fer
ëlle
, Univ
Some
Face management
lar in
rary t
ident
infer
face f
s out
ts. Ex
considered from the perspective of Relevance Theory as well as that of the Generalized
ted te
e’ a
‘all.’
mpu
ampu
informative wording that is allowed by their epistemic
state. If one knows that all students stayed on campus,
then one ought to say so, rather than using the less infor-
mative (yet logically consistent) wording (1a). Horn
(1984) identified this derivation as an instance of the more
in some contexts. Another account (e.g., Carston, 1998;
0010-0277/$ - see front matter  2009 Elsevier B.V. All rights reserved.
* Corresponding author. Tel.: +33 5 61 50 35 31; fax: +33 5 61 50 35 33.
E-mail address: bonnefon@univ-tlse2.fr (J.-F. Bonnefon).
1 Although we group them for convenience, there are important differ-
ences between the accounts offered by Chierchia and Levinson. In
particular, although Chierchia holds that scalar implicatures usually occur
by default, because he argues that they are grammar-driven, he predicts
that they only occur automatically in upward entailing contexts. In
downward entailing contexts they are said to be effortful and to occur
less frequently.
Cognition 112 (2009) 249–258
Contents lists available at ScienceDirect
Cognition
elsevier .com/locate /COGNITdoi:10.1016/j.cognition.2009.05.005Most adult speakers of English would assume the asser-
tion of (1a) to imply that the speaker believes (1b). More
generally, any assertion of the form ‘some people X-ed’ or
‘some Fs are Gs’ is usually taken to mean that not all people
X-ed, or that not all Fs are Gs, respectively.
Grice (1989) suggested that the derivation of (1b) from
(1a) reflects the general assumption that speakers abide by
the maxim of quantity, that is, that they use the maximally
Cutting, 2002; Bott & Noveck, 2004; Breheny, Katsos, &Wil-
liams, 2006; De Neys & Schaeken, 2007; Feeney et al., 2004;
Huang & Snedeker, 2009). Most of this empirical work
strived to resolve the dispute about the default character
of the inference (seeNoveck et al., 2008 for a broader review.
One type of theoretical account (e.g., Chierchia, 2004; Levin-
son, 2000 endorses the generalized view that the inference is
an implicature that is derived by default, but then cancelled
1Relevance Theory
1. Introduction
Individuals have a well-documen
pret the existential quantifier ‘som
negation of the universal quantifier
(1) a. Some students stayed on ca
b. Not all students stayed on cConversational Inference approach.
 2009 Elsevier B.V. All rights reserved.
ndency to inter-
s conveying the
s this weekend;
s this weekend.
general class of scalar inferences: the assertion of a sentence
containing an item belonging to an ordered informative-
ness scale such as hsome, alli or hpossible, certaini conver-
sationally implies that the speaker was not in a position to
use a stronger item from that scale.
Behavioural studies of the scalar inference from ‘some’ to
‘not all’ have been conducted with children (Feeney, Scraf-
ton, Duckworth, & Handley, 2004; Noveck, 2001; Papafr-
agou & Musolino, 2003) and adults (Bezuidenhout &Pragmatics
Scalar inference for a fact that all X-ed is always perceived as an underinformative utterance, it is also seen
as a nice and polite thing to do when X threatens the face of the listener. These findings areWhen some is actually all: Scalar in
Jean-François Bonnefon a,*, Aidan Feeney b, Ga
a Laboratoire Cognition, Langues, Langage et Ergonomie Maison de la recherche
bQueen’s University, Belfast, UK
a r t i c l e i n f o
Article history:
Received 6 November 2007
Revised 7 May 2009
Accepted 8 May 2009
Keywords:
a b s t r a c t
Accounts of the sca
between contempo
to this debate is to
We suggest that the
X implies a loss of
Experiment 2 rule
lower-bound contex
journal homepage: www.ences in face-threatening contexts
Villejoubert a
ersité de Toulouse, CNRS, 5 allées A. Machado, 31058 Toulouse Cedex 9, France
ference from ‘some X-ed’ to ‘not all X-ed’ are central to the debate
heories of conversational pragmatics. An important contribution
ify contexts that decrease the endorsement rate of the inference.
ence is endorsed less often in face-threatening contexts, i.e., when
or the listener. This claim is successfully tested in Experiment 1.
a possible confound between face-threatening contexts and
periment 3 shows that whilst saying ‘some X-ed’ when one knew

priate amounts to opening a new avenue of research
about
scalar inference. Finally, identifying a new context that
makes the scalar inference inappropriate has, on its
250 J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258Sperber &Wilson, 1995) endorses the particular view that the
inference is an explicature that is not derived by default, but
invited purely as a function of context.
From an empirical perspective, a critical difference be-
tween the two accounts is the time and effort required
by the narrowed interpretation of ‘some,’ i.e., some but
not all. Because this interpretation reflects the derivation
of the scalar inference, it should be comparatively quick
and easy according to the generalized view, whereas it
should be comparatively slow and effortful according to
the particular view. Current evidence tends to favour the
particular account. For example, adults tend to take longer
to reach the narrowed interpretation of ‘some’ (Bott & No-
veck, 2004; Noveck et al., 2003). Furthermore, the nar-
rowed interpretation is less frequent when participants
are instructed to respond quickly (Bott & Noveck, 2004),
or when they must simultaneously carry out a secondary
task (De Neys & Schaeken, 2007).
Breheny et al. (2006) proposed another technique to
test between the two accounts, by capitalising on the crit-
ical role that both accounts give to the context of the utter-
ance. The generalized account assumes that contextual
processing is involved in cancelling the narrowed interpre-
tation, whereas the particular account assumes that con-
textual processing is involved in reaching the narrowed
interpretation. As a consequence, the two accounts make
different predictions about the time required to interpret
‘some’ utterances when context is manipulated so as to
make the narrowed interpretation appropriate or not. The
particularised account predicts that interpretation should
be faster when the narrowed interpretation is contextually
inappropriate, whereas the generalized account makes the
opposite prediction.
