The effects of recursive communication

royalsocietypublishing.org/journal/rspb
dynamics on belief updating
Niccolò Pescetelli1,2 and Nick Yeung2
1
Max Planck Institute for Human Development, 94 Lentzeallee, Berlin 14195, Germany
2
Department of Experimental Psychology, University of Oxford, Anna Watts Building,
Research Radcliffe Observatory Quarter, Woodstock Road, Oxford OX2 6GG, UK

Cite this article: Pescetelli N, Yeung N. 2020 NP, 0000-0002-8826-2202; NY, 0000-0003-1905-2129
The effects of recursive communication
Many social interactions are characterized by dynamic interplay, such that
dynamics on belief updating. Proc. R. Soc. B
individuals exert reciprocal influence over each other’s behaviours and
287: 20200025. beliefs. The present study investigated how the dynamics of reciprocal
http://dx.doi.org/10.1098/rspb.2020.0025 influence affect individual beliefs in a social context, over and above the
information communicated in an interaction. To this end, we developed a
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

simple social decision-making paradigm in which two people are asked

to make perceptual judgments while receiving information about each
Received: 15 January 2020 other’s decisions. In a Static condition, information about the partner only
Accepted: 1 July 2020 conveyed their initial, independent judgment. However, in a Dynamic con-
dition, each individual saw the evolving belief of their partner as they learnt
about and responded to the individual’s own judgment. The results indi-
cated that in both conditions, the majority of confidence adjustments
were characterized by an abrupt change followed by smaller adjustments
Subject Category: around an equilibrium, and that participants’ confidence was used to arbi-
Behaviour trate conflict (although deviating from Bayesian norm). Crucially, recursive
interaction had systematic effects on belief change relative to the static
Subject Areas: baseline, magnifying confidence change when partners agreed and reducing
behaviour, cognition, neuroscience confidence change when they disagreed. These findings indicate that during
dynamic interactions—often a characteristic of real-life and online social
contexts—information is collectively transformed rather than acted upon
Keywords:
by individuals in isolation. Consequently, the output of social events is not
belief change, escalation, social interaction, only influenced by what the dyad knows but also by predictable recursive
confidence, decision-making and self-reinforcing dynamics.

Author for correspondence:

Niccolò Pescetelli 1. Introduction
e-mail: pescetelli@mpib-berlin.mpg.de People very often rely on others’ judgments to inform their decisions. We ask
our family and friends for their opinions on a particular job we are planning
to apply for, the colour of a new dress to buy, or the political candidate to
favour. When information is costly to acquire, aggregating multiple samples
of imprecise information is often more effective than spending time improving
a single sample [1]. Aggregating beliefs from diverse and independent sources
has long been known to improve judgment reliability [2,3]. How people incor-
porate others’ judgments and advice into their decisions has been of central
interest in social and organizational psychology, and more recently in compu-
tational social science [4,5]. One particularly fruitful approach to this question
is the judge–advisor system paradigm [4], in which participants (the ‘judges’)
form independent beliefs, then are asked to revise their initial judgment after
seeing the belief of one or more advisors. This approach provides precise con-
trol over the information provided by the advice to participants, as well as
precisely defined and distinct moments at which an initial decision is formed,
advice is presented, and the judge’s belief updated. Previous studies have high-
Electronic supplementary material is available lighted the factors affecting how personal judgments are influenced by others’
online at https://doi.org/10.6084/m9.figshare. beliefs, like confidence [6,7], power [8], expertise [9], agreement among different
sources [10], whether the self is involved [11], and how distant is the advice
c.5054653.
from the judge’s own view [12,13].

© 2020 The Author(s) Published by the Royal Society. All rights reserved.
 (a) (b) Alice sees ... Bea sees ... 2

stimulus

royalsocietypublishing.org/journal/rspb
Alice

social exchange confidence rating

100% 100% 100% 100%

computer
occluder

Left Right Left Right

time
dynamic static static dynamic

feedback
Bea Alice: incorrect Alice: incorrect
Bea: correct Bea: correct

Proc. R. Soc. B 287: 20200025

Figure 1. (a) Participants sat on opposite sides of a wooden occluder and used one monitor, one keyboard, and one mouse each, controlled by the same computer.
(b) After seeing a perceptual stimulus (stimulus), participants made a judgment independently from one another (confidence rating). During the social exchange
part, participants saw on their own scale their partner’s initial belief (Static condition) or their partner’s evolving belief (Dynamic condition) on alternating blocks.
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

Bidirectional black arrows along the confidence scale represent real-time continuous movement along the scale. The scale used in the actual task had 50 levels per
interval. Finally, during the feedback stage, participants received feedback on their binary accuracy (correct versus incorrect) and earnings (not shown) at the end of
the trial (feedback). As shown in the figure, the feedback made clear which participant was correct or incorrect on a given trial by using the user-names that
participants had to provide at the beginning of the experiment. Notice that although task difficulty was titrated to each participant’s performance, correct answers
(LEFT/RIGHT) were identical across participants. (Online version in colour.)

