u
GETTING PERSONAL WITH
COMPUTERS :
HOW TO DESIGN
PERSONALITIES FOR AGENTS
D. CHRISTOPHER DRYER
IBM Research, San Jose, California, USA
Recent research indicates that people respond socially to computers and perceive them as
having personalities. Software agents are artifacts that particularly embody those qualities
most likely to elicit social responses : fullling a social role, using language, and exhibiting
contingent behavior. People’s disposition to respond socially can be so strong that they may
perceive software agents as having a personality, even when none was intended. T he
following is a discussion about intentionally designing personalities for social agents. T o
design personalities, it is necessary to consider the nature of personality and its role in
interactions between people and artifacts. In addition, a case study of designing a social
software agent is presented. T he conclusions from this experience are summarized as
guidelines for future agent developers.
Personality is a fundamental linchpin of social relationships. In th e cont ext
of human interaction, people automatically and unintentionally organize the
behavior of their partners into simplifying traits (Uleman et al., 1996), and
people tend to agree about which par tners are best described by particular
traits (Moskowitz, 1988). Beyond categorization, personality shapes the very
natur e of social relationshi ps, even impacting how satisfying an interaction
is for the participants (Dr yer & Horowitz, 1997).
It is not surprising therefore that researchers and engineers have started
think ing about personal ity as essential to the design of social agents for user
interfaces. Indeed, people may perceive automatically a personality in any
social agent, regardless of whether the agent’s designer intended a person
-
ality. This suggestion presents some new opportunities for human
-
machi ne
interacti on designers. Rather than have unint ended personalities shaping
Part of this article appeared as ‘‘Ghosts in the machine: Personalities for s ocia lly adroit software
agents,’’ by D. Christopher Dryer, pp. 31 36, in K. Dautenhahn (ed.),
Socially Intelligent Agents
, AAAI
Press Technical Report F S-97-02, American Association for Arti® cial Intelligence, Menlo Park, Califor-
nia, 1997.
This wor k was in part the product of the CSLI interface laborato ry at Stanford University, the New
Technolo gy team at the Walt Disney Company, an d the WarpGuide team at IBM , including Ron
Barber, Roger Didio, Steve Idhe, Bob Kamper, Bob Kelley, M arion Lindsey, Craig Moser, Paul Rogers,
Ted Selker, John Tyler, Les Wilson, and Julie Wright. In addition, Wendy Ark, Renee Fadiman, Ronaldo
Mendoza, and three anonymous reviewers ma de important contr ibutions to an earlier draft. IBM and
OS/2 Warp are trademarks of the I BM Corporation.
Address correspondence to D. Christopher D ryer, Intern at ional Business Machines Corporation,
Almaden Research Center, 650 Harry Road, San Jose, CA 95120, U SA. E-mail : dryer@ almaden.ibm.com
Applied Arti® cial Intelligence, 13 : 273 295, 1999
Copy righ t 1999 Taylor & F rancis
Ó
0883-9514/99 $12.00
1
.00
273
274 D. C. Dryer
people’s interaction s with machin es, designers can intentionally create per
-
sonalities that support satisfying social interactions. This task depends on
understa nding the nature of per sonality, the role of personality in people’s
interacti ons with arti
®
cial agents, and how agent personalities can be
created.
THEORETICAL BACKGROUN D : THE N ATURE OF
PERSO N ALITY
A personality is the collection of individual di erences, dispositions, and
temperaments that are observed to have some consistency across situations
and time. Beyond this basic de
®
nitio n of pers on ality , di erent researchers
have thought ab out the p henomenon in various ways. For example, perso n
-
ality can be seen as only the measurable consistencies in behavior (e.g.,
Mischel, 1968). Person ality also can be seen as the perceived consistencies in
people, independent of whether those consistencies are actual (e.g., Shweder,
1975). Alternatively, personality can be seen as an interaction between the
expectation s of an observer and the behavior of the observed (e.g., Cantor &
Mischel, 1979). Where intentionally created personalities are concerned,
however, the behavior of the observed has been designed wit h the expecta
-
tions of an o bs erver in mind. T he personal ity of a user interface agent does
not matter outside of the perception of an observer. From this perspective,
the actual and perceived personalities can be a ssu med to be the same.
People unconsciously use personality as a tool to org an iz e otherwise
overwhelmin g information about social partners. Successful social behavior
negotiates the goals, beliefs, and emotions of multiple social partners. These
negotiations occur in most domains of hum an life, including collaborative
work, recreation, and family life. People use their perceptions of th eir
partn er’s personality to mak e predictions about what that person will do
and h ow that is di erent from what other people will do. Without person
-
ality and the ability to use personality to organize these predictions, social
behavior would break down.
The variety of human personalities is in
®
nite. Mapping the space of all
of the po ssible personalities seems like an impossible task. Indeed, there are
thousands of nonsynonymous trait adjectives, such a s shy, adventurous,
obsessive, domineering, feminine, and anxious.
The task, however, is simpler than it seems. While there are thousan ds of
di erent facets to person alities , only a few factor s really matter at a more
abstra ct level. The
®
ve important factors (for review, see M cCrae and John,
1992) can be described as (1) agreeable (cooperative to competitive), (2)
extroverted (outgoing to withdrawn), (3) neurotic (anxious to calm), (4) con
-
scientiou s (organized to lax), and (5) open (curious to closed minded). Cer
-
Designing Personalities for Agents 275
tainly , there are other things to know about a partner, but nearly all of them
covary with one of these
®
ve dimensions or some combination of them.