Such contextual manipulations certainly allow power-
ful empirical tests of pragmatic theories; but they also re-
quire the identification of contexts that make the scalar
inference inappropriate. There is, however, surprisingly
little work on this issue, which has only addressed so far
the case of lower bound contexts. Lower-bound contexts
are such that it is enough to know that some X-ed, and
knowing that all X-ed has no additional informational
value.
For example, imagine that a company considers build-
ing a factory near a lake, but that legislation prohibits
building factories anywhere near the natural habitat of
any endangered species. An environmental expert tells
the company lawyer that:
(2) Some birds at that lake belong to endangered
species.
In this situation, knowing that all birds at the lake be-
long to endangered species is of no further import than
knowing that some birds do. The bottom line, in both cases,
is that the factory cannot be built there. In such a context,
(2) does not seem to warrant the inference that not all the
birds at the lake belong to endangered species.
Lower-bound contexts were empirically investigated
by Breheny et al. (2006), who found that they decreased
the processing time of scalar terms, as predicted by the
particular account. This investigation was an importantown, theoretical implications for current theories of sca-
lars. Accordingly, we will consider in the General Discus-
sion section whether (and at what cost) current accounts
can be reconciled with our own findings. In that discus-
sion, we will give special attention to Relevance Theory,
the main representant of the particular view on scalars,
and probably the most successful account of scalars so
far.
2. Face-threatening contexts
Consider the following exchange:
(3) a. What impression did I make during dinner?
b. Some thought you drank too much.
We argue that (3b) leaves open the possibility that all
guests at the dinner thought so, as well as the possibility
that the speaker knows for a fact that all guests at the din-
ner thought so. More generally, we suggest that the infer-
ence from ‘some X-ed’ to ‘not all X-ed’ is made less
available when X threatens the face of the listener. Brown
& Levinson (1987) define face as a sense of positive identity
and public self-esteem that all humans project and are
motivated to support in social interactions. Many actions,
called face-threatening acts, can induce a loss of face for
the speaker or for the listener (e.g., apologizing to other
people, criticizing them). Performing such an action often
requires a linguistic strategy that mitigates the face threat.
One of these strategies is to use probability expressions,
not to convey uncertainty, but rather to sugar-coat a
face-threatening act (Bonnefon & Villejoubert, 2005; Bon-
nefon & Villejoubert, 2006):researchers who already wish to manipulate context so
as to make the narrowed interpretation inappropriate;
but it also creates the need to generalise to this new
context previous findings about the time course, the
development, and the processing requirements of thescalars. Not only does it offer a solution tofirst step in a new direction: testing different accounts
of scalars in contexts wherein the scalar inference is inap-
propriate. To move the field further along this path, how-
ever, it is necessary to identify new contexts wherein the
scalar inference is inappropriate, beyond the only known
case of lower-bound contexts. Our objective in this article
is to introduce one new such class of contexts. In three
experiments, we will show that face-threatening contexts
make the narrowed interpretation of ‘some’ less appropri-
ate, and that this phenomenon is unrelated to lower-
boundedness.
The data we report in this article are not in any way
meant to definitively prove or disprove the particular or
the generalized views; rather, they are meant to move
the field forward by broadening the theoretical and
empirical scope of this debate. At the empirical level,
finding a context that makes scalar inferences inappro-

is tested in Experiment 1.
lowin
J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258 251told you, do you think it is possible that everyone hated
[loved] your poem [recipe]?’g Yes/No question: ‘From what this fellow member3. Experiment 1
3.1. Method
Participants were 53 undergraduate students at the
University of Toulouse (nine men, 44 women, mean age
21.5, SD = 1.6). Each participant read two stories (Poem
and Recipe), one in the face-boost condition, one in the
face-threat condition (the two stories were rotated across
condition, and the order of the conditions was counterbal-
anced across questionnaires). The face-threat version of
the Poem scenario read:
Imagine that you have joined a poetry club, which con-
sists of five members in addition to you. Each week,
one member writes a poem, and the five other
members discuss the poem in the absence of its
author. This week, it is your turn to write a poem
and to let others discuss it. After the discussion, one
fellow member confides to you that ‘Some people
hated your poem.’
In the face-boost version, ‘Some people hated your
poem’ was replaced with ‘Some people loved your poem.’
The Recipe story was very similar to the Poem story, except
that it involved a cooking club rather than a poetry club.
After reading each scenario, participants answered the fol-(4) a. My decision was possibly unfair.
b. You are probably going to become deaf soon.
When they read statements such as (4a) or (4b), indi-
viduals interpret the expressions ‘possibly’ and ‘probably’
as denoting very high probabilities, because they construe
them as politeness markers, i.e., as a way to preserve the
faces of those involved in the conversational exchange.
For example, in (4a), using ‘possibly’ may either save the
face of the listener, by not directly contradicting her opin-
ion that the decision was unfair, or that of the speaker, by
not directly admitting that the decision was unfair. In such
cases, individuals disregard the scalar inference from ‘pos-
sibly’ or ‘probably’ to ‘not certainly’ out of politeness con-
cerns: In particular, when X in ‘possibly X’ poses a face-
threat to the listener, there is a powerful reason why, con-
tra the maxim of quantity, the speaker would use the term
‘possibly’ although she was in an epistemic position to use
the term ‘certainly.’
Similarly, we suggest that individuals are likely to con-
strue the quantifier ‘some’ in ‘some X-ed’ as a polite device
meant not to hurt the feelings of the listener, when X-ing is
something that threatens the face of the listener. In this sit-
uation, there is again a powerful reason why, contra the
maxim of quantity, the speaker would use the term ‘some’
although she was in an epistemic position to use the term
‘all.’ If our analysis is correct, individuals should be less
likely to consider that ‘some X-ed’ means ‘not all X-ed’
when X threatens the face of the listener. This prediction3.2. Results and discussion
Overall, 83% of participants answered ‘No’ when asked
if it was possible that everyone loved their poem/recipe
when told that ‘some’ did. This result reflects the classic
scalar inference from ‘some’ to ‘not all.’ However, a reliably
lower 58% of participants answered ‘No’ when asked if it
was possible that everyone hated their poem/recipe when
told that ‘some’ did (Wilcoxon, Z ¼ 3:36; p < :001, two-
tailed, h ¼ 0:36). To be precise, 30 out of 53 participants
derived the scalar inference in both contexts, eight did
not derive it in either context, 14 derived it only in the
face-boost context, and 1 derived it only in the face-threat
context.
