DataPDF Available

Recipient design in human communication: simple heuristics or perspective taking?

published: 25 September 2012
doi: 10.3389/fnhum.2012.00253
Recipient design in human communication: simple
heuristics or perspective taking?
Mark Blokpoel*,M arlieke va n Kes teren ,Arjen Stolk ,Pim Haselager ,Ivan Toni and Iris van Rooij
Radboud University Nijmegen, Donders Institute for Brain Cognition and Behaviour, Nijmegen, Netherlands
Edited by:
Bert Timmermans, University
Hospital Cologne, Germany
Reviewed by:
Colin Camerer, California Institute
of Technology, USA
Emily Falk, University of Michigan,
Mark Blokpoel, Department of
Artificial Intelligence, Donders
Institute for Brain, Cognition and
Behaviour, PO 9104, 6500HE,
Nijmegen, Netherlands.
These authors contributed equally
to this work.
Humans have a remarkable capacity for tuning their communicative behaviors to different
addressees, a phenomenon also known as recipient design. It remains unclear how this
tuning of communicative behavior is implemented during live human interactions. Classical
theories of communication postulate that recipient design involves perspective taking,
i.e., the communicator selects her behavior based on her hypotheses about beliefs and
knowledge of the recipient. More recently, researchers have argued that perspective
taking is computationally too costly to be a plausible mechanism in everyday human
communication. These researchers propose that computationally simple mechanisms,
or heuristics, are exploited to perform recipient design. Such heuristics may be able to
adapt communicative behavior to an addressee with no consideration for the addressee’s
beliefs and knowledge. To test whether the simpler of the two mechanisms is sufficient
for explaining the “how” of recipient design we studied communicators’ behaviors
in the context of a non-verbal communicative task (the Tacit Communication Game,
TCG). We found that the specificity of the observed trial-by-trial adjustments made by
communicators is parsimoniously explained by perspective taking, but not by simple
heuristics. This finding is important as it suggests that humans do have a computationally
efficient way of taking beliefs and knowledge of a recipient into account.
Keywords: heuristics, recipient design, communication, computational intractability
Imagine that a person on the street comes up to Ann and asks
her: “Where can I find a supermarket?” Ann’s reply may depend
in subtle ways on a multiplicity of cues such as whether or not the
person speaks with a foreign accent, the person is speaking hastily,
or the person is by car. In the presence of such cues she may, for
instance, speak more clearly, use simpler words, make shorter sen-
tences, and give directions specifically how to drive there by car.
As a result of these adjustments Ann may construct a message that
the addressee is more likely to understand than otherwise. This
adaptation of a communicative signal—such that it is tuned to
the addressee—is known as recipient design (Sacks et al., 1974).
Classical theories of communication consider recipient design
as constitutive of genuine or intentional communication (Grice,
1975, 1989; Levelt, 1989), as opposed to mere accidental or non-
intentional forms of communication. Yet, recently a debate has
ensued on the presumed ubiquity of recipient design in every-
day communication (Clark, 1996; Horton and Keysar, 1996;
Keysar et al., 1998), as well as on the nature of the cognitive
mechanisms underlying the phenomenon (Epley et al., 2004;
Shintel and Keysar, 2009; Galati and Brennan, 2010). With this
paper we aim to contribute particularly to the second topic of
debate: i.e., the nature of the mechanisms underlying recipient
design in everyday (interactive) communication.1Specifically, we
1As the focus of the debate in the literature has been on testing
the extent to which people display egocentric bias when communicating
consider two proposed explanations of the “how” of recipient
design and present evidence that the computationally simpler
of the two cannot by itself account for the subtle and context-
sensitive ways in which humans fine tune their messages to
Traditionally, recipient design is thought to involve a mecha-
nism that forms hypotheses about, among other things, beliefs,
and knowledge of the addressee, and uses these hypotheses to
optimize the message for the addressee (Grice, 1975, 1989; Clark
and Carlson, 1982; Levelt, 1989). Such a perspective taking mech-
anism can explain several of the adaptations made by Ann in our
example. For instance, observing the addressee’s accent, Ann may
infer that English is not his first language and therefore that he
is unlikely to know low frequency words and understand gram-
matically complex English sentences. She may in turn use this
(inferred) information to construct simpler sentences that she
believes are understandable for the addressee.
In more recent years, researchers have argued that a perspec-
tive taking mechanism for recipient design is computationally
too costly to be plausibly invoked automatically in everyday
communication (Epley et al., 2004; Shintel and Keysar, 2009;
(Horton and Keysar, 1996), it is important to point out that our research does
not set out to directly contribute to that debate. In fact, considerations of the
nature of the mechanisms underlying recipient design seem to be orthogo-
nal to the question of the relative frequency of ecogentric bias in everyday
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |1
Blokpoel et al. Recipient design: heuristics or perspective taking?
Galati and Brennan, 2010). These researchers propose that instead
recipient design is based on simple heuristics or rules-of-thumb
triggered by the presence or absence of certain cues.2Such a cue-
based heuristics mechanism for recipient design may achieve com-
municative fine tuning without any resort to hypotheses about
the beliefs and knowledge of the addressee. To illustrate, consider
again the example scenario: Ann may take the foreign accent as a
cue to classify the addressee as a tourist and the habitual response
triggered by this classification may be to speak more clearly, use
shorter sentences, use higher frequency words, etc. Again, as a
result of such adjustments Ann may construct a message that the
addressee is more likely to understand than otherwise. Observing
such communicative fine tuning one may think Ann designed the
message for the tourist based on what she thinks he knows and
believes, but in fact this would be a case of mere appearance of
perspective taking. Given the presumed intractability of recipi-
ent design by perspective taking, and the evident availability of
an alternative and computationally cheaper heuristics account, it
seems prudent to investigate if perhaps the computationally sim-
pler account can by itself account for recipient design in human
Understanding the computational sufficiency of different
mechanisms for recipient design is also of considerable practical
importance. For example, it can give us insight into how to cre-
ate artificial agents that can communicate in human ways (e.g.,
in the context of human-robot interaction; Breazeal, 2002; Green
et al., 2008). Imagine a situation where Ann is in a shopping mall
and is being approached by a robot who wishes to provide her
with information about an attractive sale (Shiomi et al., 2007;
Satake et al., 2009). How should the robot adapt its commu-
nicative signals such that Ann will better understand it? If the
adaptation could be achieved by a set of simple heuristics this
could make the design of such socially interactive robots much
more feasible, as compared to when the adaptation would require
the robot to engage in elaborate hypothesizing about the beliefs
and knowledge of the addressee.
In this paper, we investigate the computational sufficiency
of simple heuristics-based mechanisms for explaining recipi-
ent design as it occurs in human–human communication. We
specifically set out to identify situations in which humans adapt
communicative signals in ways that cannot be explained by simple
heuristics. As our examples illustrate, it can be difficult to tease
apart perspective taking and heuristics in natural language con-
versation. For this reason, we study recipient design in the context
of a communication game in which players create novel commu-
nicative signals in the absence of previous conventions. The form
of communication occurring in this game can be compared to
real-world situations where two agents act without a completely
shared lexicon, such as when speaking to a tourist or when sig-
naling something from a distance or behind a window. The game
that we use is called the Tacit Communication Game (TCG, De
Ruiter et al., 2010) and it has been previously validated in several
2This heuristics account may take inspiration from the fast and frugal heuris-
tic program in decision-making (Gigerenzer and Todd, 1999; Marsh, 2002;
Gigerenzer and Brighton, 2009).
The TCG has been developed to study human communica-
tion under controlled experimental conditions (De Ruiter et al.,
2007, 2010; Newman-Norlund et al., 2009; Noordzij et al.,
human communication that goes under the label of experimen-
tal semiotics. This approach has been contrasted by Galantucci
(2009)withexperimental pragmatics. Whereas experimental prag-
matics focuses on spoken conversation, experimental semiotics
is concerned with human communication more generally and
the emergence of novel ways of communicating in particular
(Galantucci and Garrod, 2010). Experimental semiotics is char-
acterized by the use of games in which participants are to dis-
cover novel communicative systems. By studying communication
in experimental semiotic games it becomes possible to test for
fundamental characteristics of communication free from the con-
ventions introduced by linguistic settings. Semiotic games also
give more experimental control on the common ground shared by
participants during communication. Several semiotic games have
been developed and studied (Camerer, 2003; Weber and Camerer,
2003; Galantucci, 2005; Selten and Warglien, 2007; Scott-Phillips
et al., 2009; Feiler and Camerer, 2010), with the TCG being one
of the few that has been studied both from a behavioral and
neuroscientific perspective.
