to attack from

Chalmers’ Puzzle, I will show how referentialism exposes itself in Davidson’s semantics, and

Bayesianism exposes itself in the principle of charity.

§2.1 REFERENTIALISM

According to Chalmers’ (2011)10 definition, referentialism is the view that insofar as beliefs

attribute properties to individuals, the objects of those beliefs must be determined by the

individuals and the properties attributed. According to a common strand of referentialism, the

14

120434

objects of beliefs are propositions ­ which in turn are expressed by sentences or utterances.

So each sentence, utterance or belief expresses a particular proposition, whose contents are

then determined by the objects and properties to which it refers. For example, the proposition

expressed by my belief or utterance 'Ben Nevis is a mountain' has contents which are wholly

determined by the reference of its constituent parts. Since 'Ben Nevis' refers to Ben Nevis (the

mountain), the propositional contents are constituted by that mountain. And since 'is a

mountain' refers to the property of being a mountain, the propositional content is wholly

determined by what satisfies the conditions of mountainhood. This might seem trivial, or

indeed implausible, but the alternative ­ that propositional content is determined wholly by

facts about our own internal mental states ­ comes with its own set of advantages and

disadvantages.

The referentialist treats propositions concerning probability in the same way. The

propositional content of my belief to degree 0.8 (henceforth: credence) that Boris Johnson will

be elected is wholly determined by the extension of terms 'Boris Johnson' and 'will be elected',

respectively, Boris Johnson himself, and the property of being elected. But how does the

credence 0.8 play into the propositional content?

One option is to make the credence degree 0.8 a part of the proposition, such that the belief

expresses the (colloquial) proposition 'the chance that Boris Johnson will be elected is 0.8'.

This is advantageous since the extension of a credence degree 0.8 will be the property of

having a probability of 0.8. And then the property of having a probability of 0.8 determines the

propositional content. Hence it is rightly left to the probability theorist to define the terms of

correct application of this ­ and other similar ­ probabilistic properties.

A more common view is that the credence of degree 0.8 is not part of the propositional

content of the belief. Rather, the credence is a psychological property which illustrates the

degree to which the agent would be willing to assert the proposition. As Ramsey (1926)11

suggested, we might be able to quantify this conviction by seeing working out the upper and

lower bounds of odds an agent would be willing to accept in an idealised betting situation.

Regardless of whether or not credence can be measured in this way, there is little doubt that it

is possible to have more conviction in some beliefs than in others; this should be our guiding

15

120434

intuition in this strand of epistemological discussion. I will pursue this in more depth in

sections to come.

For now, we return to the subject of referentialism: the opening passages of 'Truth and

Meaning' (1967) are suitably described as a rally against 0­dimensional, or extensional

theories of meaning. Davidson finds initially that such theories cannot deal with the

introduction of intensional contexts ­ such as those elicited by propositional attitude reports ­

since they yield a failure of the coextensivity principle. After finding that 1­dimensional

intensional theories of meaning also fail ­ albeit for slightly different reasons ­ the decision is

taken to abandon meanings as the meaning­giving theorems altogether.

Consequently, we are led down the path of a 0­dimensional theory of truth as providing the

meaning­giving theorems for every possible utterance in a natural language. This is presented

not as the only way to understand a language, but as just one way to get into the same

epistemic position as a native speaker. And it is clear that if Davidson is to appeal to a

0­dimensional theory of truth as underpinning his future semantics, this will need to be

defended once more against some classic puzzles for the extensionalist.

Interestingly, if we take a closer look at Davidson’s extensionalism, we see that it is entailed

by Chalmers’ referentialism. Hence if we hold the latter, we are implicitly committed to the

former. This is because referentialism relies on a conception of the extension of terms in order

to yield propositional content. In the example above, it is only the extension of 'Boris Johnson',

namely, Boris himself, that can give the belief a content. And no belief lacking content could

be meaningful. Hence 0­dimensional meaning relies essentially on a certain conception of an

extension.

However, it is not so simple to show that Davidson’s own extensionalism entails referentialism

of the kind Chalmers’ offers. I contend that arguments for Davidson’s extensionalism as a

form of referentialism draw blood precisely because of the two positions’ related attitudes to

reference and truth­conditions. On extensionalism at the level of Davidson’s semantics, a

T­sentence looks like this:

(4) s is T in L if and only if p

16

120434

Where ‘s’ refers to a structural description of a sentence in the object language; ‘is T’ refers to

the truth­predicate; ‘in L’ is replaced by the name of the object language, and ‘p’ refers to an

extensional description of the truth conditions of ‘s’ in the metalanguage (such that it is apt for

interpretation). This kind of strict extensional theory is advantageous ­ according to Davidson ­

since it does not necessarily involve the usage of any semantical tools beyond the apparently

indispensable ‘refers to’ (1967, p.305). The key feature for present purposes is that a

prerequisite of Davidson’s T­sentences yielding a useful result is that the unknowns in the

T­sentence formula successfully refer.

(5) 'Water is wet' is T in English if and only if water is wet.

Simply put, p’s replacement in (5), ‘water is wet’, must refer to water and the property of being

wet. In particular, the propositional content expressed by the truth­conditions ‘water is wet’

must be ­ in virtue of their being truth­conditions ­ denotative of particular states of affairs in

the world. That is to say, unless referentialist determining of propositional content figures in

Davidson’s workable T­sentence formulae, T­sentences cannot underpin a Davidsonian

semantics. The version of extensionalism in play here must entail the referentialist view that

propositional content is determined by the objects and properties referred to.

To put this into the form of an example: a central claim of extensional theories is that 'Ben

Nevis is a mountain' is true in virtue of whether the extension of 'is a mountain' includes the

extension of 'Ben Nevis'. The truth­value of the sentence varies from ‘true’ in the positive

case, to ‘false’ in the negative. Referentialists hold that the objects of belief ­ or what is

believed ­ is in some sense dependent on the objects referred to. Since extensionalist

accounts require a dependence for propositional content on the objects referred to, so they

require a referentialist assumption.