This result suggests that people’s tendency to draw the
scalar inference from ‘some X-ed’ to ‘not all X-ed’ decreases
when X threatens the face of the listener. We were con-
cerned, however, that our face-boost and face-threat con-
ditions may correspond to an upper-bound and a lower
bound context, respectively. Remember that a lower bound
context is one where it is important to know that some X-
ed, and where knowing that all X-ed has no additional
informational value. It might be the case that love evokes
upper-bound context whilst hate evokes lower-bound
contexts.
More precisely, when appraising the quality of a poem
or a recipe, the fact that all loved it undoubtedly carries
more information than the fact that some did. This unanim-
ity makes us more confident that the poem or recipe was
indeed truly good. But is knowing that all hated our poem
really more useful to us than knowing that some did? The
fact that some hated our poem is likely information enough
for many of us to bury it for life and not to read it again to
any other audience. Learning that everyone actually hated
the poem does not affect the decision we are already likely
to make from the fact that some hated it. If this is the case,
then sentences such as ‘some hated your poem’ would
evoke lower-bound contexts and counter the scalar infer-
ence to ‘not all,’ independently of any politeness consider-
ations. To make a stronger case for our hypothesis, we
need to rule out this alternative explanation. This can be
done by manipulating the target of the face-threat, making
it either the listener or another person, one that neither the
speaker nor the listener care for. Imagine for example that
you had a chili in a restaurant and are considering recom-
mending the place to some friends. Someone who was
there with you tells you that ‘some people hated their chili.’
Compare this with the situation where you cooked a chili
for some friends, and are considering cooking the same chili
for other guests. One of the friends who was there tells you
that ‘some people hated your chili.’
Our politeness analysis predicts that the inference is
unlikely to be derived in the second case and in the second
case only. Only in the second case will the use of ‘some’ be
construed as a politeness strategy, because neither the
speaker nor the listener care for the person whose face is
threatened in the first case. The alternative explanation
in terms of a lower-bound context predicts that the infer-
ence is unlikely to be derived in both cases. The lower-
bound context explanation assumes that it is enough to
know that some guests hated the chili to decide not to cook

it again (and that knowing that all guests hated the chili is
not going to change that decision anyway). It is also en-
ough to know that some people hated their chili to decide
not to recommend a chili place to some friends (and, again,
knowing that everybody hated their chili is not going to
change that decision anyway). Experiment 2 was con-
how likely it was that the speaker would use the word
‘some’ if she knew that the number of people who loved/
Note that this procedure makes it explicit that the
speaker is omniscient about the real number of persons
who loved/hated the trip or the Chili, whereas the omni-
science of the speaker was not explicitly asserted in Exper-
iment 1. This procedure elicited the fuzzy membership
function (Zadeh, 1965) of the concept of ‘some’ as a func-
252 J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258hated the chili (or the way the trip was organised) was in
fact 1. The question was then repeated for all numbers
up to 6.2
2 As a subsidiary question, participants also indicated whether the
speaker used the word ‘some’ because she knew a small number did, or
because she knew that a large number did, but she thought it would be
more polite to say ‘some.’ Many participants failed to answer that question,
and the number of missing answers made it impossible to analyse
responses.ducted to compare the two explanations. Furthermore,
Experiment 2 was conducted to extend our initial finding
to new scenarios and new measures of the interpretation
of ‘some.’
4. Experiment 2
4.1. Method
Participants were 120 undergraduate students at the
University of Durham (76 men, 40 women, four of undis-
closed gender; mean age 19.1, SD = 1.1). They were ran-
domly assigned to one of the Target groups (Listener vs.
Other). Each participant read two stories (Trip and Chili),
one in the face-threat condition, the other in the face-boost
condition (the two stories were rotated across condition,
and the order of the conditions was counterbalanced
across questionnaires). In the Target-other group, the
face-boost version of the Trip scenario read:
Imagine you went on a group trip. You are discussing
the trip with Alice, who was in the group. There were
six other people who went on this trip. You are consid-
ering whether to recommend the trip to some friends.
Hearing this, Alice tells you that ‘Some people loved
the way the trip was organised.’
In the Target-listener group, the face-boost version
read:
Imagine you organised a group trip. You are discussing
the trip with Alice, who was in the group. There were
six other people who went on this trip. You are consid-
ering whether to organise this same trip again next
year. Hearing this, Alice tells you that ‘Some people
loved the way you organised the trip.’
In the face-threat condition, the words ‘some people
loved . . .’ were replaced with ‘some people hated. . .’ The
Chili scenario was similar to the Trip scenario, except
that it involved recommending a chili restaurant to some
friends vs. cooking the same chili again for other guests.
After reading each scenario, participants rated, on a 10-
point scale anchored at totally unlikely and totally likely,tion of our experimental variables. In the context of this
experiment, the membership function assigns a number
(1–10) to each quantity (from 1 out of 6 to 6 out of 6) that
represents its degree of membership in the concept de-
fined by ‘some.’ A membership of 1 denotes a quantity that
is absolutely not in the concept, and a membership of 10
denotes a quantity that is a perfect exemplar of the con-
cept. Other values represent intermediate degrees of mem-
bership. The membership function method provides a
subtle and rich representation of the meaning of vague
expressions, and has been carefully validated in many
studies (see for reviews Budescu & Wallsten, 1995, Karelitz
& Budescu, 2004; but see Williamson, 1994 for an argu-
ment that fuzzy logic does not provide an adequate analy-
sis of vague expressions.)