The TCG is a communicative task where two players, a sender
(referred to as she) and a receiver (referred to as he)playagame
on a 3 ×3gridboard.Figure 1 depicts the sequence of events
in a typical communicative trial. Here, only the sender knows
a goal state that has to be reached in a cooperative fashion by
her and the receiver (e.g., the circle is to end up in the upper
left corner and the rectangle in the lower right corner, see event
2inFigure 1). The senders’ task is to signal the receiver what
the same time she is to contribute to achieving the final goal
state by moving her token to its goal position (e.g., the sender’s
circle must end up in the upper left corner of the board, but
along the way signal to the receiver that he is to place his rect-
angle in the lower right corner). Although the TCG may look
superficially very dissimilar to everyday face-to-face (linguistic)
communication, in fact it is designed to capture the fundamen-
tal problems faced by human communicators during their daily
interactions. For instance, in the TCG the sender gives direc-
tions to the receiver using non-conventional means on the basis
of limited common ground. The structure of this communica-
tive problem closely matches that of the scenario described in
the Introduction, where a tourist asks Ann for directions. More
generally, every human starts without access to the local commu-
nicative conventions. Accordingly, the TCG addresses the human
ability to quickly build new semiotic conventions, while provid-
ing strong experimental control of the communicative setting,
and precise quantification of the communicative behavior of the
Previous research has shown that in this game senders engage
in recipient design, i.e., they tune their communicative signals
to the particular receiver who is their current co-player. For
instance, De Ruiter et al. (2010) observed that game performance
(i.e., number of correct goal configurations produced by the two
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |2
Blokpoel et al. Recipient design: heuristics or perspective taking?
Sender Receiver
FIGURE 1 | The chronological order of phases in a trial of the Tacit
Communication Game, left (in blue) is the sender and right (in red)
the receiver. In phase 1 both sender and receiver are presented with their
assigned token for this trial. Next, after the sender presses a start button,
in phase 2 the receiver is presented with a blank screen while the sender
is shown the goal configuration of both tokens and she plans her
movements (unrestricted time). After the sender presses the start button
again, phases 3 and 4, both players’ screens display the sender’s (blue)
token and the sender is able to move her token for 5 sec. It is during this
phase that the sender can communicate the relevant information of the
goal configuration to the receiver using movements of her token. After the
sender is finished, she either presses the start button or the 5sec time
limit expires and phases 5 and 6 begin. Here both players’ screens display
the receivers (red) token and the receiver can move his token. Now the
receiver should move his token to the location (and orientation) that he has
inferred from the sender’s movement. Finally, after the receiver has
finished moving his token, both players receive feedback for their
performance on this trial. A green check mark denotes that both players’
tokens are in the exact same location and orientation as depicted in the
goal configuration shown to the sender in phase 2; and a red cross (not
shown in this figure) denotes that the tokens are not placed correctly.
players) improved when senders received feedback about whether
or not their signals were successful in communicating with the
receiver. This finding suggests that senders use this feedback to
better tune their signals to the receiver. Also, in a variant of the
TCG adapted to child-level complexity, Newman-Norlund et al.
(2009) observed that (adult) senders make very specific changes
to their communicative signals depending on whether or not they
believed to be playing with an adult or a child. For instance, they
observed that initially the length of the pause by the sender’s
token on the receiver’s goal location—taken to be an ostensive
signal—was significantly longer when the sender was told the co-
player was a child rather than an adult. Given that performance
of the receiver was identical in the two conditions—viz., the
receiver was played by an experimental confederate—the effect
slowly disappeared as the sender got further tuned to the current
Findings such as these show that the TCG evokes recipi-
ent design, making the game a suitable platform for our study.
Although the abovementioned findings were previously inter-
preted as evidence for a perspective taking mechanism for recip-
ient design, these observations could also be explained using
cue-based heuristics mechanisms. For instance, the finding that
performance improves with feedback can be explained by a
“take-the-best” heuristic (Gigerenzer and Goldstein, 1996), which
selects signals from a predefined list based on their cue validity
and where cue validity is updated on the basis of the received
feedback. The finding that signals are initially different when a
sender thinks she is playing with a child versus an adult, yet
become comparable when performance of the co-players turns
out to be identical, can be explained by an “anchor-and-adjust”
heuristic (Tversky and Kahneman, 1974; Epley et al., 2004).
Such a heuristic can adopt different anchors for discriminabil-
ity of a signal for different categories of addressees and adjusts
these discriminability values upon finding that lower levels suffice
as well.
Additionally, a study by Noordzij et al. (2009) showed that the
right posterior Superior Temporal Sulcus (right pSTS) is active in
both senders (during planning) and receivers (during observation
of the signal). Noordzij et al. reasoned that the right pSTS imple-
ments an intention recognition process that is used by receivers to
understand signals, but also by senders as a subprocess of recipi-
ent design. Their finding, however, does not unequivocally show
that sender’s engage in this form of perspective taking. Namely,
it is also consistent with the idea that the pSTS implements the
shared representations of senders and receivers that are activated
during communication.
These observations are not to argue that the results in the
literature are not suggestive of perspective taking. We merely
wish to point out that the evidence is not yet conclusive: the
findings do not rule out that the effects can be explained
by simple heuristic mechanisms as well. Moreover, given the
prevalent idea that perspective taking is computationally costly,
whereas heuristics are computationally cheap, the latter may
prima facie make for a more plausible explanation of the effects
than the former. By studying in more detail context-specific
dependencies between receiver behaviors and sender signals
in the TCG, we aim to contribute more convincing evidence
that recipient design also draws on mechanisms of perspective
Specifically, we set out to study adaptations made by senders
to their signals on a given trial as a function of the type of
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |3
Blokpoel et al. Recipient design: heuristics or perspective taking?
error made by the receiver on a preceding trial. Our rationale
for studying such trial-to-trial dependencies is the following: if a
receiver makes an error in interpreting a sender’s previous signal,
this may cause the sender to change her signal to make it easier to
understand for the receiver—i.e., recipient design. The adaptation
may be achieved by invoking some form of perspective taking.
For instance, observing the error made by the receiver, the sender
could form hypotheses about why the receiver misunderstood
certain aspects of the signal, and then use these hypotheses to
make her subsequent signals easier to understand for the receiver.
Alternatively, the sender may make her subsequent signals eas-
ier to understand without any recourse to perspective taking but
instead by using only simple heuristics. In the latter case, how-
ever, the nature of the adaptations should be such that they can be
explained by invoking some simple function mapping error cues
to adaptations. We test whether or not the trial-to-trial adapta-
tions made by senders in the TCG can be modeled by such simple
heuristic rules.
We report novel analyses of behavioral data collected by Stolk
et al. (2012). The aim of Stolk et al. was to study the neural corre-
lates of human intentional communication using MEG imaging.
The experiment consisted of two tasks, namely the TCG and a
comparable control task without communicative dependencies.
As the tasks were completely blocked in the design, we can focus
on the design of the TCG task by itself. In this section we present
the methods that were relevant for acquiring the behavioral data
that we analyzed.
Fifty-two participants, students and colleagues, took part in the
study. We will report analyses of the behavioral data obtained for
a selection of 46 participants. Two pairs were excluded because
of technical problems and one pair because performance was
exceptionally poor.
Participants gave informed consent according to institu-
tional guidelines of the local ethics committee (CMO region
Arnhem-Nijmegen, The Netherlands) and were either offered
a financial payment or given credits toward completing a
course requirement. The age of participants ranged between
18 and 40 years and all had normal or corrected-to-normal
Participants of each pair sat behind a 19-inch monitor on which
the game board (3 by 3 squares) and the tokens were displayed.
Participants controlled their token with a hand-held controller.
This controller contained (among others) four buttons which
3Not only were these participants successfulon just 30 out of 80 trials (37.5%),
much less than the average performance (72%), but we also observed that
these participants did not converge on a common (i.e., shared) strategy.
