A Peer Pressure Experiment: Recreation of the Asch Conformity

A peer pressure experiment: Recreation of the Asch conformity

experiment with robots.

urgen Brandstetter

, P

eter R

acz

, Clay Beckner

, Eduardo B. Sandoval

, Jennifer Hay

, Christoph Bartneck

Abstract— The question put forward in this paper is whether

robots can create conformity by means of group pressure. We

recreate and expand on a classic social psychology experiment

by Solomon Asch, so as to explore three main dimensions. First,

we wanted to know whether robots can prompt conformity in

human subjects, and whether there is a signiﬁcant difference

between the degree to which individuals conform to a group

of robots as opposed to a group of humans. Secondly we ask

whether group pressure (from human or robot peers) can exert

inﬂuence in verbal judgments, analogously to the inﬂuence on

visual judgments that is known from previous research [3], [2].

Thirdly, we investigate whether the level of conformity differs

between an ambiguous situation and a non-ambiguous situation.

Our results show that in both visual and verbal tasks,

participants exhibit conformity with human peers, but not

with robot peers. The social inﬂuence of robot peers is not a

signiﬁcant predictor of verbal or visual judgments in our tasks.

Furthermore, the level of conformity is signiﬁcantly higher in

an ambiguous (unclear) situation.

I. INTRODUCTION

When we look at the current development of human-

like social robots, it is possible to predict a future where

robots help out in the ofﬁce, teach children in the class-

room, become companions, work in advertising, or help in

our households [5], [14], [20]. During the last few years,

there has been an observable trend towards service robots

outﬂanking industrial robots in production volume. Whereas

all sold industrial robots total 2.3 million units, more than

2.5 million units of service robots were sold in 2011 alone,

with an estimated 16 million further units to be sold between

2012 and 2015 [24]. The importance of robots is not only

recognized by the science community and the industry,

but also by governments. For example, the New Zealand

government has put the development of robots on their main

future agenda.

With this increased presence of robots in domestic life, the

question remains as to how these robots will be regarded by

humans—as mindless machines, as subservients, or as peers?

Previous work demonstrates that humans treat computers

socially, for instance, by tempering unpleasant feedback to a

computer, apparently so as to be more polite [19], [9]. Such

ﬁndings raise the question of whether it is possible for robots

to prompt human conformity.

Human Interface Technology Lab, University of Canterbury, P.O. Box

4800, Christchurch, 8140, New Zealand

New Zealand Institute of Language, Brain and Behaviour, University of

Canterbury, P.O. Box 4800, Christchurch, 8140, New Zealand

Please direct correspondence to J

urgen Brandstetter at

[email protected] or

eter R

acz at [email protected]

To investigate the persuasive power of robots, we ex-

pand on two landmark psychology experiments. The ﬁrst

experiment was conducted by Muzafer Sherif, who studied a

conformity effect in an ambiguous situation. The experiment

builds on the autokinetic effect, a phenomenon in which a

person perceives a sudden movement of a light point when

no reference is given. Sherif created this effect by sitting

participants in a dark room with one small light point visible.

The participants had to look at this point, and had to say

how much the light point moved. However, the light never

actually moved, but the autokinetic effect created an illusory

sudden movement. This movement is perceived differently

by different people. After he did the experiment with one

person, Sherif set two or three people in the same room

and asked them to say out loud how much the point moved.

Astonishingly, after three rounds all the participants said the

same number, even though everyone perceived a different

movement. This effect is called informational conformity or

social proof and describes the effect in which people in an

uncertain situation look at their neighbors to see what is

probably the right answer and conform with them. [7], [21],

[15].

A second experiment is reported by Solomon Asch in his

inﬂuential paper “Effects of group pressure upon the mod-

iﬁcation and distortion of judgements”. Asch builds on the

ﬁndings of Sherif to ﬁnd out whether humans also conform in

non-ambiguous situations. His experiment simulated a simple

visual line test. The participant saw three lines of different

heights, labelled A, B, C on the left side of a board, and one

reference line on the right side labelled with “?” (see Figure

1). The task was to say which line matched the reference line.