4.2. Results
Membership functions were computed by averaging
membership judgments across participants. Fig. 1 depicts
the function values depending on the Target group (Lis-
tener vs. Other) and the polarity of X in ‘some people X-
ed’ (face-boost vs. face-threat). Two pieces of information
are of special interest to our purpose: the membership of
the number 6 (i.e., to what extent does ‘all’ enter in the
concept of ‘some’?), and the peak of the function (i.e., what
quantity is the best exemplar of the concept of ‘some’?).
The peak is computed by averaging for each participant
the quantities that received the highest membership rat-
ings, then averaging across participants the values so
obtained.3
4.2.1. Membership of 6
As expected, there was no detectable main effect of
face-boost vs. face threat (Fð1;118Þ ¼ 1:1; p ¼ :30;g2 <
:01), nor of target group (Fð1;118Þ ¼ 1:5; p ¼ :22;g2 ¼
:01). However, as expected, an interaction effect was de-
tected, Fð1;118Þ ¼ 3:9; p ¼ :05, g2 ¼ :03. This interaction
reflected the fact that face-threatening contexts only had
a detectable effect in the Target-listener group. In this
group, the membership of 6 was only 4.1 (SD ¼ 3:5) in
the face-boost condition, but went up to 5.6 (SD ¼ 3:7) in
the face-threat condition, Fð1;64Þ ¼ 12:8,
p < :001;g2 ¼ :17. In contrast, in the Target-Other group,
the membership of 6 was 3.9 (SD ¼ 3:2) in the positive
condition, and 4.3 (SD ¼ 3:3) in the negative condition.
This difference did not achieve statistical significance,
Fð1;54Þ ¼ 1:7; p ¼ :20;g2 ¼ :03.
3 Note that the apex of the curve is not a visual representation of the
peak of the membership function. The apex of the curve is the response
option that received, on average, the highest rating. The peak of the
membership function is the average of the response options that received the
highest rating. The graphical representation is very useful to assess trends
in membership, but the value of the peak cannot be read off the graph.

et:L
rget:
J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258 253ip
d
eg
re
e
in
th
e
m
ea
ni
ng
o
f S
om
e
7
Targ
Ta
Some hated XAnother relevant statistic is the proportion of partici-
pants who circled ‘10’ to indicate the membership degree
of the number 6. This proportion was 10% in the Target-
Other group, in each face condition. But in the Target-lis-
tener group, this proportion went from 17% in the face-
boost condition to 29% in the face-threat condition.
4.2.2. Peak
As expected again, there was no detectable main effect
of the face condition (Fð1;118Þ < 0:01; p ¼ :96;g2 < :01) or
target group (Fð1;118Þ ¼ 1:1; p ¼ :29;g2 ¼ :01. However,
the interaction was marginally significant, Fð1;118Þ ¼
3:1, p ¼ :08;g2 ¼ :03. This comes close enough to signifi-
cance to justify running separate analyzes in the two target
groups. These analyses reveal that face-threatening con-
texts only have a detectable effect in the Target-listener
group. In the Target-Other group, the peak was 4.3
(SD ¼ 1:1) in the face-boost condition, and 4.2 (SD ¼ 1:1)
in the face-threat condition. This difference did not achieve
statistical significance, Fð1;54Þ < 0:01; p ¼ :95, g2 < :01. In
contrast, in the Target-listener group, the peak was 3.8
(SD ¼ 1:4) in the face-boost condition, but went up to 4.2
Number of peo
M
em
be
rs
h
2
3
4
5
6
1 2 3
Fig. 1. Membership functions of ‘some people hated X’ and ‘some people loved
unimportant stranger did.istener
2
3
4
5
6
7
Other
Some loved X(SD ¼ 1:3) in the face-threat condition, Fð1;64Þ ¼ 5:3,
p < :025;g2 ¼ :08.
4.3. Discussion
Results of Experiment 2 further support our hypothesis
that the inference from ‘some X-ed’ to ‘not all X-ed’ is less
likely to be derived when X threatens the face of the lis-
tener. When this is the case, ‘some’ is understood to mean
a larger proportion – but also to leave open the possibility
that the speaker knew everyone X-ed. For example, about
one-third of participants responded that it was absolutely
likely that a speaker would say ‘some people hated your
chili’ when she knew for a fact that everyone hated the chi-
li cooked by the listener. Results of Experiment 1 are thus
replicated in Experiment 2, using different materials and a
more sophisticated measure of the interpretation of ‘some.’
Results rule out a lower-bound context explanation of
the findings. Such an explanation assumes that, e.g., the
fact that some hated the chili is decisive enough, and that
no useful information is gained by knowing that everyone
ple (out of 6)
4 5 6
X’ where X is either something the listener did or something that some

listeners acknowledge this intention, and recognise the
Imagine you gave a speech at a small political rally. You
As expected, Precision was rated much higher in the 3/6
254 J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258are discussing your speech with Denise, who was in the
audience. There were six other people in the audience.
You are considering whether to give this same speech
to another audience. Hearing this, Denise tells you:
‘Some people loved your speech.’ Denise knows that 6
people out of 6 loved your speech.
Note that, just as in Experiment 2, this procedure makes
it explicit that the speaker is omniscient about how many
people loved/hated the speech. The other stories (Chili,
Poem, Trip) were introduced in the Method sections offact that whilst using ‘some’ when one knew all is a breach
of the maxim of quantity (or of the principle of relevance),
it is also a nice and considerate thing to do in face-threat-
ening contexts.
Whether X in ‘some X-ed’ is a face-boost or a face-threat
to the listener, we expect that individuals find it dishonest
and inaccurate to use ‘some’ in situations where the speak-
er knew for a fact that all X-ed. In contrast, we expect a
critical interaction effect regarding the extent to which
individuals find it nice and considerate to use ‘some.’ More
precisely, we expect that using ‘some’ when one knew en-
ough to use ‘all’ is judged as kind and considerate in face-
threatening contexts, but unkind and inconsiderate in face-
boosting contexts.