Although one may argue that these participants still engaged their commu-
nicative abilities, we cannot investigate recipient design using our measures
for their data, because our measures were defined on particular adaptations of
particular common strategies.
were positioned left, right, up, and down from one another, these
corresponded with the four directions in which a token on the
board could be moved. Additionally, one of the shoulder but-
tons on the right side of the controller was used to perform a 90
clockwise rotation of the token. Another shoulder button on the
left side of the controller could be used to indicate the beginning
and/or the end of a movement interval.
In the experiment 80 goal configurations were used. We distin-
guish six classes of configurations of presumed different difficulty.
These classes are graphically illustrated in Figure 2.
Participants first read and signed an informed consent form,
received standardized written instructions for playing the TCG
and the control task, and of each pair one participant was
prepared for the MEG measurements (in total approximately
20 min). After having been given opportunity to ask ques-
tions about the instructions pairs practiced using the controller
(approximately 15 min). In both tasks a task-specific practice ses-
sion of 20 min preceded the 80 recorded trials which took about
45 min, resulting in a total duration of the experiment of about
Participants of a pair were in separate rooms when they played
the game. Each pair played the TCG for the 80 goal configura-
. The ordering of
trials was identical for all pairs of players. Trials were ordered in
such a way that trials became progressively more difficult toward
the end of the experiment. Ta b l e 1 lists the different configura-
tions and their distribution over the 80 trials. The role of sender
and receiver alternated every trial, such that each participant was
sender in 40 trials and receiver in the other 40 trials. The order
of events within a given trial of the TCG game is illustrated
in Figure 1. Participants receiver no performance-based rewards
other than positive and negative feedback (see Figure 1, event 7).
Consistent with previous research on the TCG (De Ruiter et al.,
2010), we found that senders typically develop a communication
strategy in which a part of the sender’s movement is designed
to signal the goal location of the receiver’s token and another
(potentially overlapping) part of the movement is designed to
signal the orientation of the receiver’s token. Such compositional
structure is also characteristic of everyday intentional communi-
cation. The most common strategy for communicating location
is what we refer to as a pause, i.e., the sender’s token spends
relatively more time at the goal location for the receiver’s token
as compared to the time it spends on other squares of the board.
De Ruiter et al. (2010) have previously suggested that such a pause
can be seen as an ostensive signal. This pause signals its own sig-
nalhood by being dysfunctional in the sense that it deviates from
the most efficient way of moving. In a similar vein, apparently
dysfunctional movements were used by sender to signal the goal
orientation for the receiver’s token, but the variation of types of
signals constructed was much larger than for signaling location
(see Appendix B for an overview). The most common strategy
that we observed is what we will call a wiggle.Thisstrategyis
illustrated in Figure 3.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |4
Blokpoel et al. Recipient design: heuristics or perspective taking?
Same token for sender
and receiver, orientation is
the same for both tokens
Different tokens: rect-
angle (sender) and trian-
gle (receiver)
Different tokens: circle
(sender) and triangle (re-
ceiver) - pointing inwards
Different tokens: circle
(sender) and triangle
(receiver) - pointing out-
Same token for sender
and receiver, orientation
Different tokens: circle
(sender) and rectangle
FIGURE 2 | Examples of the six different types of goal configurations. The difficulty of a game is determined by the combinations of tokens; the boards are
ordered in increasing difficulty. In these examples the sender controls the blue token while the receiver controls the red token.
Table 1 | Overview showing the number of times that the different
types of goal configurations occurred and how these were distributed
over the time course of the experiment (indicated by trial number).
Goal configuration
(sender, receiver)
Example Trial number
Same shape, orientation not
2(a) 1–4, 10, 16
Same shape, different
2(b) 5–9, 17
Circle—rectangle 2(c) 11–15, 18–25
Rectangle—triangle 2(d) 26–27
2(e) 28–45, 48–49, 51, 54,
58, 61–77, 80
2(f) 46–47, 50, 52–53, 55–57,
59–60, 78–79
Overall performance on the task (trials resulting in correctly
achieved goal configurations) ranged between 31 and 75 trials
correct (Mean % correct =72%, SD =14%). In section 4.1 we
analyze adaptations made by senders to their own location signals
(i.e., pauses) after receiver errors and in section 4.2, we do the
same but then for senders’ orientation signals (i.e., wiggles). As
explained in section 4.2, we will specifically set out to test if the
nature of the adaptations can be explained by simple heuristic
We analyze changes to the sender’s communicative signal for
the receiver’s location—i.e., the pause on the receiver’s goal
location—after three types of preceding errors on the part of
the receiver (only location error, only orientation error, or both
location and orientation error).4To define our dependent mea-
sure we assume that the longest (most discriminable) pause on
the receiver’s goal location is used by the sender to communicate
location to the receiver. To measure the degree to which a sender
increases or decreases the relative duration (or discriminability)
of the longest pause, we use a normalized measure of change
in duration of pausing on the goal location. We denote
4For consistency with our analysis for orientation signals (see section 4.2) we
base the location signal analysis on the same type of trials, i.e., sender and
receiver have different shaped tokens.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |5
Blokpoel et al. Recipient design: heuristics or perspective taking?
FIGURE 3 | This example movement in trial xillustrates how the
intervals Tx=x
11)that are part of a wiggle movement are
divided over the three types of locations. The goal location is the
bottom-right square, the non-goal locations are the rest of the squares, and
the adjacent location is the middle-right square. This means that
10),and Wx=x
explained in sections 4.1 and 4.2.
this measure as (p:N)and its mathematical definition is
explained next.
We define an ordered list Tt= t
n1)of intervals
between individual moves (i.e., “times spent on locations”) for
the entire movement of a sender’s token in trial t,excluding
the start and end intervals (i.e., τ1and τn). See Figure 3 for an
illustration. We further distinguish two types of locations on the
board: the receiver’s goal location and the receiver’s non-goal
locations (i.e., the rest). The following two sublists of Ttcon-
tain the times that the sender’s token spent on these two types
of locations:
GtTt,suchthatGtcontains all “times spent on” the goal
NtTt,suchthatNtcontains all “times spent on” non-goal
The length of longest pause on the receiver’s goal location is
defined as follows:
It is not the absolute value of ptthat determines the discrim-
inability of the pause for a receiver, but how much longer ptis
as compared to the times spent at other locations. To capture
this discriminability we normalize ptwith respect to the aver-
age time spent at other locations nt(Equation 2). The normalized
5Here the symbol : means normalized with respect to, e.g., a:bis ais
normalized with respect to b.
measure pt:Ntdivides ptby the average time spent on non-goal
locations nt:
Our interest is in how pt:Ntchanges on trial tas a function of
the type of error made by the receiver on trial t2 (recall, sender
and receiver roles switch every trial; therefore the last trial pre-
ceding trial ton which the sender was in the sender role is trial
t2). We therefore define a measure that computes the size of
pt:Ntrelative to the size of pt2:Nt2.Weusethelog
as this minimizes the effect of variability in overall movement
speed and allows us to treat the amount of (normalized) increase
and decrease symmetrically.6,7The resulting measure is defined as
We computed statistics for the measure (p:N)separately for
those trials where the receiver on trial t2placedhistokeninthe
incorrect location but in the correct orientation (location error),
placed his token in the correct location but in the incorrect ori-
entation (orientation error), and placed his token both in the
incorrect location and incorrect orientation (combined error).
In this analysis we ignore trials where on t2 no receiver error
seems to have been made, which would be either because the trial
was successful or because the error seemed to have been due to
the sender rather than the receiver. Appendix A describes in detail
how we filtered those trials.
Ta b l e 2 gives an overview of the relevant statistics after removal
of outliers. As the assumption of normality was violated for the
three distributions of the change in pause measure, we performed
a non-parametric Wilcoxon signed rank test for independent
samples to test whether or not the change in normalized pause
length differed from zero in the three conditions. Here val-
ues larger than 0 correspond to an increase in the length of
the pause and values smaller than 0 correspond to a decrease
in pause length. We found a significant increase in the length of
the pause after a receiver had previously made a location error
(Mean =0.17, Median =0.17; Percentage of trials with increased
pause time =68%, p<0.04), but no significant change after
an orientation error or after a combined error (p>0.67 and
p>0.37, respectively).
We note that it is quite remarkable that we observe this recipi-
ent design effect after location errors despite potential variability
introduced by the intervening trial (t1) on which the sender
was in the receiver role. This suggests that the effect is quite
6The log2transforms the ratio such that it is 0 when there is no change,
and the increase ((p:N)<0) and decrease ((p:N)>0) in speed are
equidistant from 0, i.e., if a pause is two times shorter it has the same distance
from 0 than if it was two times longer.