When the participants were alone in a room to perform the

test they almost always gave the correct answer. In a second

round Asch placed the participants in the same room with

other participants who were all actors and who all gave the

wrong answer. Even though the real participant – presumably

– knew the correct answer, in 32% of all tasks the participants

went along with the group if the group size was bigger than

four. In this case, Asch provided evidence that conformity

is not only inﬂuential in ambiguous situations but also in

non-ambiguous ones [2], [3], [4].

In the current study, we recreate the Asch experiment and

use aspects of the Sherif experiment as a model for further

research. We investigate a visual judgment task along the

lines of Asch [2], [3], [4], and extend the methodology

to a verbal production task. Verbal tasks may provide an

especially fruitful domain for conformity studies, because

language is inherently social, and speakers inﬂuence one

2014 IEEE/RSJ International Conference on

Intelligent Robots and Systems (IROS 2014)

September 14-18, 2014, Chicago, IL, USA

Fig. 1: An example of the lines task, where the participants

have to identify the corresponding line. Based on Solomon

Asch’s experiment.[2]

Fig. 2: An example of a word set presented in the word task.

The four confederates each produce the past tense form of

the ﬁrst four verbs. Participants then have to provide the

correct past tense form of the ﬁfth

another’s linguistic behaviors on an ongoing basis. The

perceived social characteristics of a speaker seem to play

a role in speech processing, and the social status of speakers

affects processes of language change[12], [13]. Moreover, it

is known that speakers tend to copy one another in natural

settings; individuals align their actions and representations,

interactively and subconsciously [10]. Speakers have many

choices available when they speak; for instance they may

often choose from different pronunciations of the same

word. In a conversational context, one speaker’s choice of

a particular linguistic variant typically increases the likeli-

hood that other speakers will choose the same variant. In

most cases, speakers accommodate to each other, and their

linguistic styles “converge”, as is found by observing many

linguistic domains including speech rate [22], word choices

[6], and choices of grammatical constructions [23]. Speaker

convergence is subject to social variables; speakers tend to

accommodate to one another unless they want to establish

social distance from their interlocutor(s) [1], [17], [18].

In the current experiment, we explore a form of linguistic

convergence between speakers with respect to morphological

variation, speciﬁcally involving the English past tense. In

some cases speakers may choose between various past tense

patterns, as in “dived” or “dove” for the past tense of “dive”,

or “slayed” or “slew” as the past tense of “slay”. Such

morphological pairs are known as morphological “doublets”;

individual speakers may alternate between different choices

in different circumstances [11], and some competing mor-

phological items have co-existed side-by-side for centuries

[8]. For instance, the Oxford English Dictionary indicates

that the past-tense forms “leaped” and “leapt” have coexisted

in English since at least the 1500s, as have “dreamed” and

“dreamt”. Since no single past tense form is universally

preferred, morphological doublets offer a verbal analogue to

ambiguous visual judgements. The proposal of the current

study is that the choice of one verb variant over another

is a behavior that may be inﬂuenced by other speakers.

That is, speakers who inﬂuence one another socially should

be expected to converge on the same past tense form (or

the same past tense verb pattern). Building on the above

experiments and our interests in Human Robot Interaction

and verbal communication, we developed an experiment

which includes the following conditions: ambiguous and

non-ambiguous line conditions, and ambiguous and non-

ambiguous verbal conditions. We ran these conditions with

four confederate robots (for one group of participants) or

four confederate human actors (for the other). Data in the

peer conditions is compared against matched baseline data,

gathered from participants in isolation. The ﬁrst question we

wanted to answer was whether participants conform more (or

less) to a robot group than to a human group. The second

question is whether conformity to verbal judgments can be

observed in a peer-pressure experiment, comparable to the

classic conformity studies by Asch. In the last question we

wanted to understand whether participants conform more to

the peer group in an ambiguous task than in a non-ambiguous

task.

For clarity, we summarize our questions here again:

Q1: Do participants conform more to the human group than

the robot group?

Q2: Do participants conform to a group’s verbal judgments

in addition to visual judgments?

Q3: Do participants conform more to the the group in an

ambiguous task than in an non-ambiguous one?