5.1. Method
Participants were 50 undergraduate students at the
University of Durham (eight men, 42 women, mean age
20.4, SD = 2.4). They read four stories wherein X in ‘some
people X-ed’ was either a face-boost (‘some people loved’)
or a face-threat (‘some people hated’), and where the pro-
portion of people who X-ed, as known by the speaker at the
time of the utterance, was either 3/6 or 6/6. For example, in
the face-boost, 6/6 condition, the Speech scenario read:hated the chili. If this explanation was correct, then the
inference would also be less likely to be derived in the sit-
uation where the listener seeks advice about whether to
recommend a chili place to some friends. But what we ob-
served was the critical interaction we expected: The infer-
ence is derived less often only in situations where X is a
face-threat to the listener. When seeking advice about
whether to recommend a chili place to friends, the asser-
tion that ‘some people hated their chili’ does seem to invite
the inference that not all did.
5. Experiment 3
So far, we have shown that the interpretation of ‘some
X-ed’ depends on whether X threatens the face of the lis-
tener. If it does, then ‘some’ is less likely to be interpreted
as implying ‘not all.’ We predicted this effect from a face-
management analysis of statements including ‘some.’ We
argued that when X-ing is face-threatening, listeners may
consider that ‘some X-ed’ was meant to stand for ‘all X-
ed,’ out of kindness and consideration for their feelings.
In our third experiment, we test the assumption thatParticipants thus appeared to acknowledge the conflict
that arises between being a good Gricean speaker and
being a nice person. They appeared to realise that when
X threatens the face of the listener and one knows that
everyone X-ed, saying that ‘some X-ed’ is a breach of Gri-
cean or relevance-theoretic expectations, witness their re-
sponses to the Accuracy and Honesty questions. But they
also appeared to realise that this linguistic formulationcondition (M ¼ 7:1; SD ¼ 1:7) than in the 6/6 condition
(M ¼ 3:0; SD ¼ 1:4), Fð1;49Þ ¼ 182:6; p < :001, g2 ¼ :79.
This was true in both the face-threat and face boost condi-
tions, although the effect was slightly stronger in the for-
mer, as detected by the interaction effect (Fð1;49Þ ¼
7:5; p ¼ :008, g2 ¼ :13).
In contrast, Kindness ratings clearly showed the ex-
pected interaction effect (Fð1;49Þ ¼ 94:0; p < :001;g2
¼ :66). In the face-threat condition, Kindness ratings were
higher in the 6/6 condition (M ¼ 7:3; SD ¼ 2:2 for 6/6 vs.
M ¼ 5:7, SD ¼ 1:8 for 3/6), but the opposite was true in
the face-boost condition (M ¼ 3:7; SD ¼ 2:1 for 6/6 vs.
M ¼ 6:7; SD ¼ 1:6 for 3/6).
Besides the interaction effect, two main effects were de-
tected. Both effects appear to be due to the extremely low
kindness ratings given in the face-boost, 6/6 story. For this
reason, and because a very large interaction effect was de-
tected, we are inclined not to over-interpret these two
main effects, which we nevertheless report. Kindness rat-
ings were higher in the face-threat condition
(M ¼ 6:5; SD ¼ 1:7 vs. M ¼ 5:5, SD ¼ 1:4), Fð1;49Þ
¼ 20:2; p < :001;g2 ¼ :29; and they were higher in the 3/
6 condition (M ¼ 6:2; SD ¼ 1:4 vs. M ¼ 5:5; SD ¼ 1:5),
Fð1;49Þ ¼ 7:8; p ¼ :007, g2 ¼ :14.
5.3. Discussion
Results of Experiment 3 show that saying ‘some X-ed’
when one knew enough to say ‘all X-ed’ is always per-
ceived as inaccurate and even dishonest wording, whether
or not X threatens the face of the listener. However, using
‘some’ in such a situation is also perceived as a nice and
considerate thing to do when X threatens the face of the
listener.Experiments 1 and 2. After reading each story, participants
rated how accurate, considerate, honest, and nice it was of
the speaker to use the word ‘some’ in that context, using
four separate 10-point scales. There were four different
versions of the booklet. In each version, the experimental
conditions were assigned to a different story, and the con-
ditions were presented in a different random order.
5.2. Results
An index of Kindness was computed by averaging the
ratings for ‘considerate’ and ‘nice’ (between which the cor-
relation was higher than .80), and an index of Precision was
computed by averaging the ratings for ‘accurate’ and ‘hon-
est’ (between which the correlation was higher than .80).
Fig. 2 depicts the distribution of the two indices as a func-
tion of our experimental variables.

24
6
8
10
Some loved X
Precision of
J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258 255Pr
ec
isi
on
a
nd
K
in
dn
es
s
of
th
e
Sp
ea
ke
r
6
8
10
Some hated X
Kindness of
Some hated X
Precision ofwas the act of a kind speaker, witness their responses to
the Niceness and Considerateness questions.
6. General discussion
The scalar inference from ‘some X-ed’ to ‘not all X-ed’ is
so intuitively compelling that some theorists (but not all)
have argued that it constitutes a default reaction, which
is then cancelled in some contexts (Levinson, 2000). Other
theorists have argued that this inference is purely a matter
of context, just like any other (particularised) conversa-
tional inference, whose derivation is a matter of applying
general principles of communication, such as the maxim
of quantity (Grice, 1989), or the principle of relevance
(Sperber & Wilson, 1995).
Both accounts assume that the scalar inference should
be available in some contexts but not in others, but make
different predictions as to how such contextual aspects
are processed. An important contribution to future experi-
mental tests of the theories then consists of identifying
contextual manipulations that can turn the scalar infer-
ence from available to unavailable. Very few such manipu-
lations, though, have been proposed so far. In this article,
we have identified a new and simple contextual manipula-
2
4
When 3 out of 6 did When 6 out of 6 did
Fig. 2. Precision and Kindness attributed to the speaker asserting ‘some people X
(X = hated) to the listener, and of the true proportion of people who X-ed (3/6 vSome loved X
Kindness oftion that affects the availability of the scalar inference:
When X in ‘some X-ed’ threatens the face of the listener,
individuals are less likely to infer that the speaker meant
or knew that not all X-ed – and this is because they con-
sider the possibility that the speaker might want to be nice
more than to be precise.4 Although, as previously stated, the
purpose of this research was not to disprove either the gen-
eralized or the contextual account of scalar inferences, these
findings already have significant implications for both ac-
counts, which we now consider in turn.