Whilst the principle of charity does provides a satisfactory answer to the problem of speaker

knowledge, thus allowing radical interpretation to go through, this comes at a cost. Since

radical interpretation relies on an extensionalism, so it relies on a referentialism. The

important consequence for the extended project is that Davidson needs referentialism in order

to ground T­sentences as meaning­giving, T­sentences in order to ground radical

17

120434

interpretation, and radical interpretation in order to ground charity. Each of these is required in

order to fulfil the extended project. This staking in referentialism is the first tenet in the paper’s

core argument, and can be represented thus:

(A) The principle of charity enables Davidson’s semantics to have an empirical

application by solving the problem of speaker knowledge. But this solution necessarily

involves a commitment to referentialism. Hence Davidson’s project is staked upon a

commitment to referentialism.

§2.2 BAYESIANISM

The next section deals with a second consequence ­ Bayesianism ­ which we conjecture is

brought about by Davidson commitment to the principle of charity. In particular, the

consequence is a commitment to a Bayesian decision theory as central to a conception of

rationality. On a Bayesian decision theory, certain laws of probability theory have a role to

play in rationality in that they provide normative constraints on belief. In general this leads to

Bayesian decision theorists cashing out warranted belief, or indeed correctness of belief, in

terms of the probabilities assigned to individual outcomes as per the experience of an

individual subject. Such an emphasis on an agent’s perceived probability of outcomes as a

fundamental kind of evidence has led to Bayesian decision theory being widely taken to invite

epistemological subjectivism.

Putting aside questions of subjectivism, Bayesians differ with respect to which laws of

probability they take to provide normative constraints on rational belief. Further, they take the

laws of probability to constrain not the contents of beliefs, but the degrees of belief (or

credences) in particular propositions. The least controversial of these constraints are the

probability axioms, most famously propounded by Kolmogorov (1933)12. But we will also

consider some others, including conditionalisation.

18

120434

First we will sketch a Bayesian decision theory in sufficient detail that we might extract a few

key commitments. In order to do this, we will draw on Strevens (2006)13 in order to present the

Bayesian position in just enough detail to extract its key commitments. Then we will

investigate the extent to which Davidson’s work on the principle of charity might be seen to

encapsulate the Bayesian position.

Strevens (2006) sets out three basic principles which underpin the Bayesian decision theory.

The first of these is that the subject assigns credences or subjective probabilities to

propositions, outcomes or events. Essentially, these credences constitute the degree to which

the agent believes the various propositions to be true ­ and are represented by a real number

which is greater than zero (total lack of credence) and less than one (certainty). One

important thing to note is that not only are they psychologically real in that they describe an

attitude of ‘holding true’, they also map subjects to propositions (or outcomes) via a relation of

degree. This stands in progressive opposition to classical theories of confirmation, which

typically attribute relations between subjects and propositions relative to three possible

outcomes: asserts, denies or neither.

The mathematical properties which the Bayesian attributes to credences bear intuitive

resemblance to probability notation. However, given that credences are psychological objects,

as opposed to mathematical, problems arise concerning how these are to be measured or

quantified in psychological subjects. As we mentioned, Ramsey (1926) developed a model

according to which the worst odds at which an individual accepts a bet on an outcome

constitutes the value of their credence. For example, suppose that I think there is a good

chance that it will rain tomorrow, having seen the weather forecast. A hypothetical method by

which I can assign value to my credence might involve an interlocutor asking me if I would be

willing to accept odds pertaining to receiving £15 if it rains or giving away £10 if it does not.

On each dismissal of the odds he offers, the interlocutor gradually alters his position until I

accept. Plausibly, if after some number of offers, I accept receiving £10 if it rains and paying

£10 if it does not, then he has discovered that I possess a credence in the proposition that it

will rain of approximately 0.5.

Ramsey’s approach may be flawed in multiple ways, not limited to problems involving

idealised betting situations, aversions to betting and degrees of accuracy, but these are not

19

120434

areas this paper will explicitly pursue. However, it is important to take stock of the distinction

between probability at the level of mathematics, as opposed to credences which occur at the

level of psychology. Although the two may appear to behave in the same way, equating them

is ­ at least prima facie ­ a dangerous and dogmatic path to pluralism. For now, we can take

Strevens’ first principle thus: that Bayesianism distinguishes between credences (at the

psychological level) and probabilities (at the mathematical level).

The second of Strevens principles of Bayesianism is a development of his distinction between

credences and probabilities. In particular, both probabilities and credences appear to conform

to probability axioms ­ such as those sets developed by Kolmogorov (1933) and Savage

(1954)14. Before we can begin a suitable explication of these axioms, some definitional

groundwork is in order. In Kolmogorov’s 'Foundations of the Theory of Probability', he defined

the terms of the debate such that the probability value P, of some event E, takes place within

probability space (Ω, F, p). This probability space in turn is composed of sample space Ω (the

set of all possible outcomes); event space F (a space in which a set composed of more than

zero outcomes may occur); and p (a probability function which maps possible events to

probabilities).

Given this framework for the mathematical study of probabilities, we can express the axioms

as constraint upon the value of p. The first axiom states that every possible outcome within

probability space must be mapped by the probability function to a real number between 0 and

1. At root, this is a mere notational point. The second axiom states that any outcome which is

logically certain, or unavoidable, must be expressed in the notation as p(E) = 1. Conversely,

any event which is logically impossible must be represented as p(E) = 0. Thus the rest of the

set of possible probability values represents degrees of possibility within probability space.

The third axiom states that if outcomes are mutually exclusive (such as in the case of a single

(ideal) coin toss ­ it necessarily lands either heads or tails), then p(E v D) = p(E) + p(D).

With the three axioms of probability in hand, Strevens introduces two further definitions. The

first of these is the notion of a conditional probability ­ specifically, the probability of a

particular outcome condition on another’s occurrence. The probability of E given that D is

expressed in the notation as follows:

20

120434

p(E|D) = p(ED) / p(D)

We can use the definition in the following way. If a university student estimates that his

chances of getting into University are 0.5, then we can represent this outcome, D, as p(D) =

0.5. Further, if for every student gaining a place, there is a 0.4 chance of their gaining

University accommodation, then we can represent the probability of this event, E, as p(E) =

0.4. What is the probability of gaining University accommodation conditional on gaining a

University place? Or in other words, what is the value of p(E|D)?

p(E|D) = ((0.4)*(0.5)) / p(0.5) = 0.4

Hence the conditional probability of E given D is 0.4. Contrast this with the more conventional

unconditional probabilities that these same events will occur, where unconditional probability

is informally defined as the probability an event will occur regardless of evidential factors. The

unconditional probability of outcome E, that the student will get into University

accommodation, is identical to the probability of (Getting into University AND gaining

accommodation), which is computed by p(E*D) or p(0.4*0.5) = 0.2. Hence we can say that

outcome E has an unconditional probability of 0.2 and a conditional probability (on E) of 0.4.