II. METHODS

We performed a 2x2x2 mixed-factor experiment. The

within-subject variables were the modality of the stimulus

(lines vs. words), the ambiguity of the stimulus (ambiguous

vs. non-ambiguous), and the ordering of the stimuli (lines

ﬁrst vs. words ﬁrst). The across-subject variable was the

setting: a group of human peers, a group of robot peers,

or alone (the baseline). The number of times the participants

conformed to their peers was the dependent variable. In the

baseline setting, an error in the visual task was the dependent

variable, and in the verbal task the selection of the regular

(-ed-ﬁnal) past tense form was the dependent variable. Such

responses match the group behaviors in the peer conditions,

as described below, and thus their occurrence in isolation

provides us with information about baseline behavior.

A. Design

We aimed to re-create the Asch and Sherif experiments

introduced above, expanding the visual decision task with

a verbal task and the group of human confederates with a

group of robots.

1336

The line projection consisted of three horizontal parallel

lines labelled A, B, C on the left side of the projection and

one line labelled with ? on the right side. An example can

be seen in Figure 1. The line on the right, marked with

a ? matches one of the lines on the left. The task for the

participant was to say which line, A, B, or C, matched the

line labelled ?, following the model of Asch’s experiment.

The word projection consisted of groups of ﬁve words.

The task was to say the word on the projection and its past

tense form, for example like and liked. Each participant had

to say the word corresponding to their position. This means

participant one would say the ﬁrst word and its past tense,

participant two would say the second word and its past tense,

and so on. An example word set is given in Figure 2.

Each experiment block (involving either lines or words)

was divided into three parts: a warmup (3 subject re-

sponses), an ambiguous-item set (15 subject responses), and

a nonambiguous-item set (15 subject responses). Each ex-

periment block thus consisted of 33 trials, although warmup

items are not analyzed, leaving 30 responses of interest in

each block, per subject.

In the warmup part, consisting of three unambiguous

tasks, the peer group would say the correct answer—thus

providing the correct visual judgment in the lines task, or

providing the normative past tense in the verbal task. In

this case we wanted to give the participants the feeling that

the robots/actors understood the question, and could provide

correct responses.

In the second and third parts of each experiment block, the

experiment confederates (robots or human actors) gave con-

sistent answers, with the intention of swaying participants’

responses away from their default choices. In the visual

task, the confederate actors/robots would always choose the

incorrect line response. In the verbal task, the confederates

always provided the regular past tense form, that is, a form

which ends in -ed.

In the verbal task, ’ambiguous’ items are stimulus verbs

for which variation exists in English usage, and the target

words can have a regular or irregular past tense form, for

example, dive — dived, or dive — dove. Non-ambigous

verbal stimuli are items which have only an irregular past

tense form, for example, run—ran. Note that this means that

in the peer-group conditions, the robots/actors all provided

verb forms which are not normatively in use, that is, over-

regularized forms such as runned as the past tense of run.

Additional details are provided in the ’Baseline’ sec-

tion below regarding about the designation of ambigu-

ous/nonambiguous items in the visual task.

We counter-balanced the order of the line and word

experiment blocks. We used pre-recorded introductions in

all conditions.

B. Apparatus

The setting consisted of a projector, a high quality wireless

microphone, a table with ﬁve chairs, four customized Alde-

baran NAO robots or four confederate actors, and a laptop to

control the recordings and experiment. The projecting area

Fig. 3: Plan of the physical setting of the test chamber.

had a dimension of 243x177cm, although the maximum line

length was 118cm. The table was exactly parallel to the

projecting area at a distance of 200cm. The dimensions of

the table were 80x250cm (see Figure 3).

To make our NAO robots more individualistic, we gave

every robot different outﬁts and voices. We did this in-

dividualization process on the assumption that if robots

were distinctive, participants would more easily see them

as individuals rather than as copies of each other. Since

we wanted to make sure that the robots’ appearance rather

than their voices constitutes the only difference between the

actors and the robots, we used recordings of the actors for

the robot voices. We also made sure that each robot (in the

robot condition) sat in the position corresponding to the actor

(in the human condition) whose voice it shared.

C. Procedure

So as to make the human and robot conditions as consis-

tent as possible, the start of the experiment was designed

so as to discourage spontaneous interactions between the

participant and the confederates. In both conditions, the

participant signed consent forms in a room separate from

the experiment room. The participant was then informed

that the other participants were already in the experiment

room and the study was ready to begin. When the participant

entered the second room, all of the confederates greeted him

or her (with a simple ’Hello’, ’Hi there,’ etc.) In the robot

condition, the robot confederates turned their heads toward

the participant when making the greeting, to give a sense

that the robots were aware of the participant’s presence. The

research assistant told the participant to ﬁnd an empty chair;

since the confederates were already seated, the participant

had no option but to take the last chair in the row (See Fig. 3).