However, in daily life, advisors’ beliefs rarely remain inde- polarization [21–23]. Recently, attention has been devoted to
pendent from each other or from the advisee’s own judgment, understand belief polarization in more realistic contexts, like
as is typically implemented in judge–advisor system studies. networks and social media platforms, due to the far-reaching
Instead, information is collectively transformed and manipu- consequences that these phenomena have in society [24–26].
lated until it converges to a decision. For example, you might Bail et al. [27] found that being exposed to disagreeing opinions
say to the friend recommending the yellow dress that you actu- can polarize individual beliefs even further. This is counter-
ally don’t like yellow, you like green; to which she replies that intuitive considering that, from a Bayesian perspective,
she thought green was kind of nice too and when, at that point, disagreeing evidence should always decrease the strength of
a member of staff at the shop interrupts to say that green one’s own beliefs. This finding may suggest that dynamic inter-
is going to be next year’s fashionable colour, a purchase is actions among individuals might differ from static social
made. Here, the processes of decision, advice, and update exposure as often studied in the laboratory. Little is known
take place all at the same time, and the line between judge about the direct comparison between uni- versus bidirectional
and advisors blurs such that all partners affect each other’s social exchange. To address this issue, in the present study we
beliefs without a clear distinction between cause and effect. investigate the effect of recursive social dynamics—typically
Thus, an important class of social interactions involves bidirec- seen in free social interactions—using a carefully controlled
tional information flow, as in face-to-face conversations and judge–advisor paradigm. We investigate differences between
online chats, raising the possibility of recursive dynamics static versus dynamic interactions in social decision-making,
emerging among people’s beliefs. Even online, where com- specifically in a task in which the information available to
munication happens in discrete rather than continuous steps participants to make a decision was kept constant across
(like tweets or replies), interactions allow for a bidirectional interactions, thus allowing us to characterize the specific
back and forth between parties who are thus able to mutually impact of interaction dynamics themselves, above and
adjust according to each other’s positions. We characterize beyond the information brought to the interaction by the
this class of recursive interactions as ‘dynamic’, as opposed individuals involved.
to ‘static’ social exposure. While the latter has typically been In our study, pairs of participants performed simple
investigated in micro-scale experimental studies of advice- perceptual judgements in parallel (figure 1). On each trial,
taking, the former has been investigated within the opinion participants first made private, independent judgments
dynamics research program largely adopting macro-level (Confidence rating) about which of two boxes contained more
analytic and simulation-based approaches [14,15]. dots [28] (Stimulus). They then were asked to continuously
As such, a full picture of social decision-making may monitor and update their judgement, expressed on a semi-
require analysis of interaction dynamics, moving beyond the continuous confidence scale, based on observing the other
very valuable simplification of a participant working alone as participant’s belief on the same scale (Social exchange). Each
an observer of social information (cf. [16–18]). Consistent participant was thus both judge and advisor. Importantly, we
with this reasoning, previous work found that individuals compared conditions characterized by static information
performing a task together become more confident and align exchange—in which participants only saw their partner’s initial
their linguistic expressions when they are allowed face-to- judgment—with conditions characterized by dynamic infor-
face verbal interaction, but not when confidence is shared with- mation exchange—in which participants saw their partner’s
out verbal interaction [19,20]. This finding relates back to moment-by-moment belief change and thus how they
earlier work in social psychology on ‘risky shifts’ and group themselves influenced their partner’s views. By keeping
 decision-relevant information constant in the two conditions, informed about their partner’s belief. At this point, belief updates 3
we isolate the contrast between static and dynamically evolving were recorded continuously. In contrast to the standard judge––
advisor system paradigm [4], where belief updates happen in a

royalsocietypublishing.org/journal/rspb
social information, and thus characterize the way information is
shared and collectively transformed between individuals. Based single step, here we recorded beliefs continuously as they evolved
over time: the mouse x-position along the confidence scale was
on previous findings—which show that face-to-face and other
recorded every 200 ms for 5 s (4 s in Experiments 2 and 3). After
dynamic group interactions tend to escalate belief strength
this social exchange phase, the trial ended with feedback provided
[20,22,23,25]—we expected that changing the nature of infor-
to both members of the dyad.
mation exchange (dynamic versus static) would make the final Our manipulation concerned only the social exchange stage.
judgments participants converge to after social exchange more Two conditions alternated across blocks. In the Static condition,
extreme, even when matching decision evidence across the the choice and confidence level (the two together representing a
two conditions. Thus, contrary to previous literature, we tightly ‘belief’) selected by each dyad member in the private phase
controlled the perceptual evidence accumulated during the appeared on their partner’s scale as a static coloured cursor.
private decision phase, which is the only information needed Dyad members were at this point asked to (and were rewarded
to successfully accomplish our task. Finally, feedback was for) continuously monitor and update their confidence by

Proc. R. Soc. B 287: 20200025

provided at the end of the trial (Feedback). moving their mouse along the scale. In the Dynamic condition,
the social exchange part started exactly as in the static one, with
Three experiments (N = 72) were performed. Experiment 1
each dyad member’s cursor appearing on their partner’s scale.
directly compares dynamic and static situations within dyads.
However, and for the whole duration of the social part (5 s), if a
We found that when participants are allowed to freely interact
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