7If a sender moves from the starting position, to the receiver’s goal location
and then to her own goal location in three moves this measure is not calcu-
lable. Such trials are excluded from all analyses based on this measure, see
Appendix A.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |6
Blokpoel et al. Recipient design: heuristics or perspective taking?
Table 2 | Overview of results for change in pause length on trial tas
compared to trial t2 for the three types of receiver errors.
NMean SD Median pIncrease (%)
Location 37 0.17 0.47 0.17 0 .04 68
Orientation 69 0.02 0.66 0.15 0.67 39
Location and 41 0.10 0 .67 0.14 0.37 56
robust. Of interest is whether the effect is best explained by a
process of perspective taking or by the application of a simple
At first it may seem that our finding that the pause length is
significantly increased after the receiver made a location error,
but not after the receiver made an orientation error, is consis-
tent with the idea that the sender could use a simple rule such
as “if location in error, then increase relative pause.” Such a rule
should indeed be triggered after a location error and not when
that cue is absent (i.e., when only the orientation was in error).
However, such a rule should be triggered always when the rele-
vant cue is present, yet we found no significant increase in pausing
length when the receiver had made an error both in location
and in orientation. It thus seems that senders interpret location
errors as different kinds of misunderstandings on the part of
the receiver than a combined error, causing them to highlight
location after a location error, but not after a combined error.
The reasons for this will become clear after our next analysis
of how senders adapt their signals for orientation after receiver
In the previous section we found a specific adaptation to the
pauses—used to signal a receiver’s goal location—of senders after
a location error. We performed a similar investigation into the
strategies used by senders to signal goal orientation to see whether
or not we would observe specific adaptations to these strate-
gies. Given that a variety of qualitatively different strategies were
used by senders to signal orientation, averaging effects over these
would make the results uninterpretable. Therefore, we decided
to focus on the most common strategy used to signal orienta-
tion: a wiggle. Wiggle strategies were observed in trials in which
sender and receiver tokens were different in shape. There were five
pairs of participants that used different strategies to communicate
orientation and their trials were excluded from this analysis (see
Appendix A).
A wiggle is a (possibly repeated) movement of the sender’s
token from the receiver’s goal location to one of its adjacent
locations on the board, and back again to the receiver’s goal
location. Figure 3 illustrates this movement characteristic. The
majority of pairs used the direction of this movement to com-
municate the orientation of a token (e.g., a wiggle toward
a location above the goal location would indicate that the
receiver’s triangle should “point” up, toward that location). A
few pairs used a different interpretation. They used the exact
number of wiggles to communicate the number of rotations
the receiver needed to perform in order to correctly orient his
We reasoned that, just like there could be a heuristic rule that
states “if location in error, then pause longer”, there could be
a heuristic rule that states “if orientation in error, then wiggle
longer (i.e., perform more wiggles)”, or alternatively, “if orien-
tation in error, then wiggle slower.” Inspection of sender move-
ments revealed that although the number of wiggles performed
varied between participants, it was practically constant within any
given participant across all trials for those participants who used
the “wiggle to point”-strategy (i.e., some participants consistently
wiggled once, some consistently wiggled twice, etc.). Hence, the
number of wiggles lacked the within-sender variability required
for recipient design. Moreover, for participants using the “wiggle
to rotate”-strategy, the number of wiggles was consistently linked
to the number of required rotations. The speed of the wiggle was
variable within participants and a meaningful measure for all wig-
gle strategies, and therefore we set out to test if it was indeed lower
after orientation errors as predicted by the second hypothesized
heuristic. To investigate this we defined a measure of change in
speed of the wiggle, which we denote by w.Wenextexplainhow
this measure is mathematically defined.
As before we use the list notation Tt=t
denote intervals between individual moves (i.e., “times spend on
locations”) for the entire sender movement in trial t,excluding
the start and end intervals (i.e., t1and tn). For our purposes
we consider a particular sublist of Tt, viz., those times spent on
the location adjacent to the receiver’s goal location visited by the
sender’s token during the wiggle:
WtTt,suchthatWtcontains all “times spent on” the
adjacent location that were part of the wiggle.
We defined the speed of a wiggle wtas the average time spent on
the adjacent field as:8
Naturally, the slower the wiggle, the longer the average time spent
on the adjacent field is and the higher wtis. We assume that the
discernibility of the wiggle is independent of the movement speed
of the sender. Therefore, no further normalization of Equation 5
is needed.
Our interest is in how the speed of the wiggle wtchanges on
trial tas a function of the type of error made by the receiver on
trial t2. We define the relative change in speed of the wiggle
8Including the average time spent on the goal location would confound the
measure because of standard pausing behavior on the receiver’s goal location.
9Similar to Equation 4, we use a log2-ratio to preserve symmetry.
10If a sender does not use a wiggle this measure is not calculable. Such trials
are excluded from all analyses based on this measure, see Appendix A.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |7
Blokpoel et al. Recipient design: heuristics or perspective taking?
When wis less than 0 this means that the sender has increased
the speed of the wiggle; when wtis 0 the speed of the wiggle is
unchanged; and when wtis greater than 0 the sender would have
decreased the speed of the wiggle.
We calculated statistics for the measure wseparately for
those trials where the receiver on trial t2madealoca-
tion error, an orientation error, or a combined error. Similar
to the analysis in section 4.1, we ignore trials with the prop-
erty that the error on t2 was not unambiguously due to the
receiver (see Appendix A for details on how these trials were
Ta b l e 3 gives an overview of the relevant statistics after removal
of outliers. We performed a non-parametric Wilcoxon signed
rank test for independent samples to test whether or not the
change in wiggle speed differed from zero in the three condi-
tions. Here values larger than 0 correspond to an increase in
the speed of the wiggle and values smaller than 0 correspond
to a decrease in the speed of the wiggle. We found a signif-
icant increase in the speed of the wiggle after a receiver had
previously made a combination of a location error and an ori-
entation error (Mean =−0.13, Median =−0.10; Percentage of
trials with increased wiggle speed =24%, p<0.03), but no sig-
nificant change after orientation errors alone, or after location
errors alone (p>0.27 and p>0.57, respectively).
Our results do not support the type of heuristic we hypothe-
sized for signal adaptation after orientation errors, as no change
in wiggle speed was observed after those type of errors. Also, no
such change was observed after a location error. Interestingly,
though, if the receiver previously made an error in both loca-
tion and orientation we did observe a change in speed, but this
change was in the opposite direction than we had anticipated.
That is, after a receiver had made a combination of location and
orientation error on trial t2 the speed of the wiggle was sig-
nificantly increased, rather than decreased, by the sender on trial
t. Inspecting the trials on which receivers made these combined
error revealed that the increase in speed of the wiggle served a
purpose in disambiguation.
A typical error made by the receivers on these trials was to
mistake the field adjacent to the goal location with the goal loca-
tion itself. Conditional on this (mistaken) inference, the wiggle
signaled an orientation in the opposite direction of the correct
orientation, in effect causing the receiver to also incorrectly infer
orientation. This situation is sketched in Figure 4. It seems that
upon observing this combined error made by the receiver, the
sender realizes that the misinterpretation was caused by an ambi-
guity in the signal making it difficult for the receiver to discern
Table 3 | Overview of results for change in wiggle speed on trial tas
compared to trial t2 for the three types of receiver errors.
NMean SD Median pIncrease (%)
Location 19 0.07 0.33 0.03 0.57 42
Orientation 39 0.04 0.20 0.01 0.27 54
Location and 25 0.13 0.26 0.10 0.03 24
FIGURE 4 | A frequent error which occurs for the “wiggle to
point”-strategy is that the wiggle and pause are confused. In this
example, the figure on the left depicts the goal configuration and the
sender’s movement. The figure on the right depicts the receiver’s incorrect
which of the two locations visited during the wiggle is the goal
location. The sender then makes the discriminability between
goal location and its adjacent field higher, not by increasing
the relative pause on the goal location (see section 4.1), but by
decreasing the (average) time spent at adjacent field. The context-
sensitive nature of this adaptation of the signal for orientation
(i.e., the adaptation does not occur after an orientation error, but
it does occur after a combined error) suggests genuine perspective
taking on the part of the sender.