When 3 out of 6 did When 6 out of 6 did
-ed,’ as a function of whether X is a face-boost (X = loved) or a face-threat
s. 6/6).
4 Note however that the speaker cannot be formally accused of lying. If
one believes that all X-ed, then one also believes that some X-ed, and one
cannot be accused of lying by saying that some X-ed. We may talk about
this strategy as ‘sugar-coating,’ but we must not confuse this kind of polite
sugar-coating (Bonnefon & Villejoubert, 2006; Demeure, Bonnefon, &
Raufaste, 2008; Demeure, Bonnefon, & Raufaste, 2009) with another kind
of face-saving move, informally known as the white lie, wherein what is
asserted is downright false. These face-saving moves should not be
confused either with the self-serving use of ambiguous language. It is quite
possible, though, that contextual manipulations based on self-serving uses
of language may yield comparable results to our face-management
manipulation. We leave that discovery to future research.

their affective impact (the way they make the listener feel).
256 J.-F. Bonnefon et al. / Cognition 112 (2009) 249–2586.1. Implications for the generalized account
The generalized account rests on the plausible assump-
tion that deriving the scalar inference by default (and then
cancelling it, in the rare cases when it is not appropriate)
allows for quick and efficient communication. As noted
by Noveck & Sperber (2007), though, this gain in efficiency
decreases with the frequency of the situations in which the
default inferences must be cancelled. If this frequency
reaches some critical level, it can become less efficient to
derive the inference by default.
Explaining our results within the generalized account
would amount to considering that the default inference
from ‘some’ is cancelled in face-threatening contexts. The
problem with this solution is that these contexts are likely
to be far from rare. To get a rough order of magnitude for
their frequency, consider that Youmans (2001) analyzed
taped American English conversations (23,000 words,
including about 500 probability terms) and reported that
probability terms were used as face-management devices
(as in ‘My decision was possibly unfair’) at the rate of 4
out of 10 occurrences. If the polite use of ‘some’ is any-
where near as frequent, then the rationale for turning the
scalar inference from ‘some’ into a default is called into
question, because too many exceptions to the default
quickly make it counter-productive.
6.2. Implications for Relevance Theory
From a relevance-theoretic point of view (Sperber &
Wilson, 1995), the scalar inference from ‘some’ to some
but not all is the result of a presumption of optimal rele-
vance. According to the original formulation of the pre-
sumption of optimal relevance, an optimally relevant act
of communication (a) achieves cognitive effects that are
large enough to make it worth processing; and (b) is the
most relevant one that the communicator could have used
to produce these effects (e.g., they cannot be achieved by
another act of communication that would require less pro-
cessing effort). To interpret a statement, listeners are as-
sumed to follow a path of least effort, and to stop as soon
as an interpretation of the statement meets their expecta-
tions of optimal relevance.
Outside lower-bound contexts, the unenriched inter-
pretation some and possibily all X-ed does not yield enough
cognitive effects to justify the effort of processing ‘some.’
Listeners then move on to the enriched interpretation some
but not all X-ed. This interpretation yields enough effects to
justify the additional processing effort; and these effects
could not be achieved by a simpler statement than ‘some
X-ed.’ Listeners thus settle on this interpretation that
meets their expectation of optimal relevance.
The problem is that this explanation should apply the
same whether X-ing is a face-boost or a face-threat to the
listener. In contrast, our results clearly show that the scalar
inference is less frequent when X-ing is a face-threat to the
listener; and that listeners may even consider the interpre-
tation all X-ed in that situation. Remember, for example,
that 29% of participants in Experiment 2 gave a rating of
10 (out of 10) to the likelihood of saying ‘‘some people
hated your chili” when knowing for a fact that everyoneAs a matter of fact, Haugh (2003) argues that this focus on
epistemic rather than affective effects seriously limits the
ability of Relevance Theory to account for politeness phe-
nomena. It remains to be seen whether Relevance Theory
will undergo this major transformation in order to account
for the effect of face-threatening contexts, but this evolu-
tion seems unlikely. Indeed, Relevance Theory does not
seem to be moving towards extending the definition of
cognitive effects, but, as we shall see, rather towards
extending the definition of optimal relevance. We now
consider how this evolution may account for the effect of
face-threatening contexts.
6.2.2. Redefining optimal relevance
In the postface to the 1995 edition of Relevance, Sperber
and Wilson introduced a revised formulation of the pre-
sumption of optimal relevance. Just as the original formu-
lation, this revised version has two clauses. The first clause
is kept unchanged: an optimally relevant act of communi-
cation achieves cognitive effects that are large enough to
make it worth processing. The novelty is in the second
clause. The original formulation required that the cognitive
effects could not be achieved by another, more relevant act
of communication. The revised clause weakens this de-
mand; it only requires that the cognitive effects cannot
be achieved by a more relevant act, that would be compat-
ible with the communicator’s abilities and preferences.
This revised formulation helps to make one important
step towards explaining our findings, but also creates a
serious difficulty. Before we discuss these issues, we illus-
trate the workings of the revised formulation through
Sperber and Wilson’s own example (p. 275): Peter and
Mary are planning a trip to southern France, and Peter is
suggesting they could go and meet their old friend Gérard,
if that would not take them too far from their way. Peter
inquires about the whereabouts of Gérard: ‘Where doeshated the chili. Relevance theorists may consider two ways
out of this difficulty, which we will discuss in turn.
6.2.1. Redefining cognitive effects
It could be tempting to handle face-threatening con-
texts by directly factoring face-management into the com-
putation of cognitive effects. This would amount to
considering that, ceteris paribus, an utterance that pre-
serves the face of the listener has greater cognitive effects
than an utterance that hurts the feelings of the listener.