But what difference should the above observations make to Philosophy? Mathematicians

differ with respect to the precise details of conditional probability, Bayes’ Theorem, and to a

lesser degree, the probability axioms. But where the above principles become interesting is

where they are said to constrain rational belief. This is the intuition enshrined in the principle

of conditionalisation, which is characteristic of Bayesian position. On this epistemic principle,

the laws of probability ­ including both the axioms and the definitions of conditional probability

­ constrain rational thought. In particular, the definition of conditional probability governs the

way in which we ought to update our beliefs in light of new evidence. This brings us to

Strevens’ second definition.

The principle of conditionalisation begins with the mathematical formulae above, but pitches

instead at the level of epistemology. Hence the first stage in the conditionalisation position is

that rational agents hold both conditional and unconditional beliefs in propositions. The

student, for example, has both unconditional beliefs p(D) = 0.5 that he will get into University,

21

120434

and p(E) = 0.2 that he will gain student accommodation. However, the student also has

conditional beliefs that if he successfully gains entrance into University, then the probability of

his successfully gaining student accommodation is increased, to p(E) = 0.4.

According to the principle of conditionalisation, the student’s unconditional credence p(D),

once he has acquired total evidence p(E) = 1, should be updated so as to match his previous

conditional credence in p(D|E). The same goes for all new unconditional credences. On

acquisition of a piece of total evidence E, such that p(E) = 1, unconditional credences should

be updated to match the value of any and all credences previously conditional upon E. Put

simply, one’s belief in a proposition were an event to occur, should match one’s new belief in

the proposition once the event has occurred. This is the colloquial definition of

conditionalisation going forward.

§2.3 BAYESIANISM IN RADICAL INTERPRETATION

Zilhão (2003)15 and Rescorla (2013)16 have argued that Donald Davidson’s radical

interpretation is grounded in a version of Bayesian decision theory, and is therefore

necessarily committed to the principle of conditionalisation. According to Rescorla, we need to

look closely at Davidson’s application of the principle of charity in order to see this.

Rescorla (2013) cites three themes in Davidson’s work on radical interpretation which suggest

it is grounded in a Bayesian decision theory. The first of these is the principle of charity. On

the principle of charity, in particular, the principle of coherence, an interpreter should hold that

native speakers conform to basic norms of rationality. Rescorla (2013, p.) notes LePore and

Ludwig’s emphasis that Davidson defines charity as finding enough rationality in others that

we assume they exhibit ‘norms of logical consistency, of action in reasonable accord with

essential or basic interests, and the acceptance of views that are sensible in the light of

evidence’ (2005, p.319)17 This he takes to constitute evidence in favour of the claim that the

principle of charity commits Davidson to a version of Bayesian decision theory.

22

120434

Rescorla also points out that Davidson reminds us in his (1973) paper ‘Radical Interpretation’

that if a native applies high probability to a sentence, the principle of charity then entitles us to

take this as evidence that the sentence is true. Insofar as we take charity literally, we can

derive the claim in Davidson that a native’s assigning high probability to a sentence

constitutes evidence, but not total evidence for truth of the utterance. In particular, our

ascription of particular norms to subjects should ‘track genuines relations of evidential

support’ ­ such as those seen on a Bayesian decision theory. In fact, this requirement is a

version of probabilism ­ which states in the broadest terms that thought conforms to the basic

probability axioms. A common derivation is that probabilism then entails the conditionalisation

principle on rational thought, since the probability axioms entail it as a constraint on

mathematical probabilities.

The final theme is that the model described above enshrines the constitutive ideal of

rationality ­ central to all intentional ascription. The constitutive ideal makes explicit a range of

normative claims. And given that these claims act as essential enablers for the possibility of

ascribing intentionality to creatures ­ assuming them can be seen as fundamental to

interpretation. The consequence is that without ascribing our own norms to all communicative

creatures, we cannot hope to interpret them. As Rescorla (2013, p.8) puts it, ‘We can attribute

intentionality to a creature only if we treat the creature as largely conforming to rational norms.

Which rational norms? Our own, because those are the only ones we have.’ Since an agent

need not always satisfy rational norms, provided that they are in general rational, it follows

that we must act in interpretation as if an agent conforms to norms (Bayesian and otherwise)

insofar as we do ourselves.

Based on the above analysis, I take Bayesianism to be (minimally) committed to the following

core claims. First, that there is a distinction between the probability of an event at the

empirical level, and credence in a proposition at the psychological level. Second, that

probability axioms outlined in Strevens (2012) constraint the value of probability function p.

Third, that a principle of conditionalisation establishes a normative constraint on rationality. In

particular, this tells us that our credence in a proposition p conditional on evidence E before

we have E, should be updated to match our new unconditional credence in p once E has

been acquired. Hence the Bayesian project consists in two broad projects: to explain a

descriptive mathematical theory of probability and evidence, and establishing that this

23

120434

descriptive theory might provide normative guidance. Rescorla (2013) and Zilhão (2003) have

given us strong reasons to suggest that Davidson endorses these claims, which form premise

(B).

(B) Radical interpretation is grounded in ­ and hence reliant on ­ a Bayesian decision

theory. In particular, the reliance is on a Bayesian principle of conditionalisation.

However, as we have shown, radical interpretation is required in order to give his

semantics an empirical application. Hence Davidson is in no position to give up radical

interpretation, or the principle of conditionalisation.