After the participant sat down, the experiment immediately

started.

D. Baseline

We ran two baseline studies to provide reference points

for judgments in the visual and verbal tasks, in the absence

of any peer pressure.

The baseline setting for the visual task consisted of 107

line conﬁgurations. The setup of the room was exactly the

same as in the group experiment, the only difference being

that the participants did not say out loud which line best

1337

matched; they used the buttons 1,2,3 on the keyboard. We

assume there is no difference between saying the answer

out loud or pressing a button when no one is in the room

other than the participant. Out of these 107 settings, we

picked exactly 30 for our group experiment. The corpus

of 30 lines was divided into two parts. The ﬁrst part (15

settings) contained all the settings where people made most

of the mistakes. These lines were ambiguous and not easy

to distinguish. The second part of the lines contained a set

of lines where the two closest lines had no difference bigger

than 30% and people made only one or zero mistakes. We

chose this setting because Asch used a similar setting to

prove his hypothesis [2]. We used baseline responses to these

30 targets as a comparison to the group settings.

The baseline setting for the verbal task was the same as the

group setting, except that the participant had to produce past

tenses for all the forms on the screen. All items that were

presented in the peer conditions (to confederates as well as

participants) were included in the baseline, to control for pos-

sible priming effects. However, for purposes of quantitative

comparison between baseline and peer conditions, we only

analyze the 30 verbs for which participants in the peer groups

gave responses. We used consistent coding for the baseline

and the group settings, labeling only regularized responses

as ‘conforming’.

Unlike the group settings, the baseline settings for the

lines and the words were run with two different sets of

participants.

E. Participants

We had 78 total participants; 23 participated in the group

of human confederates, 19 in the group of robots, 18 in the

lines baseline, and 18 in the verbal baseline.

All participants were college students at the University of

Canterbury, and all were native speakers of New Zealand

English. It was important to use speakers from one dialect

region only, as there may be regional variation in preferences

for past tense forms.

No participants reported any previous familiarity with the

Asch experiment during debrieﬁng. All participants were

paid $10 for their participation. The experiment lasted about

20-30 minutes.

III. RESULTS

Conformity rates for all conditions are summarized in

Figure 4.

Inspection of this ﬁgure reveals a few interesting ﬁnd-

ings, starting with the baseline regularization of verbs —

’conformity’ in the verbal task when there are no peers

to conform to. Note that among non-ambiguous verbs, 5%

of baseline participants’ responses were in fact regularized.

Such responses represent spontaneous over-regularizations

of verbs which are not expected to deviate from their

irregular status: ﬁnd—ﬁnded, go—goed, know–knowed. Thus,

the regularizations in the baseline verbal task are markedly

more common than in written corpus data, based on searches

of large online print corpora such as the Google Books

Fig. 4: Percentage conformity for all conditions. The ’con-

formity’ in the baseline conditions represents the behavior

of interest (error in the lines task, and regularization in

the verbal task) by isolated subjects, and is thus used as

a reference point for matched items in the robot and human

peer conditions.

conform ∼ group x ambiguous + (1 + ambiguous | subject) + (1 + group

| item)

coef st error z sig

Intercept (baseline) -5.46 0.58 -9.396 ***

group=humans

3.91 0.74 5.267 ***

group=robots

0.55 0.83 0.668 n.s.

ambiguous=T

4.57 0.54 8.403 ***

group=humans x ambiguous=T

-2.51 0.65 -3.889 ***

group=robots x ambiguous=T

-0.18 0.73 -0.248 n.s.

TABLE I: Mixed Effects Model Summary

corpus [16]. This veriﬁes that for purposes of quantitative

comparison, a laboratory baseline study was essential for

interpreting our peer-group results. Production of a sequence

of English past tense forms is an admittedly artiﬁcial task,

whether performed in the presence of peers or in isolation,

and speakers may also prime themselves to regularize verbs.

In all conditions, there is a clear and strong effect of

conformity induced by the human peer group. In three of

the four conditions, the robots appear to elicit slightly higher

rates of the conforming judgements than the baseline data,

although the rates are markedly lower than for the humans.