member updated their belief, this would instantly appear also on

in real-time, recursivity in the belief update is observed. Exper- their partner’s scale and vice-versa. This led to a situation where
iments 2 and 3 (reported in electronic supplementary material) participants were not only informed of their partner’s original
largely replicate the findings of Experiment 1, adding controls beliefs, but also how those beliefs changed in real-time as a func-
for potential confounding factors and showing robustness tion of their own belief changes (figure 1b). In both conditions,
across different instructions and confidence scales. Exper- participants could update their decision and/or confidence level,
iments 2 and 3 tested a simple alternative explanation of the thus potentially completely reversing their belief (e.g. switching
interactive effects reported below for Experiment 1, namely from ‘certainly left’ to ‘certainly right’). The recursive nature of
that participants quickly forgot their initial decisions and the Dynamic condition thus captured dynamics typically observed
in formal approaches of opinion dynamics [14,15] (see electronic
were only left with their current expressed beliefs as a basis
supplementary material for detailed methods).
for continued belief updating. We therefore included confi-
dence anchors to remind each participant of the initial,
independently expressed confidence in their own judgment
(Experiment 2) and of their partner (Experiment 3). We largely 3. Results
replicated results found in Experiment 1, suggesting that this
alternative explanation did not account for the data. The
(a) Confidence changes asymmetrically for consensus
study was approved by the University of Oxford ethics and condition
board (MS-IDREC-C1-2015-075). During the social stage, the time-series of confidence changes
was recorded continuously for 5 s after each participant’s
decision. The largest changes occurred about 1 s after first
exposure to their partner’s belief (electronic supplementary
2. Methods material, figure S1). Confidence updates were typically a
(a) Participants single, unidirectional transition, towards increased confidence
Twenty-four dyads (N = 48, 37 females, age: M = 23 years, s.d. = when partners agreed and reduced confidence when they
3.2) were recruited in pairs—potential volunteers were invited to disagreed (figure 2a). The influence that social information has
bring a friend—for money or course credits. All participants gave on each dyad member is quantified by the distance δC between
informed consent before taking part in the study. their post-exchange signed confidence and pre-exchange signed
confidence, with larger absolute values corresponding to
greater impact of social information. The sign of δC is expected
(b) Procedure to be positive in agreement trials and negative in disagreement
Participants sat on opposite sides of a desk divided by an occlud-
ing screen (figure 1a), each with a separate monitor, keyboard, and trials. The range of possible confidence change however varies
mouse. The pair performed a series of trials together, each consist- in agreement and disagreement: while in agreement the maxi-
ing of a private perceptual decision phase followed by social mum absolute distance between two beliefs is 50 (remember
exchange (figure 1b). During the private perceptual decision, our response scale had 50 points per interval), the maximum
participants in parallel performed a dot-count decision task [28]: absolute distance between two disagreeing beliefs is 100, corre-
two boxes containing dots randomly arrayed on a 20 × 20 grid sponding to an extreme situation when both participants
were briefly (160 ms) flashed on each participant’s screen to the are maximally confident, but on opposite intervals (e.g. one
left and right of a central fixation cross. Participants had to indicate responding ‘sure LEFT’ the other ‘sure RIGHT’).
which box contained most dots and their confidence in this A 2 × 2 ANOVA on confidence change with factors of
decision. Thus beliefs were operationalized as signed confidence
consensus with respect to the initial, private decisions (dis-
on a continuous scale, ranging from −50 (confident Left) to +50
agreement versus agreement) and condition (static versus
(confident Right). Conversely, confidence was the absolute value
of the belief, ranging from 1 to 50. The difficulty of the task was dynamic) revealed significant main effects of both consensus
titrated to participants’ perceptual threshold, using a staircase pro- (F1,47 = 150.26, p < 0.001, h2G ¼ 0:7) and condition (F1,47 = 9.40,
cedure that converged at 72.52% ± 0.013 accuracy [29]. As soon as p < 0.01, h2G ¼ 0:005), but no significant interaction (F < 1). As
both members confirmed their answer by pressing the space bar, expected, average δC was negative in disagreement and posi-
the social exchange stage started, where each dyad member was tive in agreement (figure 2b). The main effect of condition
 (a) (b) 4
15 0.6
static agree

royalsocietypublishing.org/journal/rspb
10 dynamic 0.5 disagree

average confidence change 5 0.4

root density
0 0.3

–5 0.2

–10 0.1

Proc. R. Soc. B 287: 20200025

–15 0
disagreement agreement –100 –50 0 50
confidence change
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

(c) (d) correlation between confidence changes

10 0.4
static
0.3 dynamic
average Pearson’s coefficient
confidence (arb. units)

0.2
5
0.1

0
0
–0.1

–0.2

–5 –0.3
0 10 20 30 40 50 disagree agree

iterations
Figure 2. (a) Average confidence change in agreement and disagreement trials, plotted for each condition. (b) Confidence change root density distribution in
agreement and disagreement. The corresponding histogram or raw frequencies (no error bars) is shown in electronic supplementary material, figure S5. (c)
Toy model emulating how confidence can increase in a Dynamic disagreement trial. y-axis represents belief as signed confidence where the sign represents a
decision interval (LEFT/RIGHT) and the absolute value represents confidence. In a Static condition, the agents update their initial belief with their discounted
partner’s belief (dashed lines). In the dynamic interaction, the same update rule is applied on every iteration until equilibrium is reached (solid lines). (d )
Effect of condition on the correlation between absolute confidence changes of the two participants across trials. Average Pearson’s correlation coefficient is plotted
as a function of consensus and condition. One dyad removed for a missing cell. Error bars represent s.e.m. (Online version in colour.)

indicated that participants’ final level of confidence was greater using a trial-level mixed regression approach (electronic
in the Dynamic condition than in the Static condition, with sep- supplementary material, table S1).
arate paired t-tests confirming this effect held regardless of Figure 2b shows the density distribution of confidence
consensus: participants increased their confidence more changes broken down by consensus (agreement versus dis-
when in agreement (t47 = 2.69, p = 0.009) and decreased their agreement). Both agreement and disagreement distributions
confidence less when in disagreement (t47 = 2.08, p = 0.04), in peaked around zero, which was by far the most common
dynamic compared to static blocks. The results indicate that change (notice that the y-axis in panel b uses a square root
belief change is dominated by the information content con- scale), indicating that very often participants ignored social
veyed (i.e. agreement versus disagreement), with the nature information. Also of note is that on some disagreement trials
of interaction (static versus dynamic) modulating rather participants actually increased their confidence, and agree-
than fundamentally altering this pattern. Nevertheless, the ment trials they decreased it. This is a surprising result if we
observed increase in confidence in the dynamic case is non- consider that, from a normative (e.g. Bayesian) perspective, dis-
trivial when considered in some parts of the belief space agreement with an independent observer should always lead
(figure 3) and when considering that it represents an average to a reduction of confidence and agreement should always
including trials when participants did not change their confi- lead to an increase of confidence (if we assume that the partici-
dence at all (figure 2b). Similar results were obtained when pant believes that their partner performs better than chance).
 confidence change of dominant participant 5
(a) (static) (b) (dynamic) (c) (dynamic – static)

royalsocietypublishing.org/journal/rspb
50 30 50 30 50 15

confidence of dominated participant 20 20 10

x

10 10 5

0 0 0 0 0 0

–10 –10 –5

–20 –20 y –10

Proc. R. Soc. B 287: 20200025

–50 –30 –50 –30 –50 –15
0 50 0 50 0 50
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

confidence of dominant participant

confidence change of dominated participant

(d) (static) (e) (dynamic) (f) (dynamic – static)
50 30 50 30 50 15
confidence of dominated participant