§2.4 CHALMERS’ PUZZLE

In his (2011) paper, Chalmers presents a version of Frege’s Puzzle along probabilistic, rather

than semantic lines. Where the original formulation tended to emphasise the semantics of

belief ascriptions as the source of the trouble, the probabilistic version focuses on a certain

conception of decision­making, namely, Bayesian decision theory, in order to generate the

inconsistency. Purportedly, there are a number of advantages to framing the puzzle in these

terms. For example, it is now open to the Bayesian to defend herself using the new­found

theoretical explanatory power probabilistic­orientated discussion granted. Since the puzzle

shows how Bayesian explanation can function at the psychological level of credence, this

provides a presumption against the referentialist. The objects of utterances also become less

important than objects of credence (strength of belief, conviction), which need to play a

particular kind of role in order to figure in an adequate theory of credence. This shift

encourages a kind of debate not previously seen in the debate against referentialism.

Chalmers takes referentialism to be (minimally) committed to the following views. First, insofar

as beliefs are said to concern the world, that is to say, insofar as beliefs can be framed as

attributing properties to objects, the content of those beliefs must be determined (at least in

part) by those individuals and properties. Despite trivial appearances, this claim is of central

importance to a semantics of belief ascriptions, since it entails that when I have a belief, say

that Bertrand Russell is dead, the content of my belief is decided by the objective object of

Bertrand Russell himself, and the objective property of being dead. Incidentally, these are the

commitments of Russell’s own view of propositions ­ which states that the objects designated

24

120434

in a sentence are literal constituents of proposition expressed by the sentence. That is to say,

my utterance 'Bertrand Russell is dead' contains as constituent part Bertrand Russell, and the

property of being dead.

One consequence of the referentialist view about semantics is that provided A and B are both

name the same object, it must be possible to switch them salva veritate (preserving truth)

within containing sentences.

(1) Hesperus is a planet.

(1*) Phosphorus is a planet.

Since ‘Hesperus’ and ‘Phosphorus’ both name the planet Venus, we should be able to (and

indeed can) switch them without altering the truth value of the sentence. In the example

above, both sentences are true, regardless of which name we use. The contemporary debate

in favour of referentialism has tended to focus on cases like the below ­ involving belief

ascriptions ­ since substituting coreferential terms in such contexts has proved a more difficult

business.

(2) Peter believes that Fatboy Slim is a musician.

(2*) Peter believes that Norman Cook is a musician.

Since 'Fatboy Slim’ and ‘Norman Cook’ both name the same individual, the referentialist

would expect that we would be able to switch the terms without altering the truth­value of the

sentence overall. However, in belief contexts, it is fairly simple to imagine situations where

sentence (2) is true, but (2*) false (or vice versa). For example, I (Peter) might have met a

man named ‘Norman Cook’ at a local bar playing a secret concert but never realised that his

public alias is ‘Fatboy Slim’. Then (2*) would be true description of the state of affairs, but (2)

not. Conversely, I might have been to a ‘Fatboy Slim’ concert years previously, but so far

away from both performance and screen that I failed to identify that the man called ‘Fatboy

Slim’ is actually Norman from the bar down the road. In that case (2) would be a fitting

description of my belief state, less (2*). The important point is that when belief­ascriptions are

introduced, the referentialist maxim that coreferential terms must be substitutable salva

veritate appears to fail.

25

120434

The above version of Frege’s Puzzle is familiar, and referentialist responses well­rehearsed.

But slightly different intuitions appear to govern our reactions to the problem when we frame

the issue in terms of probability. Chalmers invites us to consider a case in which genetic

scientist Olivia is conducting clinical research on the genetic basis of a disease, which

displays no symptoms until after the death of the carrier. The clinical importance of her study

is secured with prior research which has shown that the disease might diagnosed by the

presence of particular genetic structures ­ which we will call gene A and gene B. Olivia tests

for these A and B in an attempt to assess the role these two genes play with respect to

indicating the presence of the disease D. Her research methodology is that she assesses, for

a given subject, which of gene A, gene B and disease D they possess. With this statistical

basis, she can then derive claims about the percentage likelihood of an individual having the

disease, based on their genetic makeup. Prior research in the area has shown that subjects

with gene A have a 10% chance of having the disease, and those with gene B have a 20%

chance of having the disease. However, participants who exhibit both gene A and B have a

90% chance of having the disease.

Olivia believes that there are sixty participants in the study; in fact, there are only fifty­nine.

Such appearances are explained by the fact that she is being deceived by one individual, who

checks in as Dr. Jekyll on Tuesday morning, and then again as Mr. Hyde later that afternoon.

At no point in the study is Olivia aware that Dr. Jekyll and Mr. Hyde are one and the same

person. For a given subject, we can indicate their having gene A, gene B or the disease D by

respectively affixing A, B or D to their initials. Thus, for Dr. Jekyll, a belief in his having A, B or

D can be represented as follows: JA, JB, JD. The analogue for Mr. Hyde is then: HA, HB, HD.

On Tuesday evening, Olivia examines Dr. Jekyll’s A swab, and Mr Hyde’s B swab (amongst

other things), finding evidence that JA and HB. Given that JA, she forms both an

unconditional credence that p(JD) = 0.1, and a conditional credence that p(JD|JB) = 0.9. The

same goes for Mr. Hyde’s B swab. Since the probability of a B­gene owner having the disease

is twenty percent, the unconditional probability p(HD) = 0.2. However, the conditional

probability of HD based on evidence HA is represented p(HD|HA) = 0.9.

26

120434

In the morning examination, Olivia acquires evidence JA, and in the evening, she acquires

HB. For now, we will focus on HB. If referentialism is true, then given that Dr. Jekyll and Mr.

Hyde are the same person, the proposition that HB is the same proposition that JB. Hence to

acquire total evidence that HB and JA, is to acquire total evidence HB, JB, JA, HA. So the two

pairs HB and JB, as well as HA and JA, should be substitutable salva veritate within

containing probabilistic propositions.

Recall the two sets of conditional probabilities which we derived from Olivia’s morning and

evening examination, p(JD|JB) = 0.9 and p(HD|HA) = 0.9. Now that evidence JB and HA has

been acquired, Olivia should be in a position to update these credences, in accordance with

the principle of conditionalisation. However, charges Chalmers, Olivia is in no such position,

since she is unaware that she possesses evidence JB and HA, in virtue of her being unaware

that Dr. Jekyll and Mr. Hyde are identical. Hence Olivia ­ though rational ­ is ill­informed to

update her credences in accordance with rational norms.