We set out to investigate whether or not recipient design in
human communication can be fully explained by simple “fast and
frugal” heuristics (Gigerenzer and Goldstein, 1996; Gigerenzer
and Todd, 1999; Gigerenzer and Brighton, 2009; Shintel and
Keysar, 2009). To this end, we studied trial-to-trial changes made
by players in the context of a communication game. In this game,
players had to mutually achieve a goal configuration that only one
of the players knew (the sender). The sender was to communicate
to her co-player (receiver) the goal location and orientation of the
ses we tested changes in movement characteristics of the sender’s
token movement after a receiver had made one of the following
three possible errors in a preceding trial: the receiver had placed
his token in an incorrect location but correct orientation (location
error); the receiver had placed his token in the correct location but
in an incorrect orientation (orientation error); or the receiver had
placed his token in an incorrect location and incorrect orientation
(combined error).
First, we found that after a receiver had made a location error,
senders tended to pause relatively longer on the receiver’s goal
location. This change in the sender’s movement can be inter-
preted as making the pause more discriminable from the rest
of the movement, making it in effect “clearer” or “less ambigu-
ous” which of the locations on the board was marked by the
sender as the receiver’s goal location. Second, we found no such
increased emphasis on the goal location after an orientation error,
nor after a combined location and orientation error. Particularly,
the absence of an increased pause in the latter case is important,
as it demonstrates that the adaptation is not guided by a sim-
ple heuristic rule such as “if location in error, then pause longer”.
After all, such a rule should also be triggered when both location
and orientation are in error, because its precondition would be
satisfied in that case as well.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |8
Blokpoel et al. Recipient design: heuristics or perspective taking?
It may be argued that the pattern of data could be explained
by a heuristic rule “if location in error and orientation not in
error, then pause longer.” Putting aside that such a heuristic rule
seems to be rather ad hoc, it also violates the condition of fru-
gality that is generally taken as the hallmark of simple heuristics
(Gigerenzer and Goldstein, 1996; Gigerenzer and Todd, 1999;
Gigerenzer and Brighton, 2009). Namely, the extra condition
“orientation is not in error” is here set to function as a context
for when to apply the simple rule “if location in error then pause
longer” and when not. When a heuristics program allows for this
type of context sensitivity there seems to be no bound to the pos-
sible (potentially arbitrary) interactions it can code between cues
and the adaptations they trigger. Mappings encoding context-
sensitivity, potentially even to arbitrary levels, can hardly be said
to be simple in the sense of frugal as they are not ignoring much
The abovementioned ad hoc heuristic would also not be able
to account for another finding we did. After a receiver had made a
combined location and orientation error, senders tended to wig-
gle their token relatively faster. Here, a “wiggle” was a (potentially
repeated) movement of the sender’s token from the receiver’s
goal location to an adjacent field and back to the receiver’s goal
location (see Figure 3 foranillustration).Thewigglemove-
ment was used by some senders to signal the “direction” of the
receiver’s token (the movement direction aligning with one of
the main axis of the receiver’s token, triangle or rectangle) and
by others to signal the “number of rotations” to be performed
by the receiver with his token (senders always knew the start
orientation of the receiver’s token). Inspection of the situations
in which receivers made the combined location and orientation
error revealed that it arose from a confusion on the receiver’s
part between the starting point of the wiggle (the goal loca-
tion) and the end point of the wiggle (the location adjacent to
the goal location). The confusion sometimes caused the receiver
to mistakenly infer that the location adjacent to the goal loca-
tion was the actual goal location; an error in orientation was
then caused as a side-effect by the receiver correctly interpret-
ing the direction of the wiggle movement conditioned on the
erroneously inferred location (see Figure 4 for an illustration).
The increase in the sender’s wiggle speed can be understood as
the sender disambiguating which of the locations visited dur-
ing the wiggle is the goal location and which not, by spend-
ing on average less time on the location adjacent to the goal
Notably, we did not find any adaptation of the speed of the
wiggle after a receiver had made a location error alone, nor after
a receiver had made an orientation error alone. Again, particu-
larly the absence of an increased wiggle speed in the latter case is
important, as it demonstrates that the adaptation is not guided
by a simple heuristic rule such as “if orientation in error, then
wiggle faster. After all, such a rule should also be triggered when
only orientation is in error. Another important observation is the
following: if a sender would have applied the simple “if location
in error, then pause longer,” she would in effect also have dis-
ambiguated which location visited during a wiggle is the goal
location. The fact that senders could have used the simple rule
to achieve the same effect, but do not, suggests that they do
not use such simple rules in this case at all. After all, assum-
ing that they do use the rule when receivers make a location
error, but turn it off when receivers also make an orientation
error and determine some other way to achieve the disambigua-
tion, suggests that the sender would be unnecessarily expending
more than necessary cognitive resources. This is important to
emphasize because one could argue that our specific experimen-
tal paradigm promotes more effortful processing than involved in
typical everyday communication. Even if that were the case, the
results would then suggest that communicators spend the nec-
essary resources for engaging in perspective taking even when
cognitive resources are already heavily taxed by the task and even
when a heuristic would have been sufficient to achieve the same
Rather than postulating ad hoc heuristics, we think our results
are better explained by the hypothesis that the sender employs
a mechanism of perspective taking. On this view, errors on the
part of the receiver tell the sender something about the way in
which the receiver is (mis)interpreting the communicative inten-
tions driving her token movements. In effect, senders may treat a
“combined location and orientation error” as an entirely different
event than simply a location error plus an orientation error. The
sender uses the errors that the receiver makes to form hypothe-
ses about the “why” of the receiver’s misinterpretations, and uses
these hypotheses to adjust her movements on subsequent trials.
preted as a form of “clarification” or “disambiguation, where
the sender has realized what ambiguity had caused the receiver’s
earlier mistake and she adjusts the signal to ensure the same mis-
take is prevented from then on. The context-sensitive nature of
these disambiguations suggest that they are not rote rules, but
quite sophisticated forms of fine tuning. For instance, imagine
the receiver had placed the token on some different location than
the goal location. On a subsequent trial the sender then pauses
longer on the goal location to distinguish it more clearly from
all the other locations she visits. Yet, if the receiver had con-
fused the location adjacent to the goal location visited during a
wiggle for the goal location itself, then the sender wiggles faster
to distinguish more clearly the goal location from the adjacent
In sum, the recipient design that we observe in the context
of our communicative game is not straightforwardly explained
by simple heuristics, yet it is parsimoniously explained by per-
spective taking. Of course, we cannot rule out complex heuristics
for recipient design. However, a heuristics account that allows
for arbitrary interactions between cues runs into the same com-
putational intractability problem that motivated the critique of
a perspective taking mechanism for recipient design in the first
place. Namely, the number of possible combinations of cues
grows exponentially in the number of possible cues. If rules can
be triggered by arbitrary combinations of cues then an expo-
nential number of rules will need to be stored. Such a heuristic
model does not obviously scale to situations of real-world com-
plexity with more than a few possible cues (cf. Gigerenzer, 2008),
because an exponential number of rules either needs astronomi-
cal amounts of space to be stored, or—if the list of rules is stored
in compressed form—it takes astronomical amounts of time to
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |9
Blokpoel et al. Recipient design: heuristics or perspective taking?
find the right rule to apply (cf. Newell, 2005; van Rooij et al.,
Does this mean that recipient design is computationally
not believe that. The fact that communicators in the game engage
in recipient design shows that they have some efficient way of
doing so. Moreover, the nature of the signal adaptations sug-
gests that they draw upon a mechanism of perspective taking,
suggesting that the communicators had some efficient way to
invoke and use such a mechanism to their ends. As some of
us have extensively argued elsewhere (van Rooij, 2008; Blokpoel
et al., 2010), intractability of a cognitive model should be taken
as an indication that the model has so far failed to specify the
right set of situational constraints under which the modeled cog-
nitive capacity is displayed. Hence, theories of communication
in general, and recipient design in particular, must incorpo-
rate a set of situational constraints that allows such theories to
explain how perspective taking computations—that are otherwise
intractable—can be efficiently performed under the conditions in
which we observe it. Specifying such constraints is best considered
a long-term research program, though some promising initial
theoretical results have been obtained (Blokpoel et al., 2011; van
Rooij et al., 2011).