20 20 10
x

10 10 5

0 0 0 0 0 0

–10 –10 –5

–20 –20 y –10

–50 –30 –50 –30 –50 –15

0 50 0 50 0 50
confidence of dominant participant
Figure 3. Confidence change observed for every social situation. On a given trial, the x-axis represents the initial confidence of the trial-dominant member, and the
y-axis the initial confidence of the trial-dominated member in relation to the former (+50 indicates confident agreement, −50 indicates confident disagreement).
Horizontal and vertical coordinates thus represent the pre-exchange dyad state, while the colour (z-axis), represents confidence change. Panels a–c represent con-
fidence change in dominant trials in Static condition, Dynamic condition, and their contrast, respectively. Panels d–f represent confidence change in dominated trials
in Static condition, Dynamic condition, and their contrast, respectively. Panels c and f represent contrasts between the two conditions (contrasts (b,a) and (e,d),
respectively). Regions of the space labelled ‘x’ and ‘y’ correspond to regions of the belief space where the effect of condition is strongest as described in the main
text. Trials within each condition and dyad were linearly interpolated, due to data sparseness. Each panel contains 175 data points. (Online version in colour.)

To explore this surprising pattern of confidence change confidence changes were 4.5 times more frequent in disagree-
more formally, a two-way repeated measures ANOVA was ment than agreement (0.018 versus 0.004 of trials) and that
computed on the probability of an irrational confidence the Dynamic condition was 33% more likely to produce
change, defined as confidence decreases in cases of agreement irrational confidence increases than the Static condition (0.020
and confidence increases in cases of disagreement, again with versus 0.015 of trials, t47 = 2.59, p = 0.01), as well as (numeri-
factors of consensus and condition. To avoid including trials cally) 25% less likely to produce irrational confidence
where increases/decreases in confidence were simply due to decreases (0.003 versus 0.005 of trials, t47 = 1.98, p = 0.05). The
involuntary cursor movements (a ‘trembling hand’), we interaction between condition and consensus was however
defined a confidence ‘change’ as a shift larger than 5 confidence not replicated in Experiment 2 (electronic supplementary
points in the unexpected direction. The findings were, material), indicating that the result was not robust to changes
however, consistent across cut-offs greater than zero. Results in the use of the confidence scale.
show a significant effect of consensus (F1,47 = 7.88, p = 0.007, Although not normatively prescribed (e.g. in a Bayesian
h2G ¼ 0:07) but not of condition (F1,47 = 2.56, p = 0.11, framework), belief aggregation strategies described in the
h2G ¼ 0:001) and a significant interaction between the two literature [3,30,31] can explain irrational decreases in confi-
(F1,47 = 9.90, p = 0.002, h2G ¼ 0:005), indicating that irrational dence in cases of agreement. For example, averaging of
 confidence would lead to this outcome when a partner agrees represented by a pair of coordinates, while collapsing across 6
but is much less confident than the participant, such that the the particular side of participants’ choices (left versus right

royalsocietypublishing.org/journal/rspb
participant concludes that they should not have been so confi- box). In figure 3, pixel colour indicates the median change in
dent in the first place. Of more interest, therefore, are irrational confidence from pre- to post-exchange of the dominant
increases in confidence after disagreement, which occurred (upper panels) and dominated (lower panels) member of the
more frequently than irrational decreases but which are dyad on each trial. The trial-dominant and dominated partici-
difficult to reconcile with any obvious confidence-update strat- pants’ confidence change can be combined into a single vector
egy. We notice that this irrationality could occur through field (electronic supplementary material, figure S2) visualizing
recursive dynamics introduced by real-time interaction. dyadic transitions in state space [31]. Dynamic animations of
Consider an example trial in which a participant starts off on confidence transition in this space for each condition are
a confidence level of 6 while their partner weakly disagrees provided via an Open Science Framework (OSF). Figure 3
with a confidence level of -4 (the negative sign indicates dis- shows once again the overall increase in confidence seen in
agreement). Suppose next that both participants use a simple dynamic versus static interactions, with dynamic interaction
update strategy, namely summing their own initial signed con- characterized by greater increases in confidence when partners

Proc. R. Soc. B 287: 20200025

fidence with their partner’s weighted signed confidence (here: agreed (cf. larger red area in the upper half of figure 3b than 3a)
weight = 0.80) (cf. [31]). In a situation without recursive inter- and smaller decreases when they disagreed (cf. smaller blue
action, participants can only use their partner’s initial belief area in lower half of figure 3b than 3a). The contrast plots
to update their beliefs (dashed lines). If recursive dynamics (figure 3c,f ) more precisely identify the locus of these
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