Intuitively, Olivia’s circumstances are such that her beliefs in p(JD) = 0.1, and p(HD) = 0.2.

This is because although she acquires JA and HB, since she is not in the epistemic position

required in order to know that JA and HB are identical with HA and JB respectively ­ nor

should she be. Referentialism fails to acknowledge ignorance of identities across evidential

implications, and so generates an implausible result when combined with Bayesian decision

theory. Hence the paper’s core premise (C):

(C) According to Chalmers, Bayesianism is inconsistent with referentialism since when

the two views are combined, they yield counterintuitive results on probabilistic versions

of Frege’s Puzzle.

§3.0 RATIONALITY AS CONSTITUTIVE IDEAL

In the final sections of this paper, I provide an argument to the effect that Chalmers’ Puzzle

does not draw blood from Davidson’s ‘extended project’ because it only applies to versions of

Bayesianism which are descriptive. In order to establish that Davidson’s project need not

necessarily encapsulate a descriptive element, I extend our previous analysis of Zilhão in

order to accommodate his argument that Davidson’s descriptive conception of rational agency

27

120434

is incompatible with truth as bridging the gap between mind and world. Since Davidson’s

concept of truth bridging the gap between mind and world satisfactorily answers the problem

of speaker knowledge, we contend that we should accept the idea that Davidson’s descriptive

conception of rational agency does not go through. Hence provided we can construe a

version of rationality as constitutive, but not descriptive, Davidson may avoid the objection

from Chalmers’ Puzzle. The closing sections detail precisely how Chalmers’ Puzzle might be

sidestepped.

§3.1 CONSTITUTIVE, NORMATIVE, DESCRIPTIVE

As we have seen, Davidson’s conception of rationality is that it is constitutive of thought. Just

as our conforming to basic levels of rationality is a prerequisite for our having thoughts at all,

so it is a prerequisite for any creature’s having thoughts or intentionality. Radical interpretation

is supposed to enshrine the idea that normative constraints necessarily govern all possible

intentional ascription. (2004, p. 128)17. We have seen that it does this by showing how

interpretation is only made possible by assuming our own set of rational norms in other

creatures. And necessarily, given that we have no translation scheme other than our own,

creatures must conform to our own set of rational norms, since these are all we have. For

Davidson, propositional attitudes are ordered logically and systematically, and hence must be

partly constituted by such a system such that ‘the satisfaction of conditions of consistency and

rational coherence may be viewed as constitutive of the range of application of such concepts

as those of belief, desire, intention and action’ (1980, p.237)18. Not only does this guarantee

the possibility of communication, it also legitimises ­ as per charity ­ the reading in of Bayesian

decision theoretic interpretations into the very structure the speech and behaviour of other

agents. In an important sense, Bayesian norms ­ or at least our attribution of them ­ constitute

the conditions of the possibility of interpretation.

Rescorla, on the first page of his paper ‘Rationality as Constitutively Ideal’ (2013), outlines

three general conceptions of rationality: descriptive, normative and constitutive, which we are

then shown are reflected in the Philosophy of Kant, Carnap, Quine and Davidson. In general

terms, these three positions can be described as follows. On a descriptive conception, logic

(here used interchangeably with ‘rational thought’) describes the ways in which human

subjects actually think, and as such might be confirmed or denied on the basis of empirical

28

120434

evidence. This differs from a normative conception insofar as logic here describes a system of

thought and actions to which human beings ought to live up to. In a certain sense the

normative conception is also descriptive, albeit in a slightly different way; it describes a

system which we should act in accordance with. Finally, the constitutive conception states

that logic or rationality fundamentally informs what it is to think ­ that is to say, logic neither

fundamentally describes nor prescribes the structure of human thought. Rather, acting in

accordance with rational norms is a process which enables human intentionality to take place;

Thus rationality is not just a prerequisite for interpretation, it also enables its possibility.

According to Rescorla’s own analysis, rationality as constitutive ideal entails the normative

reading. This is because insofar as rationality as constitutive ideal enables conformity to

Bayesian norms, it also prescribes their use ­ at least insofar as we want to attribute thoughts,

beliefs, desires and actions unto ourselves. However, it does not entail the descriptive

conception, since this constrains actual human thoughts ­ which are not necessarily bound

into accordance by psychophysical laws.

§3.2 THE GAP BETWEEN MIND AND WORLD

Zilhão (2003) has argued that the Bayesian conception of human agency is incompatible with

the idea that truth provides the link between mind and world. Since we have already

described the Bayesian conception, we now provide an analysis of truth as providing the link

between mind and world. The claim made by Zilhão is that empirical semantics of the kind

pursued by Quine and Davidson must satisfy two separate conditions: (1) there must be an

essential link connecting environment with action; (2) there must be a behavioural core

common to both interpreter and speaker. According to Zilhão, Davidson satisfies the first

condition by means of an appeal to intersubjective truth. In particular, Davidson is said to use

an intersubjective notion of truth to connect environment with action, ‘the truth conditions of

any of our utterances do belong to this public world from the outset’ (2003, p.231). This

intersubjective truth allows us to move from environmental stimulus ­ via the principle of

charity ­ towards T­sentences as the building blocks of an interpretation scheme.

Our framing of the problem of speaker knowledge had previously framed the problem as one

of moving between environmental stimulus to a radical interpretation T­sentence, and then

29

120434

from that radical interpretation T­sentence, to the kind required in order to form part of a

Tarskian semantics. The kind of T­sentence gathered on the ground looks like this:

(6) 'Llueve' is ‘held true’ in L if and only if it is raining.