However, in statistical analyses, we do not ﬁnd evidence

of signiﬁcant effects of peer inﬂuence by robots. A binomial

mixed-effects model was stepwise ﬁt on the data using max-

imal random effect structure. Non-signiﬁcant effects were

dropped from the model. The model summary is presented

in Table I. For purposes of economy, we only present the

overall model here, which encompasses conformity in both

visual and verbal tasks as the dependent variable. However,

separate models ﬁt independently to the lines and verbs data-

sets also show the same pattern of main effects for peer group

(human or robot) and ambiguity.

The mean per-subject conformity rates for the relevant

across-subject and within-subject settings can be seen in

1338

Figure 5. The degree of conforming responses is signiﬁcantly

higher for humans than for robots and baseline. Robots

and baseline are not signiﬁcantly different. This is true for

both the ambiguous and the non-ambiguous settings. The

degree of conforming responses are signiﬁcantly higher for

the ambiguous setting than for the non-ambiguous setting.

This difference is signiﬁcantly larger for the human group.

Whether lines come ﬁrst is not signiﬁcant. Neither is task

type (lines or words).

●

●●

●

●●

●

0.0 0.2 0.4 0.6 0.8 1.0

mean rate of conformity

basel.

robots

humans

basel.

robots

humans

non−ambiguous ambiguous

***

Fig. 5: Subject means for group setting and ambiguity: both

line and word tasks

IV. CONCLUSIONS/DISCUSSION

In this study, our main goal was to ﬁnd out if robots could

create conformity and if so, how much difference there is

between human peers and robot peers. The study uses the

idea from Sherif, who studied conformity in ambiguous situ-

ations and Asch, who studied conformity in non-ambiguous

situations.

In our ﬁrst question (Q1) we wanted to know if robot peers

elicit the same conformity as human peers. The analyses

point to a clear null result with respect to peer inﬂuence from

robots. When surrounded by robot peers, individuals’ behav-

ior shows some trend toward being inﬂuenced in the antici-

pated direction, but this behavior is not signiﬁcantly different

from the baseline behavior in isolation from peers. Subjects’

behavior alongside robot peers patterns more closely with

the baseline condition than it does with the human-peer

condition.

Of course, null results are always more difﬁcult to interpret

than positive results. One might call into question various

features of our setup: it is always possible signiﬁcant results

could arise if the sample size were larger, if the robots looked

or acted more human-like, or if the experimental instructions

were different. Additional investigation in these areas would

certainly be appropriate. We should reiterate, however, that

signiﬁcant results are evident across the board in our study

when the peers are human confederates. Moreover, our

analyses demonstrate that participants’ behavior among robot

peers is signiﬁcantly different from behavior among human

peers. Individuals clearly interpret robot peers differently

from human peers.

In our second question (Q2) we wanted to know if

participants conform to a group in verbal tasks as well

as in visual tasks. Our data indicates that individuals are

susceptible to social pressure in choosing past tense forms

— again, when the peers are humans, although not when they

are robots. This result supplements a wide range of ﬁndings

that speakers interactively align during dialogues, adding

morphological convergence to the inventory of different

linguistic levels for which convergence has been observed.

In our third question (Q3) we wanted to know if there

is more conformity in an ambiguous condition versus a

non-ambiguous situation. The ambiguous situation is often

referred to as social proof [7]. Social proof means that when

people are in an uncertain situation, they tend to look around

to see what their peers do. Conformity is clearly higher in

ambiguous cases, as veriﬁed by the regression model.

A. Future Work

There are many aspects of this work that warrant further

investigation. We are particularly interested in conducting

further work into understanding individual variation across

participants, and the types of peer pressure that different

individuals may be vulnerable to. We also intend to attempt

a conformity task which is closer to “natural” interaction, in

which participants are not just giving responses in isolation.

ACKNOWLEDGMENTS

This project was made possible through the support of a

Sub-award under a grant to Northwestern University from the

John Templeton Foundation (Award ID 36617). The opinions

expressed in this publication are those of the author(s) and

do not necessarily reﬂect the views of the John Templeton

Foundation. Hay and Beckner were also supported by a

Rutherford Discovery Fellowship awarded to Hay.

This project was approved by the University Canterbury

Ethics Committee: HEC APPLICATION 2013/111 and the

Northwestern University Ethics Committee: STU00085449.