With this assumption, listeners would stop at the unen-
riched interpretation some and possibly all hated your
poem, because this unenriched interpretation already
yields sufficient ‘effects’ (redefined) to justify the effort of
processing it; thus the non-derivation of the scalar
inference.
This solution is intuitively compelling, but its far-reach-
ing implications are hard to calculate, for it would involve a
major redefinition of a core notion within Relevance The-
ory, that of a cognitive effect. Up to now, cognitive effects
have been diversely but consistently defined in terms of
their epistemic impact on the listener (the extent to which
they increase the listener’s knowledge), to the exclusion of

Gérard lives?’ To which Mary replies: ‘Somewhere in the
south of France.’
Does Mary’s reply implicate that she does not know
where exactly Gérard lives? Such a claim would follow
from the original formulation of the presumption of opti-
At this point, the only way we see for Relevance Theory
to account for our results is to make an ad hoc assumption
about processing effort. This assumption would be that
everyone X-ed is usually a costly interpretation of ‘some
X-ed,’ but that, for some reason, it gets cheaper when X-
ing is a threat to the face of the listener; and even cheaper
than some (but not all) X-ed. Note that we do not currently
know what that reason might be. This ad hoc solution has
readily testable consequences in terms of response time.
Indeed, it predicts that interpreting ‘Some X-ed’ as some
J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258 257mal relevance: If Mary knew that Gérard lived, say, in Tou-
louse, her vague reply would not have been optimally
relevant. But the revised formulation allows another anal-
ysis. If Mary is dead against seeing Gérard, then her vague
reply is optimally relevant even if she knew that Gérard
lived in Toulouse; because the more relevant reply ‘Gérard
lives in Toulouse’ would go against her preferences.
Consider now the utterance ‘Some people hated your
poem.’ Can this utterance convey everyone hated your poem
in an optimally relevant way? According to the original
formulation of optimal relevance, the answer is No:
Although the cognitive effects of this interpretation are
most likely worth the processing efforts, the same effects
could be achieved for less effort by simply saying ‘Everyone
hated your poem.’ However, the revised formulation of
optimal relevance takes into account that this simpler for-
mulation would go against the preferences of the speaker,
in this case the plausible preference for being gentle rather
than inconsiderate. In turn, this permits interpreting ‘Some
people hated your poem’ as everyone hated your poem to
meet expectations of optimal relevance.
This is an important step in order to account for our
findings, but it also raises a difficult problem for the theory.
Whilst there is an argument that the utterance ‘Some peo-
ple hated your poem’ optimally conveys everyone hated
your poem, there is a parallel argument that it optimally
conveys some (but not all) hated your poem. Indeed, as per
the standard analysis of ‘some,’ the cognitive effects of
some (but not all) hated your poem are worth the processing
efforts; and there is no simpler way to convey some (but
not all) hated your poem than ‘Some hated your poem.’
Since saying ‘Some people hated your poem’ can be
optimally relevant to convey either some (but not all) hated
your poem and everyone hated your poem, how do listeners
settle on one communicative intention or the other? Rele-
vance Theory assumes that listeners follow a path of least
effort, and stop when an interpretation meets their expec-
tations of relevance. In particular, the theory is adamant
that listeners never continue looking for another, more ef-
fort demanding interpretation, when they have reached an
interpretation that meets their expectations of optimal rel-
evance (Sperber & Wilson, 1995). Because both interpreta-
tions of ‘Some hated your poem’ meet expectations of
optimal relevance, we arrive at the conclusion that listen-
ers will settle on the interpretation that requires the least
processing effort. Now, generally speaking, it is hardly dis-
putable that some (but not all) X-ed is an easier interpreta-
tion of ‘some X-ed’ than everyone X-ed;5 but this is not the
interpretation that participants appear to consider when X-
ing is a threat to the face of the listener.
5 Prior experimental data that we reviewed in the introduction estab-
lished that some (and possibly all) X-ed is an easier interpretation of ‘some X-
ed’ than some (but not all) X-ed. The assumption that we qualify as hardly
disputable is that the interpretation of ‘some’ as everyone is more difficult
than either the some (but not all) or some (and possibly all) interpretations.(but not all) X-ed will take longer than interpreting it as
everyone X-ed when X-ing is a threat to the face of the lis-
tener; but that the opposite is true when X-ing is a boost to
the face of the listener.6 Of course Relevance theorists may
adopt a different explanatory strategy to the one we have
outlined. Nevertheless, a test of our prediction would make
clear how useful to the theory our explanatory strategy is
likely to be.
7. Conclusion
The findings we have reported raise new theoretical
challenges for pragmatic accounts of scalar implicatures,
but they also open a new avenue of empirical research into
scalars.7 Researchers now have at their disposal an easy way
to manipulate whether a scalar inference is contextually
appropriate, in order to observe the effect of this manipula-
tion in reading time or response time studies. In parallel
with chronometric issues, the question is opened whether
mental workload moderates the effect of face-threatening
contexts, and in which direction (De Neys & Schaeken,
2007). If mental workload were to disrupt the effect of
face-threatening contexts, the question would be raised
whether the processing of face-threatening contexts is cor-
related with cognitive ability (Feeney et al., 2004). Finally,
an important task that was not within the scope of the pres-
ent paper, but which deserves careful attention, is to address
the developmental issue of whether the sensitivity to face-
threatening contexts comes to children simultaneously with
their understanding of the scalar implicature, or whether it
occurs at a later developmental stage. Overall, the investiga-
tion of face-threatening contexts has the potential to give
rise to a new body of empirical data, which will pose new
constraints on the way we theorize about scalars.
Acknowledgement
Jean-François Bonnefon and Gaëlle Villejoubert, Labora-
toire Cognition, Langues, Langage et Ergonomie (CNRS,
UTM, EPHE), Toulouse, France. Aidan Feeney, School of Psy-
6 The generalized account predicts that the interpretation everyone X-ed
should take longer than the default interpretation some (but not all) X-ed,
independently of whether X-ing is a threat or a boost to the face of the
listener.