Whereas the kind of T­sentence required in order to form the basis of a Tarskian theory of

truth looks ­ in template form ­ like this:

(4) s is T in L if and only if p

The problem is that (6) contains ‘held true’, whereas (4) contains ‘is T’ ­ which in effect

amounts to ‘is true’ provided we satisfy Morris’ expression of Tarski’s Convention T. The

problem was that it is not obvious how to safely move from ‘is ‘held true’’ [by a speech

community] to ‘is true’ [according to a definition of truth]. However, Zilhão has given us strong

reason to believe that an intersubjective reading of ‘is T’ in (4) enables us to make these

coextensive. The idea that truth­conditions ‘belong to this public world from the outset’ (2003,

p.231) in effect supposes that truth­conditions as determined intersubjectively. This is similarly

enshrined in Davidson’s own writings, where he proclaims that an agent ­ simply in his

belonging to a speech community ­ has mostly true beliefs:

(GE) (x)(t) (if x belongs to the German speech community then (x holds true ‘Es

regnet’ at t if and only if it is raining near x at t) (1973, p.315).

According to Zilhão, this conception of intersubjective truth is incompatible with Davidson’s

theory of rationality, which is effectively the one found in a Bayesian decision theory. On a

Bayesian decision theory, presupposed is a framework according to which we have (i) beliefs

­ which are truth­evaluable attitudes towards states of the world, assumed relative to desired

outcomes (2003, p.233); and (ii) desires ­ ‘pro­attitudes towards representations of outcomes

of actions’ (2003, p.233). Credences represent the real­numerical degree to which different

outcomes are considered possible by the agent. We can use these credences to construct a

principle of ‘maximal expected utility’ according to different outcomes and desires. Essentially,

this principle nominates, for any set of beliefs, desires and credences, a course of action

30

120434

which optimal according to the circumstances. On a Bayesian decision theory, the ‘maximal

expected utility’ principle nominates the decision Bayesian agents actually make.

However, as with any theoretical principle which purportedly governs human action ­ in virtue

of also being part of experience ­ it should be confirmed or denied on the basis of empirical

evidence. The problem with the principle of ‘maximal expected utility’ is that it is not so

empirically verifiable, even in principle. One is faced with a major problem ­ since

Bayesianism assumes the system of beliefs, desires and utility in its initial statement ­ one

cannot empirically confirm or deny the existence of such a system except relative to it. That is,

it is not possible to empirically test, since to empirically test involves assuming its truth from

the outset.

Further, Zilhão contends that when one looks closer, the Bayesian decision theory is based

on a description of the ideal behaviour of a perfectly rational gambler ­ in accordance with the

axioms of probability. This is the idea we touched on in our discussion of Ramsey (1926) in

section §2. The claim that needs to be empirically verified is then ­ does the behaviour of

actual human agents conform on any analogous level to the behaviour of the perfectly rational

gambler in the idealised betting situation? In order to affirm this, agents need to act according

to the probability axioms not just in an idealised game situation, but in most (or all) aspects of

their daily lives.

§3.3 THE PROBABILITY AXIOMS AS CONSTITUTIVE, NORMATIVE.

How far do the axioms prescribed for use in an ideal game situation inform our daily decisive

behaviour? Zilhão suggests that his axioms A and B face (respective) and devastating

problems. According to Axiom A, a relation of ‘is at least as preferred as’ holds between any

two coordinates open to an agent’s consideration. This is a binary relation which satisfies

transitivity, reflexivity and connectedness and allows us to place any or all outcome options

open to us in a linear scale with an assignment of real­numerical values to each. Tversky

(1969)19 undermines axiom A by showing that in practice, preference patterns do not satisfy

the constraints on axiom A, since more often than not psychological studies have shown

preferences ranked in the way described above do not exhibit transitivity. At worst, these

31

120434

empirical observations suggest that agents do not conform to the basic probability axioms

most of the time.

Axiom B drives us to the same result, albeit due to a different problem. According to axiom B,

given outcomes A, B, C and D. If A is at least as preferred as B, and C and D result

respectively from the same change in their common outcomes, then C must be at least as

preferred as D. These are enshrined in ­ and hence justified by ­ the following decision

strategy. If an agent is considering the most advantageous outcome to choose, the agent

should consider the mutually exclusive and jointly exhaustive set of events within the

probability space as his options. Hence possible states with the same outcome arising from

those options available to him should not be considered twice. However, Axiom B has again

been shown to be false by Allais’ problem (1953)20. On empirical tests of Allais’ problem,

subjects are given two decision situations each involving two gambles. In choice one, agents

are given the opportunity to choose between winning a certain large amount of money

outright, as opposed to gambling that same amount of money against a small probability of

winning five times that amount, and an even smaller probability of winning nothing. In the

second choice, agents must choose between two gambles where they are most likely to win

nothing. Importantly, in the second choice, agents choose the gamble in which the highest

prize is at stake, despite the fact that the chance of winning is less than in the first choice of

gamble. Hence axiom B is not warranted to have status as a descriptive axiom of actual

gambling behaviour in human agents.

Quite clearly, these gambling situations provide strong reasons against the view that the

axioms of probability constitute descriptions of actual decision­theoretic behaviour in game

play, even under idealised conditions. And equally clearly, Davidson’s conception of rationality

as constitutively idea dispels the presumption that any such descriptive project might be

possible. Rather, the view favoured is one where Axiom A and Axiom B describe behaviours

the assumption of which makes possible communication and interpretation in the radical

interpretation situation. Without the assumption that agents do conform for the most part to

Axiom A and B, interpretation would not be possible. Thinking of rationality as constitutively

ideal in this way essentially absorbs the objection ­ agents’ acting out of step with the rational

norms under idealised circumstances is not objectionable in itself. All that is required is that

32

120434

agents conform most of the time, and that we are in a position where charity permits us to

assume this.

§3.4 CHALMERS’ PUZZLE REVISITED

The discussion in this chapter so far has aimed to demonstrate beyond reasonable doubt that

Davidson’s position is dependent on the principle of charity, enshrined in the constitutive ideal

of rationality, in order to complete his ‘extended project’. A corollary of this discussion is that

the constitutive conception ­ in virtue of rationality being a prerequisite for interpretation ­ need

not entail a descriptive conception of rationality. A constitutive conception aims to show that

an agent need only assume that other agents are by and large rational in order to interpret

them. This does not require agents to conform in a rigid way to rational descriptions; a useful

feature of an account given that this requirement has been repeatedly empirically

undermined.