are introduced, each participant can use his/her partner’s cur- between-condition differences, while also highlighting the
rent belief to update their own. Figure 2c shows that this simple magnitude of the effects in certain conditions. Thus, dynamic
strategy leads to an oscillatory update (solid lines) that stabil- interaction leads to particularly marked confidence increases
izes for the more confident participant (in red) on a higher when partners began the social stage in agreement but with
confidence (distance from 0) than initially held. The effect low confidence (points marked ‘x’ in figure 3). Confidence
reflects the fact that if the low-confident partner (in blue) crosses change in the Dynamic condition (panel b,e) in these conditions
the decision boundary, disagreement turns into agreement— of uncertain agreement is 20–30 confidence points, and
thus supporting one’s initial belief—instead of providing con- approximately 15 points greater than in static interaction
tradictory evidence, and therefore leads to an increase in blocks. Thus, when interacting dynamically, but not statically,
confidence (electronic supplementary material, figure S9). two uncertain partners tended to reinforce each other’s belief
To test for recursive dynamics in our behavioural data, we so that together they converged on the maximum possible con-
counted, for each condition, the average number of vacillations fidence level. The other key point of interest in the contrast
in a trial, namely the number of times the direction of plots lies in the disagreement half of the belief space, specifi-
the update (i.e. stationary/increase/decrease) changed in the cally at the points labelled ‘y’ in figure 3c,f. These were trials
update window (see electronic supplementary material). Sup- in which the dominant member was highly confident and
porting our intuitions, we found that both the average number the dominated member weakly disagreed, a situation
of vacillations in a trial and the total number of irrational described in our simple simulation above. The warmer colour
increases were significantly more frequent in the dynamic than at ‘y’ in figure 3c indicates that disagreement had markedly less
Static condition. Thus, irrational increases in confidence could impact on the dominant partner’s confidence when the dyad
have arisen because participants treated the observed updates interacted dynamically rather than statically. The correspond-
of their partner (influenced by their own judgment) as if they ing point in figure 3f indicates that, similarly, larger shifts in
were independent evidence. Figures S6–S9 show individual the trial-dominated participant’s confidence toward the trial-
trial trajectories in belief space, including agreement and dominant individual’s position were observed in dynamic
disagreement, vacillations, and irrational confidence increases. compared to static blocks. Overall, therefore, this belief space
analysis identifies the conditions under which dynamic inter-
action has its largest impact—when partners agree with
(b) Dyadic interactions in belief space symmetric low levels of confidence, and when they disagree
The preceding analyses explore belief change when the data with asymmetric levels of confidence—and shows that this
are aggregated across broad categories (e.g. agreement versus impact is substantial in these conditions.
disagreement trials). To explore more nuanced behaviour as
a function of participants’ absolute and relative levels of confi-
dence, we explored our data as a function of a two-dimensional (c) Coupling of confidence changes during interaction
‘belief space’ [31] as shown in figure 3. The figure plots confi- The analyses above consider belief change for each dyad
dence change following interaction within this space. Here, member separately. To investigate how the magnitude of
the x-axis value indicates the confidence of whichever partici- partners’ belief changes co-varied across trials, we calculated
pant in the dyad is the more confident on any given trial across-trial correlations of absolute confidence change between
in the private judgment phase, henceforth the ‘dominant’ initial and final confidence (δC) between the two members of
member on the trial, thus ranging from 1 (minimum confi- each dyad. Pearson’s r coefficients were calculated for each
dence) to 50 (maximum confidence). The y-axis value gives dyad as a function of dyadic consensus (agree versus disagree)
the confidence of the less confident, or ‘dominated’, member and interaction condition (dynamic versus static). A 2 × 2
of the dyad in their initial judgment on the trial, on a scale ANOVA on the resulting values across dyads (figure 2d)
ranging from −50 (disagreement with maximum confidence) revealed a main effect of consensus (F1,22 = 20.93, p < 0.001)
to 50 (agreement with maximum confidence). This plot creates but not of condition (F1,22 = 1.71, p = 0.20), and a significant
a grid of possible social situations in which the dyad’s state— interaction between the two (F1,22 = 38.39, p < 0.001). When
both members’ choices and their confidence—is fully dyad members did not see each other’s confidence changes
 in the social exchange stage (Static condition), confidence account a host of circumstantial variables that are known 7
changes did not correlate significantly between partners (h1 = to co-vary with problem-specific evidence, but that are not

royalsocietypublishing.org/journal/rspb
r > 0, t22 < 1.28, p > 0.2, d < 0.26). This finding is not unexpected, themselves strictly task-relevant [33]. Interpreting someone’s
but nor is it trivial—for example, a positive correlation might changes of mind as another cue for confidence is sensible in
be expected in agreement trials as a consequence of a boost many daily-life situations. Indeed, if somebody’s beliefs are
to both participants’ confidence when they agreed but were fickle, we have reasons to believe s/he must be uncertain. In
initially uncertain, such that agreement led to increased confi- the case of social interaction, however, this heuristic leads to
dence for both. This did not seem to occur. By contrast, in the potentially sub-optimal self-reinforcing dynamics, when a
Dynamic condition, partners’ confidence changes became person uses the impact of his/her own belief on the other
coupled: in agreement trials, the correlation was positive, indi- person as evidence for the belief itself. The impact of this circu-
cating that the more one member increased their confidence, lar reasoning was particularly marked for low confidence
the more their partner also increased their confidence. In dis- agreement. In these cases, dyads often escalated together
agreement, the correlation was negative, indicating that the towards maximal confidence in their beliefs.
more one member decreased their confidence in their initial This micro-level effect may provide some insights into

Proc. R. Soc. B 287: 20200025

decision, the less the partner decreased theirs. Pairwise con- poorly understood group phenomena of certainty escalation
trasts showed that, compared to the Static condition, in the and ‘confidence effects’. Confidence increases are observed
Dynamic condition correlation coefficients were significantly when individuals are allowed to freely communicate in
greater for agreement (t22 = 4.89, p < 0.001, d = 1.20) and some- groups and the more people are exposed to social information,
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