We close by reflecting on how our research may have implica-
tions for social neuroscience and social robotics, and the inter-
action between these fields. First of all, our findings suggest
that the TCG game can be a fruitful empirical testing ground
for neural theories of perspective taking. We have shown that
trial-to-trial adaptations made by senders in this game seem to
directly involve perspective taking mechanisms. The results of
this study provide a quantitative, sensitive, and implicit index
of perspective taking that can form the basis of a number of
neurocognitive investigations. For instance, in contrast to tra-
ditional approaches to the study of the neural implementation
of Theory of Mind (Fletcher et al., 1995; McCleery et al., 2011)
the current index of perspective taking can provide a large num-
ber of independent read-outs (trials), ensuring sensitivity; and it
is independent from verbal reports, avoiding to rely on linguis-
tic performance. These characteristics make the current index of
perspective taking particularly suitable for studying mentalizing
abilities (and their cerebral implementation) in populations char-
acterized by large variability in performance and limited access
to meta-linguistic knowledge (e.g., children, patients with Autism
Spectrum Disorders).
Second, our findings seem to clarify the nature of a con-
siderable challenge for the design of socially interactive robotic
agents interacting with humans in real-world settings. In Artificial
Intelligence there are longstanding difficulties in devising com-
putational mechanisms for context-sensitive processes (such as
perspective taking) that are computationally tractable—i.e., that
can scale from toy domains to real-world situations in terms of
computational speed (Pylyshyn, 1987; Haselager, 1997; Dreyfus,
2007). Yet, the increasing use of robots and other artificial agents
in daily life (e.g., in offices, care-giving institutions, shopping
malls, and musea) will require at least a reasonable functional
implementation of a recipient design capacity. Imagine a robot
guide in a large museum or city taking a tourist on a tour that
may last for an entire afternoon or day. To ensure the robot’s
efficacy it seems necessary that it can adapt to individual com-
municative characteristics of the tourist so as to avoid huge adap-
tations on the tourist’s part. If fast and frugal heuristics would
suffice for this interactive task then computational tractability
would be guaranteed. However, if we are right in our sugges-
tion that fast and frugal heuristics will not suffice to emulate
the level of adaptation characteristic for human communication,
then the computational complexities associated with perspec-
tive taking—or equivalent contex-sensitive processing—will have
to be dealt with head-on by designers of socially interactive
These observations also suggest a way in which social neuro-
science and social robotics can actually directly inform each other.
On the one hand, social neuroscience can inform social robotics:
Given (1) the apparent need for social robots to engage in (sim-
ulations of) perspective taking in order to achieve human-level
recipient design, (2) the evident ability of humans for effective
and efficient recipient design recipient design during life interac-
tions, and (3) the failing of AI so far to produce computational
models of perspective taking that scale to the real world, social
robotics may do well to look to social neuroscience for com-
putational hypotheses about how the human brain implements
the necessary perspective taking mechanisms. On the other hand,
social robotics can also inform social neuroscience, e.g., by mak-
ing the latter field (more) aware of the challenges of making
computational models that can properly scale outside the toy
domains studied in the lab. After all, for computational models
hypothesized in social neuroscience to explain everyday human
social interactions they should minimally be scalable, and hence
tractable. The scalability problem known so well to researchers in
AI and robotics is often not considered or even noticed in social
neuroscience. This is possibly because experiments in social neu-
roscience are performed in the context of simple lab tasks, and
hence computing the predictions made by said models for the
lab setting may still be feasible. Yet, the scalability of such mod-
els can be critically tested by empirical analysis and implementing
them in robots in order to test whether or not they yield similar
levels of performance in situations of real-world complexity (sim-
ilar to the real-time adaptive communicative actions performed
by humans operating in everyday settings). In this way, social
robotics can help constrain computational theories of recipient
design in social neuroscience, viz., by providing scalability as a
theoretical constraint. Given our finding that perspective taking
may be a necessary component of recipient design in humans, an
awareness of the computational complexity associated with com-
putational models of perspective taking may be more useful in the
study of communication than previously thought.
The authors thank the reviewers for useful comments on an ear-
lier version of this paper. Mark Blokpoel was supported by a
DCC PhD grant awarded to Iris van Rooij, Ivan Toni, and Pim
Haselager. Ivan Toni was supported by VICI grant #453-08-002
from NWO.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |10
Blokpoel et al. Recipient design: heuristics or perspective taking?
Blokpoel, M., Kwisthout, J., van der
Weide, T. P., and van Rooij, I.
(2010). “How action understand-
ing can be rational, Bayesian and
tractable,” in Proceedings of the
32nd Annual Conference of the
Cognitive Science Society, ed S. O. R.
Catrambone (Austin, TX: Cognitive
Science Society), 1643–1648.
Blokpoel, M., Kwisthout, J., Wareham,
T., Haselager, P., Toni, I., and van
Rooij, I. (2011). “The computa-
tional costs of recipient design and
intention recognition in communi-
cation,” in Proceedings of the 33rd
Annual Conference of the Cognitive
Science Society, eds L. Carlson, C.
Holscher, and T. Shipley (Austin,
TX: Cognitive Science Society),
Breazeal, C. (2002). Designing Sociable
Robot s. Cambridge, MA: MIT Press.
Camerer, C. (2003). Behavioral
Game Theory: Experiments in
Strategic Interaction. New Jersey, NJ :
Princeton University Press.
Clark, H. (1996). Using language.
Cambridge, UK: Cambridge
University Press.
Clark, H., and Carlson, T. (1982).
Hearers and speech acts. Language
58, 332–373.
De Ruiter, J., Noordzij, M., Newman-
Norlund, S., Hagoort, P., and Toni,
I. (2007). “On the origin of inten-
tions,” in Sensorimotor Foundations
of Higher Cognition, eds Y. Rosseti,
P. Haggard, a nd M. Kawato (Oxford:
Oxford University Press), 601–618.
De Ruiter, J., Noordzij, M., Newman-
Norlund, S., Newman-Norlund, R.,
Hagoort, P., Levinson, S., and Toni,
I. (2010). Exploring the cognitive
infrastructure of communication.
Interact. Stud. 11, 51–77.
Dreyfus, H. L. (2007). Why
Heideggerian AI failed and how
fixing it would require making it
more Heideggerian. Philos. Psychol.
20, 247–268.
Ep le y, N. , Ke ys ar, B ., Van Bov en , L. ,
and Gilovich, T. (2004). Perspective
taking as egocentric anchoring and
adjustment. J. Pers. Soc. Psychol. 87,
Feiler, L., and Camerer, C. F. (2010).
Code creation in endogenous
merger experiments. Econ. Inq. 48,
Fletcher, P., Happé, F., Frith, U., Baker,
S., Dolan, R., Frackowiak, R., and
Frith, C. (1995). Other minds in
the brain: a functional imaging
study of “theory of mind” in
story comprehension. Cognition 57,
Galantucci, B. (2005). An experimen-
tal study of the emergence of human
communication systems. Cogn. Sci.
29, 737–767.
Galantucci, B. (2009). Experimental
semiotics: a new approach for
studying communication as a form
of joint action. Top. C og n . S c i. 1,
Galantucci, B., and Garrod, S. (2010).
Experimental semiotics: a new
approach for studying the emer-
gence and the evolution of human
communication. Interact. Stud. 11,
Galati, A., and Brennan, S. (2010).
Attenuating information in spoken
communication: for the speaker, or
for the addressee? J. Mem. Lang. 62,
Gigerenzer, G. (2008). Why heuristics
work. Perspect. Psychol. Sc i. 3, 20–29.
Gigerenzer, G., and Brighton, H.
(2009). Homo heuristicus:
why biased minds make better
inferences. Top. C ogn. S c i . 1,
(1996). Reasoning the fast and
frugal way: models of bounded
rationality. Psychol. Rev. 103,
Gigerenzer, G., and Todd, P. (1999).
Simple Heuristics That Make Us
Smart. New York, NY: Oxford
University Press.
Green, S., Billinghurst, M., Chen, X.,
and Chase, J. (2008). Human-robot
collaboration: a literature review
and augmented reality approach in
design. Int. J. Adv. Robot. Syst. 5,
Grice, H. (1975). “Logic and conver-
sation,” in Syntax and Semantic 3:
Speech Acts, eds P. Cole and J.
Morgan (New York, NY: Academic
press), 41–58.
Grice, H. (1989). Studies in the Way
of Words. Cambridge, MA: Harvard
University Press.
Haselager, W. (1997). Cognitive Science
and Folk Psychology: The Right
Frame of Mind. London, UK: Sage.