Davidson’s extensional or 1­dimensional theory of truth commits him to a referentialism about

the objects of propositions. This is of fundamental importance ­ since referentialism about the

objects of propositions seems to commit one to referentialism about the objects of

propositions concerning probability. Further, since Davidson’s holistic constraint entails that

finding meaning and interpreting beliefs are part of one and the same project, he seems also

to be committed to referentialism about the objects of belief.

Chalmers’ referentialism ­ as we earlier explained ­ is a referentialism which concerns the

objects of credence. What are the objects of credence? These are the entities to which

credences are rightly assigned. For example, if I have a credence p(X) = 0.75, what is it that I

have a credence to degree 0.75 in? Conventional ­ or Russellian ­ propositions, according to

Chalmers’, need to meet certain criteria in order to do the job required such that a theory of

the objects of credence might be successful. Hence referentialism about credence is best

viewed as a constraint on the nature of the Russellian proposition. The constraint is this:

where an agent has a credence which asserts something about some object or property, the

objects of credence must be determined wholly by those objects and properties in which the

credence is had.

33

120434

What about the Chalmers’ Bayesianism? Recall that Chalmers’ definition of Bayesianism is

cast at the level of conditional and unconditional credences in propositions. His conditions on

these conditional and unconditional credences is that they must be updated in light of new

evidence in precisely the manner designated by the principle of conditionalisation ­ which

states that one’s previous credence in X conditional on acquisition of total evidence E should

match the new unconditional credence once E has successfully acquired X.

With the two definitions in hand, it is fairly straightforward for Chalmers’ to build his example

concerning Olivia. Olivia is a supposedly ‘rational’ agent who forms various conditional and

unconditional beliefs based on the hypothetical and actual acquisition of certain pieces of total

evidence. However, since she is unaware of an identity between two potential sources of

evidence ­ Dr. Jekyll and Mr. Hyde ­ once she has actually acquired total evidence E, she

does not realise her privileged epistemic position. Since Olivia has been deceived into

thinking that she actually has two smaller pieces of the evidential puzzle, she fails to update

her beliefs on acquisition of E in the way prescribed by the principle of conditionalisation. The

issue arises because referentialism is too strict a constraint on acquisition of evidence. In

particular, in failing to account for the fact that evidence might come into contact with agents,

and yet those agents be aware of its status as evidence, leads to these kinds of confusions.

Ignorance of identities plays a crucial role in Chalmers’ Puzzle, just as it does in Frege’s.

Davidson’s referentialism, despite appearances to the contrary, does not endorse the kind of

referentialism required in Chalmers. The opening passages in his (1967) paper ‘Truth and

Meaning’ appear to shift away from referential accounts of meaning with swift conviction.

However, the shift to meanings as truth­conditions does employ a referential account as a

basis for his Tarskian formal semantics. What then, are the so­called objects of truth, or those

entities to which truth­conditions are correctly assigned? And does this entail a view that

those same objects are that to which credences are correctly assigned?

In my solution to the problem of speaker knowledge, I suggested that Davidson faced a

problem in making coextensive the ‘held true’ predicate from radical interpretation with the ‘is

true’ predicate needed by a Tarskian semantics. In the third section of the paper, I outlined a

response according to which the two predicates might be made coextensive by appeal to truth

as grounded in the intersubjective ­ the same claim made in Zilhão (2003). As Davidson

34

120434

affirms in his (1973, p.315) statement of (GE), simply belonging to a speech community

allows us to afford the predicate the status of truth in a T­sentence. To revisit the earlier

example:

(6) 'Llueve' is ‘held true’ in L if and only if it is raining.

Suppose Davidson’s empirically­gathered T­sentences are referential in the sense required by

Chalmers. In that case, the ‘it is raining’ on the right hand side of the material biconditional

must refer to a state of affairs in the world. As such, the truth of (6) as a whole must depend

upon the actual status of the objects and properties ascribed. Given that referentialism is then

rendered true with respect to truth, and given that Davidsonian truth is all that there is to

meaning, this referentialism must then be said to be true concerning the propositional content

of the sentence.

However, on the intersubjective solution apprehended in Zilhão (2003), the sentence to the

right hand side of the material biconditional ‘it is raining’ depends for its truth on the actual

status of the objects and properties ascribed only insofar as these are intersubjectively

considered to be as they are according to members of the speech community. Given that in

the crucial case, referentialism about the objects of propositions does not hold, Chalmers’

objection does not go through.

As we have noted, Davidson’s Bayesianism does not extend to correct descriptive practice.

This is because the constitutive ideal of rationality entails only a normative theory of decision

as central to the possibility of radical interpretation. And empirical tests have consistently

provided plausible actual and possible counterexamples to descriptive Bayesianism.

However, Chalmers’ argument that Bayesianism is incompatible with referentialism relies on a

version of Bayesianism which requires the descriptive variant.

The counterintuitive result for Olivia’s thought experiment is generated by constraining

Bayesian rationality to consist in the principle of conditionalisation ­ updating one’s conditional

and unconditional credences to reflect the acquisition of new evidence. Referentialism

generates a problem in its failure to differentiate between acquiring new evidence, and

acquiring information (which we are unaware is evidence). The principle of conditionalisation

35

120434

then fails since the referentialist position fails to capture the nature of the evidence in play. But

the principle of conditionalisation only fails since referentialism passes it the wrong result.

Given that Davidson need not be committed to the misguided conception of referentialist

propositions, Bayesianism generates the right result. Further, given that rationality is

constitutive, not descriptive, the results of Olivia’s experiment are not counterintuitive. Now we

are in a position to state the present paper’s core argument:

(A) The principle of charity enables Davidson’s semantics to have an empirical

application by solving the problem of speaker knowledge. But this solution necessarily

involves a commitment to referentialism. Hence Davidson’s project is staked upon a

commitment to referentialism.

(B) Radical interpretation is grounded in ­ and hence reliant on ­ a Bayesian decision

theory. However, as we have shown, radical interpretation is required in order to give

his semantics an empirical application. Hence Davidson is in no position to give up

radical interpretation, nor a Bayesian decision theory.

(C) According to Chalmers, Bayesianism is inconsistent with referentialism since when

the two views are combined, they yield counterintuitive results on probabilistic versions

of Frege’s Puzzle.