what smaller for disagreement (t22 = 2.02, p = 0.05, d = 0.52). with no improvements or even damaging effects on accuracy
The negative correlation found in disagreement dynamic [20,34–36]. The effect of group polarization has long been
trials, was replicated in Experiment 3, but not in Experiment studied, and has gained renewed attention and meaning in
2 (electronic supplementary material). Further analyses (elec- the context of online interactions [23,24,27]. The present
tronic supplementary material) showed that the effect could study adds to this growing body of evidence by showing that
not be explained by participants using their partner’s reaction a potential cause of belief escalation is recursive interaction.
times (i.e. ‘unwillingness to move’) in the dynamic but not in When the interaction allows for recursive dynamics, the use
the Static condition. Coupling of confidence changes in inter- of redundant task-irrelevant information becomes more
action suggests that participants were influenced by their likely. People should use only each other’s independent beliefs
partner’s confidence change when updating their own belief, to arrive at a final decision, because this is the only information
rather than basing their change solely on their partner’s initial that carries task-relevant value. However, they also (incor-
(independent) judgment as normatively prescribed. rectly) use how much their own belief is affecting their
partners. This strategy creates dependencies that can poten-
tially create escalation dynamics. Our dynamic model of
4. Discussion belief update, even though based on dyadic interaction and
The present study compared social exchange involving highly simplified (figure 2c), shares important features with
static, one-step communication with exchange characterized models of opinion dynamics—including the operationaliza-
by dynamic and recursive interaction. We hypothesized that tion of belief and belief update as signed one-dimensional
real-time recursive dynamics, which characterize many daily- continuous values—which use formal models drawn from
life interactions of social influence [18], would have systematic engineering and physics to study the nonlinear properties of
impact on decisions made in a social context, over and above a network of individual nodes holding beliefs. The study of
the effects of the information brought by each individual to these systems, though less complex than real societies, has
the interaction, as studied in traditional judge–advisor systems nevertheless proven valuable for social scientists interested in
[4,32]. Across conditions with equal information available— emerging macroscopic influence dynamics [14,15,37]. We add
because in both the dynamic and static conditions, dyad to this literature by showing how recursive dynamics can be
members viewed perceptual evidence separately and for empirically captured in simple experimental paradigms and
160 ms only—we observed different belief aggregation strat- cognitive models of belief update.
egies according to the nature of communication between Beyond belief escalation as described above, a more subtle
partners. Specifically, dynamic interaction produced higher but nevertheless consistent effect of bidirectional dyadic influ-
confidence changes in agreement and smaller confidence ence in our data was that participants occasionally increased
changes in disagreement by breaking the independence of their confidence despite initially learning that their partner
dyad members’ beliefs: confidence changes of the two partici- disagreed with them. This surprising irrationality occurred
pants became correlated during dynamic interaction compared more frequently than the mirror effect of decreasing confidence
to a static baseline so that, in agreement, greater increases in when a partner agreed. Similar phenomena have been
confidence for one member were associated with greater confi- observed in the context of political beliefs in online envi-
dence increases for their partner, leading to belief escalation. In ronments [27]. Our results show that the effect was more
disagreement, greater changes in confidence for one member common in dynamic rather than static trials, and suggest
were associated with smaller changes in confidence for their a possible low-level mechanism for it. This phenomenon
partner, reducing the impact of disagreement on belief updat- cannot easily be explained by static aggregation rules like aver-
ing. These combined results can be understood in terms of aging, summing, or Bayesian integration [3,30,31], but is
individuals making use of their partner’s change in confidence predicted by a recursive update model. Importantly, rather
to update their belief, without taking into account that this than providing an exact description of participants’ behaviour,
change was biased (and indeed generated) by exposure to our model aimed to broadly capture the intuition that others’
their own judgment. People are known to quickly reach changes of mind can sometimes be perceived as supporting
decisions even when information is scarce by taking into evidence for one’s views. When a partner initially disagrees,
 but loses confidence in this view (or even reverses it) on sub- in line with the advice-taking literature suggesting the presence 8
sequently learning the participant’s view, the participant can of egocentric and confirmation biases [45]. More specifically,

royalsocietypublishing.org/journal/rspb
take this as evidence in favour of their initial position, and the weighting of partner’s information (i.e. use of social infor-
therefore increase their confidence. mation) seemed to follow a bimodal distribution, with peaks
Why do people make this mistake? Our interpretation around 50%, corresponding to uninformative social evidence,
implies that people seem to have limited ability to assess the and 100%, corresponding to maximally supporting social evi-
independence of evidence, a conclusion that converges with dence. These results are consistent with the hypothesis that
previous findings using the judge–advisor system paradigm people are solving a categorical inference problem (cf. [46]).
[38]. Another possibility is that when making joint decisions, Instead of using continuous social information as it is provided
we reduce our individual responsibility [39]. Feeling less by their social partners, participants seem to classify each trial
responsible, people might afford to be more confident. as ‘partner is wrong’ versus ‘partner is correct’ and, once this
A related social context explanation might apply to our categorization is performed, use social information accord-
observation that confidence actually increases on some dis- ingly to update their views. Accordingly, participants would
agreement trials: this phenomenon might occur because try to minimize situations of uncertainty (e.g. 0.25 or 0.75 evi-

Proc. R. Soc. B 287: 20200025

making a good decision is not the only goal of the decision- dence), thus maximizing the impact of social information on
maker. Another aim might be winning an argument [40]. final confidence. Notice that this is in stark opposition with tra-
Once the player knows that their ‘rival’ is less confident, they ditional opinion dynamics models of social contagion, where
will be encouraged to prevail by increasing further their confi- belief updates are modelled as a linear combination of self
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

dence. However, due to the non-verbal perceptual features of and others’ beliefs [14,47].
the task, we expect argumentative reasons to have less weight Collectively, these findings show that social influence
here. Future studies should investigate whether irrational confi- depends not only on private beliefs—here, the only task-rel-
dence increases are also observed in disagreement when logical evant information—but also on the modality in which
and linguistic arguments are exchanged. information is shared and transformed across individuals. In
Importantly, notwithstanding these differences in confi- the aggregate, the impact of recursive dynamics is subtle but
dence updating across static versus dynamic interactions, consistent, evident as a general increase in confidence in
participants’ overall accuracy showed a consistent benefit from decisions made. However, the impact is very marked in specific
social information exchange. A significant effect of decision situations, notably when shared but uncertain beliefs become
stage (pre- versus post-interaction) was found on both accuracy mutually reinforced to a state of near certainty, and when a
(F1,47 = 47.00, p < 0.001, h2G ¼ 0:16) and confidence calibration decision-maker interprets vacillation in a partner’s weak
(F1,47 = 89.58, p < 0.001, h2G ¼ 0:25) (see electronic supplemen- disagreement as positive evidence for their views. The rel-
tary material information for details). The benefits of social evance of these basic dynamics might extend beyond human
exchange were of similar magnitude across static and dynamic groups, to include other social animals, for example, in the
interaction conditions, whether this benefit was measured in case of collective motion direction [48]. Our findings contribute
terms of overall accuracy, confidence in the correct answer, or to the debate on group polarization in online and physical
the calibration between confidence and objective performance environments by providing a fine-grained description of
(see electronic supplementary material). This parity was within-subject belief dynamics in recursive and static social
observed despite different interaction dynamics across con- exchanges. Real-time interaction in many daily social situations
ditions, at least in part because these dynamics led to is recursive in nature. Effective interventions aimed at reducing
opposing effects on accuracy across trials. A standing question belief escalation online and offline will require a cognitive-level
is whether interaction might amplify errors in the presence understanding of these dynamics.
of systematic biases across members (e.g. [31]), due to belief
escalation in incorrect answers [41,42]. Ethics. The study was approved by the University of Oxford’s ethics
In the electronic supplementary material, we compare committee with Ethics Approval Ref: MS-IDREC-C1-2015-075.
empirically observed behaviour with an optimal model Data accessibility. Data available from osf.io/7b6py [49].
grounded in a probabilistic interpretation of confidence [43], Authors’ contributions. N.P. and N.Y. designed the study and drafted the
and show participants clearly departed from the Bayesian strat- manuscript. N.P. collected the data, performed the data analysis, and
egy commonly explored in the belief aggregation literature curated data availability and reproducibility. N.Y. supervised the
[30,44], in which beliefs are combined in relation to expressed overall data collection and analysis.
confidence. We used inverse Bayes theorem to compare the Competing interests. We declare we have no competing interests.
objective information conveyed from social partners with the Funding. This work was completed with the support of the Clarendon
Fund, Christ Church College and the Department of Experimental
information ‘perceived’ by a participant. Results evidenced Psychology of the University of Oxford.
that social information was distorted by the receiver in a self- Acknowledgements. We are very grateful for the constructive comments
serving manner and asymmetrically for agreement and of three anonymous reviewers, which helped shape the final version
disagreement (electronic supplementary material, figure S3), of this paper.