7 We note that this future empirical research is likely to use different
tasks and measures than those used in the present article. More precisely, it
is likely to use reading and response time measures, and the truth or falsity
judgement task, which are standard in experiments about the processing of
scalar terms. The present research, in contrast, used tasks that were geared
at establishing the effectiveness of our contextual manipulation; and it
accordingly gave priority to metalinguistic evaluations of word choice as a
function of context.

chology, Queen’s University Belfast, Northern Ireland. We
thank Bénédicte Bès, Wim De Neys, and Jean-Baptiste
Van der Henst for their thougthful comments. This re-
search was supported by grant ANR-07-JCJC-0065-01 from
the Agence Nationale de la Recherche, and by grants PN
08.010 (awarded by the British Council) and 18223VF
(awarded by the Ministère des Affaires étrangères et
Européennes) under the Alliance Franco-British Pro-
gramme 2008.
References
Bezuidenhout, A., & Cutting, J. C. (2002). Literal meaning, minimal
propositions, and pragmatic processing. Journal of Pragmatics, 34,
433–456.
Bonnefon, J. F., & Villejoubert, G. (2005). Communicating likelihood and
managing face: Can we say it is probable when we know it to be
certain? In B. G. Bara, L. Barsalou, & M. Buciarelli (Eds.), Proceedings of
the 27th annual conference of the cognitive science society
(pp. 316–321). Mahwah, NJ: Erlbaum.
Bonnefon, J. F., & Villejoubert, G. (2006). Tactful or doubtful? Expectations
of politeness explain the severity bias in the interpretation of
probability phrases. Psychological Science, 17, 747–751.
Bott, L., & Noveck, I. A. (2004). Some utterances are underinformative: The
Demeure, V., Bonnefon, J. F., & Raufaste, E. (2008). Utilitarian relevance
and face management in the interpretation of ambiguous question/
request statements. Memory and Cognition, 36, 873–881.
Demeure, V., Bonnefon, J. F., & Raufaste, E. (2009). Politeness and
conditional reasoning: Interpersonal cues to the indirect
suppression of deductive inferences. Journal of Experimental
Psychology: Learning, Memory, & Cognition, 35, 260–266.
Feeney, A., Scrafton, S., Duckworth, A., & Handley, S. J. (2004). The story of
some: Everyday pragmatic inference by children and adults. Canadian
Journal of Experimental Psychology, 58, 121–132.
Grice, H. P. (1989). Studies in the way of wordsStudies in the way of words.
Cambridge, MA: MIT Press.
Haugh, M. B. (2003). Anticipated versus inferred politeness. Multilingua,
22, 397–413.
Horn, L. R. (1984). Toward a new taxonomy for pragmatic inference: Q-
based and R-based implicature. In D. Schiffrin (Ed.), Meaning, form,
and use in context (pp. 11–42). Washington, DC: Georgetown
University Press.
Huang, Y. T., & Snedeker, J. (2009). Online interpretation of scalar
quantifiers: Insight into the semantics-pragmatics interface.
Cognitive Psychology, 58, 376–415.
Karelitz, T. M., & Budescu, D. V. (2004). You say ‘‘probable” and I say
‘‘likely”: Improving interpersonal communication with verbal
probability phrases. Journal of Experimental Psychology: Applied, 10,
25–41.
Levinson, S. C. (2000). Presumptive meanings: The theory of generalized
conversational implicature. Cambridge, MA: MIT Press.
Noveck, I. A. (2001). When children are more logical than adults:
Investigation of scalar implicature. Cognition, 78, 165–188.
Noveck, I. A., & Posada, A. (2003). Characterising the time course of an
implicature. Brain and Language, 85, 203–210.
258 J.-F. Bonnefon et al. / Cognition 112 (2009) 249–258onset and time course of scalar inferences. Cognition, 51, 437–457.
Breheny, R., Katsos, N., & Williams, J. (2006). Are generalised scalar
implicatures generated by default? An on-line investigation into the
role of context in generating pragmatic inferences. Cognition, 100,
434–463.
Brown, P., & Levinson, S. C. (1987). Politeness: Some universals in language
usage. Cambridge: Cambridge University Press (Orginal work
published 1978).
Budescu, D. V., & Wallsten, T. S. (1995). Processing linguistic probabilities:
General principles and empirical evidence. In J. Busemeyer, D. L.
Medin, & R. Hastie (Eds.), Decision making from a cognitive perspective
(pp. 275–318). San Diego, CA: Academic Press.
Carston, R. (1998). Informativeness, relevance and scalar implicature. In R.
Carston & S. Uchida (Eds.), Relevance theory: Applications and
implications (pp. 179–236). Amsterdam: John Benjamins.
Chierchia, G. (2004). Scalar implicatures, polarity phenomena, and the
syntax/pragmatics interface. In A. Belletti (Ed.). Structures and
beyond: The cartography of syntactic structures (Vol. 3, pp. 39–103).
Oxford: Oxford University Press.
De Neys, W., & Schaeken, W. (2007). When people are more logical under
cognitive load – Dual task impact on scalar implicatures. Experimental
Psychology, 54, 128–133.Noveck, I. A., & Reboul, A. (2008). Experimental pragmatics: A Gricean
turn in the study of language. Trends in Cognitive Sciences, 11,
425–431.
Noveck, I. A., & Sperber, D. (2007). The why and how of experimental
pragmatics: The case of ‘scalar inferences’. In N. Burton-Roberts (Ed.),
Pragmatics. Basingstoke, UK: Palgrave Macmillan.
Papafragou, A., & Musolino, J. (2003). Scalar implicatures: Experiments at
the semantics/pragmatics interface. Cognition, 86, 253–282.
Sperber, D., & Wilson, D. (1995). Relevance, communication and cognition
relevance communication and cognition. Oxford: Blackwell (Orginal
work published 1986).
Williamson, T. (1994). Vagueness. London: Routledge.
Youmans, M. (2001). Cross-cultural differences in polite epistemic modal
use in American English. Journal of Multilingual and Multicultural
Development, 22, 57–73.
Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8, 338–353.