Social Persuasion in Human-Agent Interaction
Yasuhiro Katagiri
Advanced Telecommunications Research Institute International
Toru Takahashi
ATR Media Integration & Communications Research Labs.
Yugo Takeuchi
Shizuoka University
Abstract
Social persuasion abounds in human-human interactions. Our attitudes and behaviors are invariably
influenced by attitudes and behaviors of other people as well as our social roles/relationships toward
them. We discuss, in this paper, the design of interface agents that considers the social aspects in
human-agent interaction. The underlying hypothesis of the design is that human social behaviors toward interface agents are on a par with those toward
humans. We report on two preliminary experimental studies focused on the nature and the effectiveness of social persuasion in human-computer interaction environments. We argue that social factors,
such as affiliation, authority and conformity, need
to be taken into account in interface agent design,
as they can have effective and persuasive power in
human-agent interaction.
1 Introduction
Interface agents that mediate between humans and computers are expected to play an increasingly important role in
“human-oriented” information systems. The research in intelligent interface agent technologies have been aiming to develop an electronic servant that serve to help people cope with
information overload created by the “machine-oriented” information technologies, by providing an effective means for
search, extraction and filtering of information [Maes, 1994;
Good et. al., 1999]. With their various personification features, interface agents are expected to alleviate the labor and
mental burden that people have been experiencing in complicated tasks of working with computers.
With the recent interest in anthromorphized agents or embodied agents, an emphasis is also placed on emotional and
relational aspects of interaction with interface agents [Nagao and Takeuchi, 1994; Ball et. al., 1997; Laurel, 1990;
Cassell, 2000; Cassell and Bickmore, 2000]. Multi-modal
information channels, e.g., the capabilities of exchanging information via voice, gestures, gaze and facial expressions in
addition to language, have the potential of providing us with
natural means to foster affection, pleasure and trust as well as
to exchange information with computers.
Reeves & Nass [Reeves and Nass, 1996] have convincingly
demonstrated that people’s spontaneous responses to computers follow exactly the principles of human-human social interactions and this social nature of human behaviors toward
computers are universal and consistent. Takeuchi&Katagiri
[1999] examined human responses toward animated computer agents and suggested the possibility for system designers to affect users’ behaviors by inducing human social interpersonal reactions toward interface agents.
We report, in this paper, on two experiments that focused
on the potentials of social persuasive functions of interface
agents. The first experiment investigated the possibilities and
effects of creating affiliative relationships between humans
and interface agents. The second experiment examined the
possibilities and effects of displaying inter-personal relationships through inter-agent interactions. Though both experiments produced only preliminary results, it seems they indicate that the social dynamics in interactions involving interface agents work as effectively and persuasively as in human
interactions.
2 Elements of Social Persuasion
2.1
Affiliation Need in Human-Agent Interaction
There is an essential need for humans to establish and maintain affinitive relationships with others. This need is called
the affiliation need. The affiliation need is important when
we consider human relations or group formation. People who
want to establish and maintain a friendly relationship with another person are apt to follow or sympathize with this person
[Schachter, 1959]. If such social relationships based on affiliation need are established between users and interface agents,
we expect that the user will try to establish and maintain a relationship with the agents.
To induce affiliation needs in users, we note the following
two types of interpersonal attraction of interface agents:
Appearance-based attraction: The interpersonal attraction
we feel based on the outward appearances (figures, vocal
qualities, etc.) and the internal traits (character, orientation, etc.) of the interface agents
Capability-based attraction: The personal attraction we
feel based on profits and benefits by interacting with interface agents
Establishing and Maintaining Affinitive Relationships
The appearance of a person has a great effect on first impressions. The appearance-based interpersonal attraction of an
agent will also have such an effect on the user’s first impression of the agent. This will induce the user’s affiliation need
in the first encounter. Therefore, the appearance of the interface agent is important in establishing the affinitive relationship between the user and the agent.
The strength of the affinitive relationships can vary. Cooperative or agreeable behaviors enhance the relationship,
whereas uncooperative or disagreeable behaviors damage it.
When people find that their partner can provide benefit for
them, they feel capability-based attraction toward their partner and try to enhance the relationship by increasing their affiliation needs. When their partner shows negative attitudes
toward them, they might even assume that the partner has
lessened the affiliation need toward them and try to compensate for this to maintain the affinitive relationship. They
will then feel an enhanced affiliation need toward the partner
and try to recover their former state of relationship by trying
to sympathize with the partner and respond positively to the
partner’s requests [Cialdini and Kenrick, 1976].