References
1. Vul E, Goodman N, Griffiths TL, Tenenbaum JB. 2. Galton F. 1907 Vox populi. Nature 75, 450–451. Manage. Sci. 52, 111–127. (doi:10.1287/mnsc.1050.
2014 One and done? Optimal decisions from very (doi:10.1038/075450a0) 0459)
few samples. Cogn. Sci. 38, 599–637. (doi:10.1111/ 3. Larrick RP, Soll JB. 2006 Intuitions about combining 4. Bonaccio S, Dalal RS. 2006 Advice taking and
cogs.12101) opinions: misappreciation of the averaging principle. decision-making: an integrative literature review,
 and implications for the organizational sciences. second-person neuroscience. Behav. Brain Sci. 36, 34. De Martino B, O’Doherty JP, Ray D, Bossaerts P, 9
Organ. Behav. Hum. Decis. Process. 101, 127–151. 393–414. (doi:10.1017/S0140525×12000660) Camerer C. 2013 In the mind of the market: theory

royalsocietypublishing.org/journal/rspb
(doi:10.1016/j.obhdp.2006.07.001) 19. Fusaroli R, Bahrami B, Olsen K, Roepstorff A, Rees G, of mind biases value computation during financial
5. Salganik MJ, Dodds PS, Watts DJ. 2006 Frith C, Tylén K. 2012 Coming to terms: quantifying bubbles. Neuron 79, 1222–1231. (doi:10.1016/j.
Experimental study of in-equality and the benefits of linguistic coordination. Psychol. Sci. neuron.2013.07.003)
unpredictability in an artificial cultural market. 23, 931–939. (doi:10.1177/0956797612436816) 35. Lorenz J, Rauhut H, Kittel B. 2015 Majoritarian
Science 311, 854–856. (doi:10.1126/science. 20. Mahmoodi A, Bang D, Ahmadabadi MN, Bahrami B. democracy undermines truth-finding in deliberative
1121066) 2013 Learning to make collective decisions: the committees. Res. Politics 2, 205316801558228.
6. Price PC, Stone ER. 2004 Intuitive evaluation of impact of confidence escalation. PLoS ONE 8, (doi:10.1177/2053168015582287)
likelihood judgment producers: evidence for a e81195. (doi:10.1371/journal.pone.0081195) 36. Lorenz J, Rauhut H, Schweitzer F, Helbing D. 2011
confidence heuristic. J. Behav. Decis. Mak. 17, 21. Kerr NL, Scott Tindale R. 2004 Group performance How social influence can undermine the wisdom of
39–57. (doi:10.1002/bdm.460) and decision making. Annu. Rev. Psychol. 55, crowd effect. Proc. Natl Acad. Sci. USA 108,
7. Sah S, Moore DA, Maccoun RJ. 2013 Cheap talk and 623–655. (doi:10.1146/annurev.psych.55.090902. 9020–9025. (doi:10.1073/pnas.1008636108)
credibility: the consequences of confidence and 142009.) 37. Galesic M, Stein DL. 2019 Statistical physics models

Proc. R. Soc. B 287: 20200025

accuracy on advisor credibility and persuasiveness. 22. Moscovici S, Zavalloni M. 1969 The group as a of belief dynamics: theory and empirical tests.
Organ. Behav. Hum. Decis. Process. 121, 246–255. polarizer of attitudes. J. Pers. Soc. Psychol. 12, Physica A 519, 275–294. (doi:10.1016/j.physa.2018.
(doi:10.1016/j.obhdp.2013.02.001) 125–135. (doi:10.1037/h0027568) 12.011)
8. See KE, Rothman NB, Soll JB. 2010 Powerful and 23. Myers DG, Lamm H. 1976 The group polarization 38. Yaniv I, Choshen-Hillel S, Milyavsky M. 2009
Downloaded from https://royalsocietypublishing.org/ on 28 April 2024