A special thanks goes to the four actors; Timothy Pomroy,

Natalie Eustace, Julian Melchert, Morgana Mountfort-Davies

and to Latifah Almanea for recruiting.

REFERENCES

[1] C. Abrego-Collier, J. Grove, M. Sonderegger, and A. C. L. Yu. Effects

of speaker evaluation on phonetic convergence. In Proc. Int. Congress

of the Phonetic Sciences, Hong Kong, pages 192–195, 2011.

[2] S. E. Asch. Effects of group pressure upon the modiﬁcation and

distortion of judgments. Groups, Leadership, and Men., pages 222–

236, 1951.

[3] S. E. Asch. Opinions and social pressure. Scientiﬁc American, pages

31–35, 1955.

[4] S. E. Asch. Studies of independence and conformity: I. A minority

of one against a unanimous majority. Psychological Monographs:

General and Applied, 70(9):1–70, 1956.

[5] C. L. Breazeal. Designing sociable robots. MIT press, 2004.

1339

[6] S. E. Brennan and H. H. Clark. Conceptual pacts and lexical choice in

conversation. Journal of Experimental Psychology: Learning, Memory,

and Cognition, 22(6):1482, 1996.

[7] R. B. Cialdini. Inﬂuence : science and practice. HarperCollins, 2009.

[8] C. Fehringer. How stable are morphological doublets? A case study

of schwa/Ø variants in Dutch and German. Journal of Germanic

linguistics, 16(4):285–329, 2004.

[9] B. J. Fogg. Persuasive technology. Ubiquity, 2002(December):2, Dec.

2002.

[10] S. Garrod and M. J. Pickering. Joint action, interactive alignment, and

dialog. Topics in Cognitive Science, 1(2):292–304, 2009.

[11] L. R. Haber. Leaped and Leapt: a theoretical account of linguistic

variation. Foundations of Language, 14(2):211–238, 1976.

[12] J. Hay, P. Warren, and K. Drager. Factors inﬂuencing speech per-

ception in the context of a merger-in-progress. Journal of Phonetics,

34(4):458–484, Oct. 2006.

[13] D. J. Hruschka, M. H. Christiansen, R. A. Blythe, W. Croft, P. Heg-

garty, S. S. Mufwene, J. B. Pierrehumbert, and S. Poplack. Building

social cognitive models of language change. Trends in Cognitive

Sciences, 13(11):464–9, Nov. 2009.

[14] T. Kanda, T. Hirano, D. Eaton, and H. Ishiguro. Interactive robots

as social partners and peer tutors for children: A ﬁeld trial. Human-

computer interaction, 19(1):61–84, 2004.

[15] S. A. McLeod. What is Conformity? - Simply Psychology, 2007.

[16] J.-B. Michel, Y. K. Shen, A. P. Aiden, A. Veres, M. K. Gray,

J. P. Pickett, D. Hoiberg, D. Clancy, P. Norvig, J. Orwant, et al.

Quantitative analysis of culture using millions of digitized books.

Science, 331(6014):176–182, 2011.

[17] M. Natale. Convergence of mean vocal intensity in dyadic communi-

cation as a function of social desirability. Journal of Personality and

Social Psychology, 32(5):790, 1975.

[18] M. Natale. Social desirability as related to convergence of temporal

speech patterns. Perceptual and Motor Skills, 40(3):827–830, 1975.

[19] B. Reeves and C. Nass. The Media Equation: How People Treat

Computers, Television, and New Media Like Real People and Places

(Center for the Study of Language and Information Publication

Lecture Notes). Center for the Study of Language and Inf, 2003.

[20] G. Shaw-Garlock. Looking forward to sociable robots. International

Journal of Social Robotics, 1(3):249–260, 2009.

[21] M. Sherif. A study of some social factors in perception. Archives of

Psychology (Columbia University), 1935.

[22] J. Webb. Studies in dyadic communication: Proceedings of a research

conference on the interview. Pergamon, pages 115–133, 1970.

[23] E. J. Weiner and W. Labov. Constraints on the agentless passive.

Journal of linguistics, 19(1):29–58, 1983.

[24] Worldrobotics.org. World Robotics - Industrial Robots 2012 Executive

Summary. Technical report, 2012.

1340