Human orientation of establishing and maintaining affinitive relationships can be adapted to interactions with interface agents. It should be possible to design the behavior of an
agent-interfaced system that implements life-like social interaction based on users’ affiliation needs.
2.2
Authority and Conformity in Inter-Agent
Interaction
Interface agents can lead us towards making accurate inferences about how an agent is likely to think, decide, and act on
the basis of its external traits such as its appearance, voice,
and communication style. Even though multiple interface
agents are simultaneously driven by the same program on a
computer, people attribute a specific identity to each agent
[Takeuchi and Katagiri, 2000]. This human response towards
agents suggests that people would regard agent-agent (interagent) interaction (such as conversations, social acts, or nonverbal communications) as having the same social dynamics
as human-human interaction. Virtual theatrical inter-agent interaction can be accepted as a natural social activity similar
to that of humans.
Expertise and Authority
Authority is inherent in the actions exercised by those who
are experts and have special knowledge and skills. People,
therefore, frequently prefer to believe news sources that have
higher expertise1 rather than others [Reeves and Nass, 1996]
even though the news itself is completely the same. This human response is entirely the consequence of people accepting
authority accompanied by expertise as an attribution of information. Furthermore, in situations where people need special
knowledge or skills, the person who is respectfully treated
with reverence seems to have higher expertise.
1
e.g. CNN, HNN, SNC, and RNN
Figure 1: Information Kiosk and PalmGuide
Conformity
The attitudes and behaviors of other people or groups frequently change our own attitudes and behaviors. Conforming
one’s attitude and behavior to a person or group who can exercise authority or to an influential power, is a sensible strategy
for receiving further benefits as a basic social skill in general [Asch, 1951]. In the public arena, moreover, we prefer to choose a strategy to conform our interpersonal behavior style with other people in order to avoid social friction.
These strategies are effective in social life. Persons or groups
have the competence to change the actions of others based on
expectations, and this is power accompanied with authority
according to its influence [Tawney, 1931].
3 Experience from Human-Agent
Interactions
3.1
Exhibition Guidance System
We first studied the affiliation need based social interaction
between humans and agents by incorporating a set of interface agents into the Exhibition Guidance System C-MAP
developed in ATR. The C-MAP (Context-aware Mobile Assistant Project) Exhibition Guidance [Sumi, 1998] features a
personal mobile assistant that provides visitors touring exhibitions with information based on user contexts. The user
of the system carries a hand-held guidance system called
PalmGuide while touring an exhibition. PalmGuide maintains user contexts such as user’s name and affiliation, temporal and spatial situations of the exhibition, history of the
tour, and personal interests (Figure 1).
A personal guide agent runs on the user’s PalmGuide and
provides tour navigation information such as introduction of
exhibit articles and recommendations of what to visit next.
This recommendation is made from the result of calculating
the user’s interest by using information of other users’ contexts.
At exhibitions using this C-MAP system, information terminals called Information Kiosks are installed at each ex-
Figure 2: Appearance of the Agents
hibit. Information Kiosks usually provide information about
the overall exhibition. When a user with a PalmGuide comes
to an Information Kiosk, the user can connect her PalmGuide
to the Information Kiosk by infrared communication. Then,
the user’s guide agent migrates to and personalizes this Information Kiosk.
The guide agent on the Information Kiosk gives interactive
guidance of the particular exhibit with animated motion and
synthesized voice. After the user finishes obtaining information with the agent’s guidance, the agent returns to the original PalmGuide and goes to the next exhibit with the agent’s
user.
3.2
Interpersonal Attraction
To solicit user affiliation needs toward interface agents, we
have designed the following two aspects of interpersonal attraction:
Appearance-based attraction
Each user selects the appearance of her personal interface agent. She can freely choose
her favorite from a selection of nine prepared
agents (Fig. 2).
Capability-based attraction
The interface agent has the capability of providing its user with integrated information
based on her context, as the agent comes along
with its user and constantly revises her personal data and interests from her tour history.
3.3
Observational Analysis
We analyzed the affiliation need based interaction by examining users’ behaviors toward recommendations made by their
personal guide agents.
Setting
We installed and ran the C-MAP Exhibition Guidance System
at ATR Open House, and observed the behaviors of the visitors who used the guide agent system. The visitors received
their PalmGuides and were allowed to freely roam around the
exhibits. A total of 21 Information Kiosks were installed at
the exhibit booths. We compared the observed users’ tour
histories with the recommended exhibits.
We designed the behaviors of the guide agents as follows:
Figure 3: Acceptance rate of recommendation (Means/SE)
1. Until the fourth access to an Information Kiosk, the
guide agent returns to the PalmGuide when the access
ends and comes along with the user to the next exhibit.