unpersuaded: the implications of power for phenomenon. Psychol. Bull. 83, 602–627. (doi:10. ‘Spurious consensus and opinion revision: why
confidence, advice taking, and accuracy. Acad. 1037/0033-2909.83.4.602) might people be more confident in their less
Manage. Proc. 2010, 1–6. (doi:10.5465/AMBPP. 24. Becker J, Porter E, Centola D. 2019 The wisdom of accurate judgments?’ J. Exp. Psychol. Learn. Mem.
2010.54493606) partisan crowds. Proc. Natl Acad. Sci. USA 166, Cogn. 35, 558–563. (doi:10.1037/a0014589)
9. Gunnar ES, Dalal RS, Sniezek JA. 2006 No decision- 10 717–10 722. (doi:10.1073/pnas.1817195116) 39. El Zein M, Bahrami B, Hertwig R. 2019 Shared
maker is an Island: integrating expert advice with 25. Del Vicario M, Vivaldo G, Bessi A, Zollo F, Scala A, responsibility in collective decisions. Nat. Hum.
information acquisition. Behav. Decis. Mak. 19, Caldarelli G, Quattrociocchi W. 2016 Echo chambers: Behav. 3, 554–559. (doi:10.1038/s41562-019-
43–60. (doi:10.1002/bdm.514) emotional contagion and group polarization on face- 0596-4)
10. Budescu DV, Rantilla AK. 2000 Confidence in book. Sci. Rep. 6, 37825. (doi:10.1038/srep37825) 40. Mercier H. 2016 The argumentative theory:
aggregation of expert opinions. Acta Psychol. 104, 26. Guilbeault D, Becker J, Centola D. 2018 Social predictions and empirical evidence. Trends Cogn. Sci.
371–398. (doi:10.1016/S0001-6918(00)00037-8) learning and partisan bias in the interpretation of 20, 689–700. (doi:10.1016/j.tics.2016.07.001)
11. Soll JB, Mannes AE. 2011 Judgmental aggregation climate trends. Proc. Natl Acad. Sci. USA 115, 41. Jasny L, Waggle J, Fisher DR. 2015 An empirical
strategies depend on whether the self is involved. 9714–9719. (doi:10.1073/pnas.1722664115) examination of echo chambers in US climate policy
Int. J. Forecast. 27, 81–102. (doi:10.1016/j.ijforecast. 27. Bail CA et al. 2018 Exposure to opposing views on networks. Nat. Clim. Change 5, 782–786. (doi:10.
2010.05.003) social media can increase political polarization. Proc. 1038/nclimate2666)
12. Schultze T, Rakotoarisoa A-F, Schulz-Hardt S. 2015 Natl Acad. Sci. USA 115, 9216–9221. (doi:10.1073/ 42. Mackay C. 1841 Extraordinary popular delusions and
Effects of distance between initial estimates and pnas.1804840115) the madness of crowds. Ware, UK: Wordsworth
advice on advice utilization. Judgm. Decis. Mak. 10, 28. Boldt A, Yeung N. 2015 Shared neural markers of Edition Limited.
144–171. decision confidence and error detection. J. Neurosci. 35, 43. Harris AJL, Hahn U, Madsen JK, Hsu AS. 2016 The
13. Sherif CW, Sherif MS, Nebergall RE. 1965 Attitude 3478–3484. (doi:10.1523/JNEUROSCI.0797-14.2015) appeal to expert opinion: quantitative support for a
and attitude change. Philadelphia, PA: W.B. 29. Treutwein B. 1995 Adaptive psychophysical Bayesian network approach. Cogn. Sci. 40,
Saunders Company. procedures. Vision Res. 35, 2503–2522. (doi:10. 1496–1533. (doi:10.1111/cogs.12276)
14. Friedkin NE, Bullo F. 2017 How truth wins in 1016/0042-6989(95)00016-X)) 44. Koriat A. 2012 When are two heads better than one
opinion dynamics along issue sequences. Proc. Natl 30. Bang D et al. 2014 Does interaction matter? Testing and why? Science 336, 360–362. (doi:10.1126/
Acad. Sci. USA 114, 11 380–11 385. (doi:10.1073/ whether a confidence heuristic can replace interaction science.1216549)
pnas.1710603114) in collective decision-making. Conscious. Cogn. 26, 45. Nickerson RS. 1998 Confirmation bias: a ubiquitous
15. Flache A, MäS M, Feliciani T, Chattoe-Brown E, 13–23. (doi:10.1016/j.concog.2014.02.002) phenomenon in many guises. Rev. Gen. Psychol. 2,
Deffuant G, Huet S, Lorenz J. 2017 Models of social 31. Pescetelli N, Rees G, Bahrami B. 2016 The 175–220. (doi:10.1037/1089-2680.2.2.175)
influence: towards the next frontiers. J. Artif. Soc. perceptual and social components of metacognition. 46. Soll JB, Larrick RP. 2009 ‘Strategies for revising
Soc. Simul. 20. (doi:10.18564/jasss.3521) J. Exp. Psychol. General 145, 949–965. (doi:10. judgment: how (and how well) people use others’
16. Auvray M, Lenay C, Stewart J. 2009 Perceptual 1037/xge0000180) opinions.’ J. Exp. Psychol. Learn. Mem. Cogn. 35,
interactions in a minimalist virtual environment. 32. Sniezek JA, Buckley T. 1995 Cueing and cognitive 780–805. (doi:10.1037/a0015145)
New Ideas Psychol. 27, 32–47. (doi:10.1016/j. conflict in judge-advisor decision making. In 47. DeGroot MH. 1974 Reaching a consensus. J. Am.
newideapsych.2007.12.002) Organizational Behavior and Human Decision Stat. Assoc. 69, 118. (doi:10.2307/2285509)
17. Dumas G, de Guzman GC, Tognoli E, Kelso JS. 2014 Processes. (doi:110.1006/obhd.1995.1040). 48. Couzin ID. 2009 Collective cognition in animal
The human dynamic clamp as a paradigm for social http://linkinghub.elsevier.com/retrieve/pii/ groups. Trends Cogn. Sci. 13, 36–43. (doi:10.1016/j.
interaction. Proc. Natl Acad. Sci. USA 111, S0749597885710400 tics.2008.10.002)
3726–3734. (doi:10.1073/ pnas.1407486111) 33. Gigerenzer G, Todd P, ABC Research group. 1999 49. Pescetelli N, Yeung N. 2020 The effects of recursive
18. Schilbach L, Timmermans B, Reddy V, Costall A, Simple heuristics that make us smart. Oxford, UK: communication dynamics on belief updating.
Bente G, Schlicht T, Vogeley K. 2013 Toward a Oxford University Press. Retrieved from osf.io/7b6py.