Horton, W., and Keysar, B. (1996).
When do speakers take into account
common ground? Cognition 59,
Keysar, B., Barr, D., and Horton, W.
(1998). The egocentric basis of
language use: insights from a pro-
cessing approach. Curr. D ir. Psychol.
Sci. 7, 46–50.
Levelt, W. (1989). Speaking:
From Intention to Articulation.
Cambridge, MA: MIT Press.
Marsh, B. (2002). Heuristics as social
tools. New Ideas Psychol. 20, 49–57.
McCleery, J., Surtees, A., Graham,
K., Richards, J., and Apperly, I.
(2011). The neural and cognitive
time course of theory of mind. J.
Neuro sci . 31, 12849–12854.
Newell, B. (2005). Re-visions of ratio-
nality? Tre n ds C og n . Sci . 9, 11–15.
Newman-Norlund, S., Noordzij, M.,
Newman-Norlund, R., Volman, I.,
Ruiter, J., Hagoort, P., and Toni,
I. (2009). Recipient design in tacit
communication. Cognition 111,
Noordzij, M., Newman-Norlund,
S., De Ruiter, J., Hagoort, P.,
Levinson, S., and Toni, I. (2009).
Brain mechanisms underlying
human communication. Front.
Hum. Neurosci. 3, 14–27. doi:
Pylyshyn, Z. (ed.). (1987). The Robot’s
Dilemma: The Frame Problem
in Artificial Intelligence. Vo l . 4 ,
Norwood, NJ: Ablex Publishing
Sacks, H., Schegloff, E., and Jefferson,
G. (1974). A simplest systemat-
ics for the organization of turn-
taking for conversation. Language
50, 696–735.
Imai, M., Ishiguro, H., and Hagita,
N. (2009). “How to approach
humans?: strategies for social
robots to initiate interaction,” in
Proceedings of the 4th ACM/IEEE
International Conference On Human
Robot Interaction,HRI ’09, (New
York, NY: ACM), 109–116.
Scott-Phillips, T., Kirby, S., and Ritchie,
G. (2009). Signalling signalhood
and the emergence of communica-
tion. Cognition 113, 226–233.
Selten, R., and Warglien, M. (2007).
The emergence of simple languages
in an experimental coordination
game. Proc.Natl.Acad.Sci.104,
Shintel, H., and Keysar, B. (2009). Less
is more: a minimalist account of
joint action in communication. Top .
Cogn. Sc i. 1, 260–273.
Shiomi, M., Kanda, T., Koizumi, S.,
Ishiguro, H., and Hagita, N. (2007).
“Group attention control for
communication robots with wizard
of oz approach,” in Proceedings
of the ACM/IEEE International
Conference on Human-Robot
Interaction, HRI ’07, (New York,
NY: ACM), 121–128.
Stolk, A., Verhagen, L., Schoffelen,
J-M., Oostenveld, R., Blokpoel, M.,
Hagoort, P., van Rooij, I. and Toni,
I. (2012) Neural mechanisms of
human communicative innovations.
42nd annual meeting of the Society
for Neuroscience, New Orleans.
Tversky, A., and Kahneman, D. (1974).
Judgment under uncertainty:
Heuristics and biases. Science 185,
van Rooij, I. (2008). The tractable cog-
nition thesis. Cogn . Sci. 32, 939–984.
van Rooij, I., Kwisthout, J., Blokpoel,
M., Szymanik, J., Wareham, T., and
Toni, I. (2011). Intentional com-
munication: computationally easy
or difficult? Front. Hum. Neurosci.
5, 1–18. doi: 10.3389/fnhum.2011.
van Rooij, I., Wright, C., and Wareham,
H. T. (2012). Intractability and
the use of heuristics in psycho-
logical explanations. Synthese 187,
Weber, R. A., and Camerer, C. F. (2003).
Cultural conflict and merger failure:
an experimental approach. Manag.
Sci. 49, 400–415.
Conflict of Interest Statement: The
authors declare that the research
was conducted in the absence of any
commercial or financial relationships
that could be construed as a potential
conflict of interest.
Received: 27 January 2012; accepted:
17 August 2012; published online: 25
September 2012.
Citation: Blokpoel M, van Kesteren M,
Stolk A, Haselager P, Toni I and van
Rooij I (2012) Recipient desig n in human
communication: simple heuristics or per-
spective taking? Front. Hum. Neurosci.
6:253. doi: 10.3389/fnhum.2012.00253
Copyright © 2012 Blokpoel, van
Kesteren, Stolk, Haselager, Toni and
van Rooij. This is an open-access article
distributed under the terms of the
Creative Commons Attribution License,
which permits use, distribution and
reproduction in other forums, provided
the original authors and source are cred-
ited and subject to any copyright notices
concerning any third-party graphics etc.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |11
Blokpoel et al. Recipient design: heuristics or perspective taking?
In this appendix we define which trials twere excluded from the
analyses in Section 4. The motivation behind the exclusion is that
we were interested to test adaptations made by senders to their
signals on trial tafter a receiver error on a previous trial t2,
the rationale being that this gives us a handle on recipient design
by the sender as a function of type of error on the receiver’s part.
For this purpose we want to only include trials tin the analyses
with the property that the error on t2 can be unambiguously
attributed to a misinterpretation by the receiver. After all, if the
miscommunication that occurred on trial t2wasduetoan
error made by the sender (e.g., she misremembered the goal con-
figuration or made a mistake in the execution of her movements)
then it is problematic to interpret changes made to the signal on
time tas the result of a form of recipient design. For instance, any
changes observed in trials following sender execution errors may
simply be an artifact of the fact that the movement was not as
planned or intended on trial t2whereasontit is.
Ta b l e A 1 gives an overview of the types of trials that were
excluded from the analyses for this reason. In the remainder of
this section we define and explain a set of criteria that we used to
judge whether a signal was unintelligible.
We list the criteria we used to judge if a sender’s commu-
nicative signal was unintelligible, dependent on the type of error
of the receiver. For instance, the type of error of the receiver
can be a location error in which case we looked at whether the
communication of location by the sender was (un)intelligible.
Besides criteria for when the receiver’s location was incorrect,
we also list criteria for when orientation was incorrect. When
both the receiver’s orientation and location were incorrect we
checked the criteria from both lists. To code which type of error
the receiver made on trial t2, we checked if the final loca-
tion and orientation of the receiver’s token matched with the goal
Dependent on the type of error of the receiver, we judge an
error as unintelligible (and thus a sender error) when:
a. The location of the receiver’s token is not correct (i.e., not on
the goal location), and:
Table A1 | Overview of which types of trials t2 were excluded from
the analyses, based on our definition of intelligible signals.
Ended in goal configuration
Sender Receiver Sender was Excluded from
intelligible analyses
Ye s Ye s
χχYe s N o
χYe s N o
χYe s Ye s
No Yes
χχNo Yes
χNo Yes
χNo Yes
1. There was no visit to the receiver’s goal location, or;
2. The experimenter (knowing both the goal configura-
tion and the strategies used by the sender) could
not recognize any movement characteristic signaling
b. The orientation of the receiver’s token is not correct, and:
1. The signal is not based on any identified strategies
(see De Ruiter et al., 2010 or Appendix B), or;
2. The signal is based on a strategy that deviates from a
previously mutually agreed strategy, or;
3. There is an error in the execution of the strategy.
c. The communicative signal corresponds to a different goal con-
figuration than was presented to the sender (suggesting the
sender has forgotten the goal configuration).
Note that those trials which were not excluded are not uncon-
ditionally included. For a trial tto be included in the analysis
of pauses, means that for both trial t2andtthe pt:Ntin
Equation 3 needs to be calculable (i.e., besides a pause on the goal
location, also at least one other field needs to be visited). For the
analysis of wiggles, a trial tis included when a wiggle on trial t2
and tis done, as required for computing wtin Equation 5. To be
clear, communicative success or failure on trial titself was not a
criterium for selection.
This section illustrates the variety of strategies we observed, and
their relative frequencies, using the most common trials (circle-
triangle, pointing inwards; see Figure 2E and Tab l e B 1 ). These
strategies were all used by senders to communicate orientation
in addition to pausing behavior, that they used to communicate
Wiggle to point
A wiggle starts with the sender’s token at the goal location of
the receiver’s token, the sender then moves to the adjacent field
and then goes back to the receiver’s goal location. The adjacent
field is the field to which the triangle “points” (e.g., the red
triangle in Figure 2A “points” up). The number of wiggles to
signal orientation varied over pairs between one and four. An
Table B1 | An overview of the various strategies that were observed
in the experiment, the number of pairs of participants using them,
and the average success-rate per strategy.