2. At the end of the fourth access, after giving the user a
recommendation for the next exhibit, the agent tells the
user that it will go ahead directly to the recommended
exhibit and wait for the user there, and disappears from
the Information Kiosk.
Depending on the user’s following behaviors, two types of
situations follow at the fifth access to an Information Kiosk.
A. The user accesses the recommended exhibit. The agent
thanks the user for her trust in the agent’s recommendation.
B. The user accesses an exhibit different from the recommended one. The agent complains that the user doesn’t
follow the agent’s recommendation.
In either case, we expect that the agent’s reaction induces its
user’s affiliation need and has an effect on the user’s subsequent behaviors. In situations of type A, we expect her affiliation need toward the agent is enhanced because of its positive
response. In situations of type B, we expect her affiliation
need to also be enhanced, because the user worries that the
social relationship with her agent would take a turn for the
worse and hopes to recover the relationship.
Results and Discussions
We separated out two subject groups ( and ) out of the
users of the system.
consists
of 22 users who accessed
a Kiosk just four times and
consists of 12 users who accessed a Kiosk more than five times. Psychological
ratings
based on questionnaire
survey
indicated
that
group
rated
higher than
group of their attitudes toward both agents
and interactions. Although statistically significant difference
cannot be observed because of the small sample size, it is
consistent with the view that the feedback of the guide agent
in the fifth access is influential in setting user attitudes.
Figure 3 compares the acceptance rates of recommendations. The leftmost
bar shows the mean acceptance
accep
tance rate for . Right two bars are rates for
before and
after each user obtained feedback from her agent based on
the user’s response to the agent’s recommendation in the fifth
access. Left two bars show a similar behavior tendency, but
the rightmost bar shows a higher rate than others. The figure indicates that people tend to behave differently toward the
guide agents’ recommendations before and after the fifth access, when the agents simply went ahead and waited for their
users at the recommended Information Kiosk. The data of the
observational
exhibited
analysis
only a statistical tendency
( ), mostly because we haven’t
been able to get a large enough number of users suitable for
analysis. Nevertheless, we believe this difference in behaviors is important as it reflects the strength and effectiveness
of the users’ affiliation needs toward guide agents in selecting behaviors.
Guide agents complained to 11 users out of 12 because
these 11 users didn’t access the recommended exhibit’s kiosk
directly after the end of their fourth access. Therefore, by
comparing the data, we can conclude that the guide agents’
action enhanced the users’ affiliation needs toward the agent.
The users hope and try to recover the affinitive relationship
with the agent. Namely, we can explain the changes in user
behaviors as being controlled by the enhancement of affiliation needs, which are effected by social interactions between
the users and their agents.
4 Experience from Inter-Agent Interactions
4.1
Interface Agents in Web-based Instruction
In order to further examine and verify the social nature of
human-agent interaction, we expand the scope of interaction
to include agent-agent as well as human-agent interaction.
We developed a prototype system of Web-based multi-agent
instruction to elucidate the effects of social factors such as
authority and conformity in human-agent interactions and to
investigate methodologies to incorporate those factors to better the design of instructional systems.
A Multi-Agent Instruction Scheme
In the traditional Web world, the authors of the Web pages
have an inventory of visual (and auditory) design parameters,
such as choice of color, sound, layout and style, to control
the salience of a particular page and to preserve coherence
across several pages. They manipulate those parameters to
make their Web contents more comprehensible and persuasive. Animated interface agents provide them with a huge
possibilities to further extend the expressiveness of Web contents.
We started exploring this extended expressiveness provided by interface agents by looking at the following simple
multi-agent instruction scheme.
1. Each user is associated with an agent called the Conductor Agent (CA). The CA always accompanies the user
and guides her through the Web pages.
2. Each set of Web pages for a meaningful unit of Web contents is associated with an agent called the Expert Agent
(EA). The EA always resides in its content pages and
gives descriptions of their information contents whenever a user visits its pages.
3. One of the tasks of the CA is to mediate the interactions
between the user and various EAs. The CA introduces
Figure 4: A Multi-Agent instruction scheme for Web contents.
the user to an EA upon entering its pages. These interagent interactions are theatrically demonstrated on the
Web pages in front of the user (Figure 4).
We assign the jobs of CA and EAs to the character agents
shown in Figure 2
The scheme above was intended to focus on the following
two tendencies of human social responses:
Expertise People prefer to accept information from sources
assumed to have higher expertise.
Conformity People prefer to choose a strategy to conform
to the interpersonal behavioral style of other people with
supposed equal social positions to themselves.