(D) Though Davidson’s theory is extensional in that the T­sentences in radical

interpretation nominate truth­conditions, they are not referential ­ as Chalmers’

requires ­ since these truth­conditions are not relative to the objects and properties

ascribed. Rather, their extensions are determined by the speech community that the

speaker is a part of.

(E) Though Davidson appears committed to Bayesianism; he is uncommitted to any

kind of descriptive Bayesian project. That is to say, the constitutive ideal of rationality

entails that rationality is normative, but not that all (actual) human agents behave

rationally.

36

120434

(F) Chalmers’ project only applies to descriptive Bayesianism because it fails to

establish that credences conform to the axioms of probability. Without this connecting

point, claims regarding normative Bayesianism are unwarranted.

§3.5 CONCLUDING REMARKS

For all the lengthy discussion, this paper’s present point is rather small: although Davidson’s

constitutive ideal of rationality commits him to a certain kind of Bayesian decision theory, it

need not necessarily commit him to a descriptive reading. And despite Davidson’s claim that

rationality describes a constitutive system, which acts as an enabling feature for

interpretation, rationality as constitutive ideal seems to fail to tell us anything descriptive about

how rational agents actually think. Indeed, Tversky’s experiments and Allais’ problem seem to

establish something of a presumption against this idea. Although a constitutive conception of

rationality describes the status of Bayesian decision­theoretic norms in rational inquiry, it fails

to answer the empirical question of how far actual interpreters conform.

Therefore, the constitutive ideal of rationality is fundamentally non­descriptive of the actual

epistemic character of an individual’s credences. So the objection from Chalmers’, which

requires only that individuals are on that theory correctly described as exhibiting conditional

and unconditional credences which are updated accordance with the conditionalisation

principle, is misguided. Where Chalmers’ Puzzle seizes upon Bayesianism as describing

actual practices embedded in the thought and action of agents, our reading of the constitutive

ideal shows how Davidson may be convicted of no such illegitimate practice.

The problem of speaker knowledge had previously looked to be devastating since it relied on

a contaminated piece of machinery ­ the principle of charity. This was essential to our

endeavours to solve the problem of speaker knowledge, but carried with it toxic commitments.

Zilhão’s (2003) first major contribution is to clarify that the principle of charity does enshrine

the constitutive ideal, but not at the cost of a commitment to descriptive Bayesianism ­ thus

sidestepping Chalmers’ objections. His second is to suggest that an intersubjective reading of

radical interpretation constitutes a suitable modification in order to solve the problem of

speaker knowledge. At least in part, this is due to its showing how the radical interpreter can

plausibly gather convincing evidence to confirm or deny a putative T­sentence.

37

120434

What to make of Davidson’s extended project in light of this discussion? We have eliminated

two potential weak points from a previously unsettled discussion: first, the problem of speaker

knowledge, which is overcome by noting Davidson’s (1973) observation that (GE) confirms

authoritative truth within an intersubjective speech community, thus releasing the onus on the

principle of charity to conflate the empirical predicate ‘is ‘held true’’ with the formal semantic

predicate ‘is T’; second, the extent of the Bayesian input into the core functioning of radical

interpretation is as part of the constitutive and normative nature of interpretation. Just as

Bayesian decision theory provides answers concerning how we should behave under

idealised game situations, it provides a prescription for the content of the principle of charity.

We suggest that the Bayesian role is to be emphasised with respect to the normative aspect

of radical interpretation, but that this cannot extend to the descriptive.

38

120434

BIBLIOGRAPHY

1. Davidson, Donald. 'Truth and Meaning.' Synthese 17.1 (1967): 304­323.

2. ‘Über Sinn und Bedeutung’, in Zeitschrift für Philosophie und philosophische Kritik, 100:

25–50. Translated as ‘On Sense and Reference’ by M. Black in Translations from the

Philosophical Writings of Gottlob Frege, P. Geach and M. Black (eds. and trans.), Oxford:

Blackwell, third edition, (1980).

3. Tarski, Alfred. ‘The concept of truth in formalized languages.’ Logic, semantics,

metamathematics 2 (1956): 152­278.

4. Morris, Michael. 'An Introduction to the Philosophy of Language.' (2006) Cambridge

University Press.

5. Lepore, E. and Ludwig, K. 'Donald Davidson: Meaning, Truth, Language and Reality.'

(2005) Oxford, Clarendon Press.

6. Davidson, Donald. ‘Radical Interpretation.’ Dialectica 27.3‐4 (1973): 313­328.

7. Quine, W. V. (1960) Word and Object. Cambridge: MIT Press.

8. Davidson, Donald. 'Thought and Talk.' (1975).

9. Malpas, Jeff E. 'Donald Davidson and the Mirror of Meaning: Holism, Truth.' Interpretation

(1992).

10. Chalmers, David J. 'Frege's Puzzle and the Objects of Credence.' Mind (2011).

11. Ramsey, Frank P. ‘Truth and Probability (1926).’ The Foundations of Mathematics and

Other Logical Essays (1931): 156­198.

39

120434

12. Kolmogorov, A. N. ‘Foundations of the Theory of Probability.’ Chelsea, New York, 1933.

Trans. N. Morrison (1956). APA

13. Strevens, Michael. ‘Notes on Bayesian Confirmation Theory.’ (2006).

14. Savage, L. J. ‘The Foundations of Statistics.’ (1954) Wiley, New York. Dover edition,

1972.

15. Zilhão, António. ‘From Radical Translation to Radical Interpretation and Back.’ Principia

(2003): 229.

16. Rescorla, Michael. ‘Rationality as a Constitutive Ideal.’ (2013) A Companion to Donald

Davidson: 472­488.

17. Davidson, Donald. 2004. Problems of Rationality. Oxford: Clarendon Press

18. Davidson, Donald. ‘Essays on actions and events’. Clarendon. (1980). APA

19. Tversky, Amos. ‘Intransitivity of preferences.’ Psychological review 76.1 (1969): 31.

20. Allais, M. 1953. ‘Le Comportement de l’Homme Rationnel devant le Risque. Critique des

Postulats et Axiomes de l’École Américaine.’ Econometrica 21: 503–46.

40