Strategy # Pairs % correct N
Wiggle to point 15 70.4 615
Exit point 4 96.3 164
Wiggle to rotate 3 78 123
Exit from center 11 75.6 141
The total number of pairs is 23 (P =23), and each strategy was observed N =
P×41 times, where 41 is the number of circle-triangle trials with inward pointing
triangle (see Figure 2E).
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |12
Blokpoel et al. Recipient design: heuristics or perspective taking?
example of a wiggle, consisting of two repetitions, is illustrated
in Figure 3.
Exit point
After indicating the location with a pause, the sender moves
to an adjacent field and then moves to her own location. The
adjacent field to which the sender exited the receiver’s goal
location indicates the direction in which the receiver’s triangle
Wiggle to rotate
The number of wiggles starting from the receiver’s goal location
indicates the number of rotations of the receiver’s triangle. At
the start the triangle always points up and any rotation of the
token is done clockwise which provides a one to one mapping
from number of wiggles to number of rotations. Zero wiggles
then means no rotation, one wiggle means “point right”, two
wiggles means “point down”, and three wiggles means “point
Exit from center
Senders communicate the goal orientation of the receiver by
exiting their starting location in the center of the board in a
particular direction. This direction indicates which way the
receiver’s triangle points. After this first move the sender moves
to and pauses at the receiver’s goal.
Some of these strategies were also observed in other trials.
The “wiggle to point”- and “wiggle to rotate”-strategies, for
example, were also sometimes observed in trials that required
senders to communicate orientation in the trial types depicted
in Figures 2B2F. Note that not all strategies reported here
were analyzed, but they were used in the trial selection process
described in Appendix A.
Frontiers in Human Neuroscience September 2012 | Volume 6 | Article 253 |13
... We indexed these dynamic adjustments in communicative proficiency as the rate of change in Efficiency over trials (Efficiency Rate). In this study, participants' abilities to comprehend novel communicative actions were quantified in a controlled and validated experimental setting; the Tacit Communication Game (TCG; Fig. 1A) ( Blokpoel et al., 2012;Newman-Norlund et al., 2009;de Ruiter et al., 2010). In this interactive task, two players are asked to recreate a spatial configuration of two simple geometrical shapes (one for the Communicator and one for the Addressee) on a digital game board. ...
... The only means the players had to communicate is with an unconventional tool, namely by moving their simple geometrical shape. Different pairs of participants solve these novel communicative problems in different ways ( Blokpoel et al., 2012;de Ruiter et al., 2010), an indication that pairs mutually converge on a common solution from a potentially infinite set of possible arbitrary solutions (van Rooij et al., 2011). This feature of the task makes it possible to isolate adjustments in communicative performance driven by the recent history of interactions between participants from overall variations in communicative performance, and their interaction with stable cognitive traits. ...
... Accordingly, we used this psychometric index to characterize the inferential processes supported by the right pSTS during referential communication. Namely, Addressees with high Raven's scores might more readily use abstract relations to quickly generate novel analogical mappings between observed actions and their communicative intentions ( Blokpoel et al., 2012;Carpenter, Just, & Shell, 1990;Volman et al., 2012). We reasoned that if rTMS over right pSTS influences Addressees' ability to quickly grasp novel communicative meanings according to recent communicative interactions (see Hypothesis #3 above), then Addressees with high Raven's scores might be particularly impaired by rTMS-induced cerebral alterations. ...
Full-text available
Wadge et al. demonstrate that individuals with ASD are able and motivated to generate and modify intelligible communicative behaviors for the benefit of a communicative partner. However, they struggle to align the meaning of those behaviors with their partner when meaning relies on knowledge derived from the shared communicative history.
Full-text available
Humans and animals make inferences about the world under limited time and knowledge. In contrast, many models of rational inference treat the mind as a Laplacean Demon, equipped with unlimited time, knowledge, and computational might. Following Herbert Simon's notion of satisficing, this chapter proposes a family of algorithms based on a simple psychological mechanism: one-reason decision making. These fast-and-frugal algorithms violate fundamental tenets of classical rationality: It neither looks up nor integrates all information. By computer simulation, a competition was held between the satisficing "take-the-best" algorithm and various "rational" inference procedures (e.g., multiple regression). The take-the-best algorithm matched or outperformed all competitors in inferential speed and accuracy. This result is an existence proof that cognitive mechanisms capable of successful performance in the real world do not need to satisfy the classical norms of rational inference.
Full-text available
Human referential communication is often thought as coding-decoding a set of symbols, neglecting that establishing shared meanings requires a computational mechanism powerful enough to mutually negotiate them. Sharing the meaning of a novel symbol might rely on similar conceptual inferences across communicators or on statistical similarities in their sensorimotor behaviors. Using magnetoencephalography, we assess spectral, temporal, and spatial characteristics of neural activity evoked when people generate and understand novel shared symbols during live communicative interactions. Solving those communicative problems induced comparable changes in the spectral profile of neural activity of both communicators and addressees. This shared neuronal up-regulation was spatially localized to the right temporal lobe and the ventromedial prefrontal cortex and emerged already before the occurrence of a specific communicative problem. Communicative innovation relies on neuronal computations that are shared across generating and understanding novel shared symbols, operating over temporal scales independent from transient sensorimotor behavior.
Many decisions are based on beliefs concerning the likelihood of uncertain events such as the outcome of an election, the guilt of a defendant, or the future value of the dollar. Occasionally, beliefs concerning uncertain events are expressed in numerical form as odds or subjective probabilities. In general, the heuristics are quite useful, but sometimes they lead to severe and systematic errors. The subjective assessment of probability resembles the subjective assessment of physical quantities such as distance or size. These judgments are all based on data of limited validity, which are processed according to heuristic rules. However, the reliance on this rule leads to systematic errors in the estimation of distance. This chapter describes three heuristics that are employed in making judgments under uncertainty. The first is representativeness, which is usually employed when people are asked to judge the probability that an object or event belongs to a class or event. The second is the availability of instances or scenarios, which is often employed when people are asked to assess the frequency of a class or the plausibility of a particular development, and the third is adjustment from an anchor, which is usually employed in numerical prediction when a relevant value is available.
We use laboratory experiments to explore merger failure due to conflicting organizational cultures. We introduce a laboratory paradigm for studying organizational culture that captures several key elements of the phenomenon. In our experiments, we allow subjects in "firms" to develop a culture, and then merge two firms. As expected, performance decreases following the merging of two laboratory firms. In addition, subjects overestimate the performance of the merged firm and attribute the decrease in performance to members of the other firm rather than to situational difficulties created by conflicting culture.
Speakers tend to attenuate information that is predictable or repeated. To what extent is this done automatically and egocentrically, because it is easiest for speakers themselves, and to what extent is it driven by the informational needs of addressees? In 20 triads of naive subjects, speakers told the same Road Runner cartoon story twice to one addressee and once to another addressee, counterbalanced for order (Addressee1/Addressee1/Addressee2 or Addressee1/Addressee2/Addressee1). Stories retold to the same (old) addressees were attenuated compared to those retold to new addressees; this was true for events mentioned, number of words, and amount of detail. Moreover, lexically identical expressions by the same speaker were more intelligible to another group of listeners when the expressions had been addressed to new addressees than when they had been addressed to old addressees. We conclude that speakers’ attenuating of information in spontaneous discourse is driven at least in part by addressees. Such audience design is computationally feasible when it can be guided by a “one-bit” model (my audience has heard this before, or not).
An important aspect of human sociality is our ability to understand the actions of others as being goal-directed. Recently, the now classic rational approach to explaining this ability has been given a formal incarnation in the Bayesian Inverse Planning (BIP) model of Baker, Saxe, and Tenenbaum (2009). The BIP model enjoys considerable empirical support when tested on ‘toy domains’. Yet, like many Bayesian models of cognition, it faces the charge of computational intractability: i.e., the computations that the model postulates may be too resource demanding for the model to be scalable to domains of realworld complexity. In this paper, we investigate ways in which the BIP model can possibly parry the charge. We will show that there are specific conditions under which the computations postulated by the model are tractable, despite the model being rational and Bayesian.