The guiding intuition was that the style of interaction between
a CA and an EA, together with the user’s tendency for conformity, influences the user’s attribution of expertise to the EA,
which then contributes to the user’s assessment of authority
in the content of the pages the EA is in.
Polite interaction The CA politely interacts with an EA. By
conformity to the interaction style of the CA and the EA,
the user recognizes the authority of the Web page contents corresponding to the EA.
Casual interaction The CA casually interacts with an EA.
By conformity to the interaction style of the CA and the
EA, the user do not attribute authority to the Web page
contents corresponding to the EA.
In order to theatrically realize the two interaction styles, we
manipulated the CA’s behavior towards EAs as follows:
Verbal expressions either deferential or colloquial to the
EA.
Modest or intimate attitude to the EA.
Distinguished verbal expressions and attitudes directed
towards the participant.
4.2
Experimental Analysis
Multi-Agent Instruction for Cooking
A prototype system was developed in the cooking instruction
domain to help the intended user accurately understand the
contents and adequately carry out the cooking steps instructed
in the contents of Web pages. Social inter-agent interactions
in the two contrasting styles were theatrically demonstrated in
the system to see if they make any changes in user attitudes
and performance.
Tips
(1)
(2)
(3)
(4)
(5)
(6)
%
%
%
%
%
%
%
%
%
%
Participants
!
%
%
"#
%
%
%
"$
%
%
Table 1: Knowledge about cooking tips.
(a) Inter-agent interaction in introduction phase
(b) Instruction giving phase
Figure 6: Custard puddings produced in the Polite condition (left)
and in the Casual condition (right).
Figure 5: Interactions in a multi-agent cooking instruction.
We prepared a series of Web pages for a custard pudding
cooking recipe. The learners can actually make custard puddings as they go through those pages and watch social interagent interactions that take place in the pages.
Four pages were created for each of the cooking steps, and
an EA was assigned to each page, as an expert of the corresponding cooking step. A CA accompanies the user as she
goes through the four steps. When the user enters a page
to initiate a new cooking step, the CA introduces her to the
EA before the EA starts the instruction. The user is required
to follow the instruction, which is given both in vocal and
document forms, to successfully complete the step. Figure 5
shows the introduction and the instruction phases in a cooking instruction step.
Method
We conducted an experiment to explore whether the behavior
and the attitudes of people are affected by social inter-agent
interaction in Web pages. The inter-agent interactions took
place in the form of greetings and requests between CAs and
EAs. At the start of each cooking step, the CA vocally speaks
to the EA in either of the following two interaction styles:
Polite condition The CA politely introduces the user to the
EA and asks him in a modest manner to give instructions
for cooking to the user.
Casual condition The CA casually introduces the user to the
EA and asks him in a colloquial style to give instructions
for cooking to the user.
The participants of the experiment were instructed to cook
a custard pudding by going through the steps in the Web instructions. A questionnaire test was performed after the ses-
sion to see if participants correctly acquired the cooking tips
given in the instructions. Performance was also assessed from
the actual cooking outcomes.
Three participants were assigned to the Polite condition
and two participants to the Casual condition. The five participants were not proficient in baking cakes, cookies, or pastries
such as custard pudding, although they could use computers
well and stated that they usually browse various Web pages.
Predictions
One’s social authority is characterized from how the surrounding people treat her as well as her appearance and behaviors. We assume that, in our experiment, people recognize
the EA’s authority and conform to the CA’s treatment of the
EA. The results can therefore be predicted as follows:
(P1) The participants assigned to the Polite condition will
score higher than those assigned to the Casual condition
in a test of knowledge in making custard puddings.
(P2) A better cooking performance will be achieved by the
participants under the Polite condition than those under
the Casual condition.
Results and Discussions
Table 1 shows the number of cooking tips given by the EAs
and also described in the corresponding Web pages which
were correctly answered in the post cooking questionnaire
test. Table 1 shows that three participants ( &$'()&+* ) assigned
to the Polite condition acquired more tips correctly than the
two participants ( , ' and ,.- ) assigned to the Casual condition.
The exemplary cooking outcomes in the Polite and the Casual conditions are shown in Figure 6. There is an obvious
difference in the performance between the two conditions.
The surface of the left pudding is smooth and visually goodlooking
/ compared to the right pudding. There were also obvious differences in taste between them. These assessments
seem to indicate that the performance of achieving the given
task was influenced by the conditions of whether CA respectfully interacted with EAs.
Both the acquired cooking knowledge result and the cooking performance result confirmed our predictions P1 and P2.
The people in the Polite condition achieved better than the
people in the Casual condition both in terms of knowledge
and performance. These results support our hypothesis that
people recognize expertise and authority in inter-agent interactions between the CA and the EAs, and adjust their attitudes
and behaviors in conformance to them.
5 Conclusions
We reported on two preliminary experiments focusing on the
persuasive potentials of interface agents. The results of both
of the experiments appear to suggest that social interaction
with and between interface agents have the persuasive power
to influence people’s attitudes and behaviors, though we need
more participants in experiments to get to statistically clear
evidences. The interface agent, designed to induce users’ affiliation needs, did actually change user behaviors in the first
experiment. Inter-personal relationship, e.g., authority, displayed by one agent toward another did change social attitudes and behaviors of people in the second experiment.
These findings have significant importance for HumanComputer interface design. Current computer interfaces frequently require us to follow explicit messages produced by
computers, e.g., “Insert a floppy disk into FDD,” “Drop these
files into trash box,” or “Click this button to confirm,” to
which users are forced to comply. Users might feel stressed
in this style of system operation, as they are always obliged
to obey the messages. The affiliation need of a user toward
her computer might be effectively applied to maintain intimate relationships. Once a reciprocal relationship between
human and computer is established, people strive to keep
their intimate relationship. When computers blame users for
their insincerity, negligence, or other unsocial behaviors toward computers, users would tend to obey computers’ requirements to compensate for the damages to their intimate
relationships, with positive feelings of cooperating with the
computer. People are also sensitive to social power such as
expertise and authority, and they would adjust their behaviors
accordingly, even if the power is merely theatrically acted out
among interface agents. Interface designers must be aware
of these potentials to enhance the quality as well as the efficiency of human-computer interactions.
Acknowledgments
We are grateful to Dr. Yasuyuki Sumi for cooperating with us
in using the C-MAP Exhibition Guidance System. We also
sincerely thank Ms. Keiko Nakao who made a great effort in
designing attractive appearances of the agents.
References
[Asch, 1951] S. E. Asch. Effects of group pressure upon the
modification and distortion of judgments. In H. Guetzkow,
editor, Groups, leadership, and men. Carnegie Press, 1951.
[Ball et. al., 1997] G. Ball et. al. Lifelike computer characters: the Persona project at Microsoft. In M. Bradshaw,
editor, Software Agents, pages 191–222. 1997.
[Cassell and Bickmore, 2000] J. Cassell and T. Bickmore.
External manifestations of trustworthiness in the interface.
Communications of the ACM, 43(12):50–56, 2000.
[Cassell, 2000] Justine Cassell. More than just another pretty
face: embodied conversational interface agents. Communications of the ACM, 43(4):70–78, 2000.
[Cialdini and Kenrick, 1976] R. B. Cialdini and D. T. Kenrick. Altruism as hedonism: A social development perspective on the relationship of negative mood state and
helping. Journal of Personality and Social Psychology,
34:907–914, 1976.
[Good et. al., 1999] N. Good et. al. Combining collaborative
filtering with personal agents for better recommendations.
In Proceedings of AAAI’99, pages 439–446, 1999.
[Laurel, 1990] B. Laurel. Interface agents: metaphors with
characters. In B. Laurel, editor, The art of humancomputer interface design. Apple Computer, 1990.
[Maes, 1994] P. Maes. Agent that reduce work and information overload. Communication of the ACM, 37(7):31–40,
1994.
[Nagao and Takeuchi, 1994] K. Nagao and A. Takeuchi. Social interaction: Multimodal conversation with social
agents. In Proceedings of AAAI’94, pages 22–28, 1994.
[Reeves and Nass, 1996] Byron Reeves and Clifford Nass.
The Media Equation. Cambridge University Press, 1996.
[Schachter, 1959] S. Schachter. The psychology of affiliation.
Stanford University Press, 1959.
[Sumi, 1998] Y. Sumi. C-MAP: Building a context-aware
mobile assistant for exhibition tours. In T. Ishida, editor,
Community computing and support systems, pages 137–
154. Springer, 1998.
[Takeuchi and Katagiri, 1999] Yugo Takeuchi and Yasuhiro
Katagiri. Social character design for animated agents. In
Proceedings of RO-MAN99, pages 53–58, 1999.
[Takeuchi and Katagiri, 2000] Yugo Takeuchi and Yasuhiro
Katagiri. Identity attribution in social interaction with
agents (in Japaese). Correspondences on Human Interface, 2(3):91–98, 2000.
[Tawney, 1931] R. H. Tawney. Equality. Harcourt Brace,
1931.