These factors result from empirical investigations into which sets of indi
-
vidual di erences (among samples of all possible i ndivi du al di erences)
covary . All of the theoretically derived personality schemes that have been
tested, such as the Myers
-
Briggs (Croom et al., 1989), can be shown to be
simply rotations of some or all of the Big Five dimensions (Furnham, 1996 ).
Of the
®
ve factors, the most important for social interactions are those
that concern individual di erences in social behavior, namely, agreeableness
and extroversion or their co mmon rotations, ‘‘friendliness’’ (friendly to co ld)
and ‘‘dominance’’ (authoritative to submissive) (Wiggins, 1979). These two
factors, proposed by Aristotle in R hetoric, ha ve been reinvented and
renamed by various theorists (e.g., Leary, 19 57 ; Benjamin, 1974 ; S tr on g &
Hills, 198 6).
The factors and some typical labels are shown in Figure 1 : the two
-
dimension al interpersonal spac e. This two
-
dimensional model has been pro
-
posed as a ‘‘nomological net’’ uniting various interperson al research
progra ms (Gurtman, 1991 ; Wiggins & Broughton, 1985). For example,
researchers have found the two
-
dimensio nal structure to be fundamental to
interpersona l traits (Conte & Plutchik, 1981 ; Wi ggi ns, 1979), needs (Wig gin s
& Broughton, 1985), problems (Horowitz, et al., 1988), relations (Wish et al.,
1976 ; Dryer et al., 1994), emotions (Knutson, 1996), and g oa ls (Dryer &
Horowi tz , 1997).
F IGURE 1. The two-dimensional interpersonal space.
276 D. C. Dryer
Because personality helps people organize t he behavior of others, it also
guides people’s responses to their social partners. People di er in the kinds
of personalities they want their partners to have, and the mat ch between two
people’s personalities impacts how enjoyable and productive the interaction
is (Dryer & Horowitz, 1997 ; Horowitz et al., 1997). In general, people like
partn ers that have ‘‘positive’’ personalities. Each of the
®
ve personality
dimension s has a positive (i.e., socially desirable) pole and a negative pole.
Personalities that are described at the positive end of the pole are better
liked, and vice versa ; all else being equal, people would rather interact with
cooperat ive, outgoing, calm, organized, and curious partners than competi
-
tive, withdrawn, anxious, lax, and close
-
minded pa rtners.
A personality, however, can be too positive. P artners with personalities
that are extremely positive on all
®
ve dimensions may be less attractive than
partn ers who have a negative personality attribute along at least one dimen
-
sion. People
®
nd it easier to identify with a ch aract er w ho has some weak
-
ness or
¯
aw that makes them seem human ; ‘‘perfect’’ characters are
unnat ural and seem less lifelike (Thomas & John so n, 1981). Similarly,
researchers (Aronson et al., 1966) have found that partners with a foible are
liked better than those without.
Personalities that are similar to a person’s ow n are also well
-
liked
(Byrne, 1971). For example, friendly people
®
nd friendly pa rtners more
attrac tive than cold people do, and cold people
®
nd cold partners mo re
attrac tive than friendly people do.
There are exceptions, however, to the maxim that ‘‘birds of a feather
¯
ock togeth er ;’’ sometimes, ‘‘opposites attract.’’ When similarity is paired
with undesirable characteristics (such as unattractiveness or evidence of
mental disturbance), people ac tu ally prefer dissimilar pa rtners over similar
partn ers (Novak & Lerner, 1968 ; Taylor & Mettee, 1971). Likewise, person
-
alities are liked if they make up for something positive that you d o not have.
People prefer partners who want to take on dissimilar rather than similar
roles in a relationship (Kroko et al., 1988). Also, when a task requir es one
person to take charge, a dominant person will prefer to work with a less
domin ant partner, and a submissive person will prefer to work with a more
domin ant partner (Dryer & Horowitz, 19 97). In special situations like these,
people prefer partners wh o are complementary to themselves rather than
similar to themselves.
EM PIRICAL IN VESTIGATION S : THE ROLE OF
PERSO N ALITY IN PEOPLE’S IN TERACTION S WITH
ARTIF ACTS
Through a number of research p rogra ms , psychologists have character
-
ized the ways in which people fundamentally perceive their social partners
Designing Personalities for Agents 277
along the dimensions of friendliness and dominance. One interesting
research question is whether people perceive the personalities of arti
®
cial
agents in the same way.
This hypothesis can be considered within the framework of social inter
-
face theory (for an overview, see Reeves & N ass, 1996). Social interface
theory is built on the results of various studies demonst rating that people
respond socially to machines. These studies suggest that people’s responses
to technologically medicated so ci al partners can be the same as their
responses to natural partners. People are inclined to treat everything as
social and natural. Therefore people automatically and sub con sci ously lever
-
age whenever possible what they know about their natural and social experi
-
ences to help th em with their technological experiences.
Softwa re agents in particular are likely to encourage social responses
because of three important features. (1) Th ey can use full
-
sentence text, in
addition to the typical user interface forms o f communication, like menus,
contr ols, and icons. Full
-
sentence text is more natural, especially for people
with a more verbal than nonverbal cognit ive style (Horn & Cattell, 1966). (2)
Softwar e agents can embody task knowledge as well as use arti
®
cial intelli
-
gence to reason about when and how to engage a person in interaction. This
gives them a compelling kind of contingent behavior. (3) Software agents can
auton omo usly perform actions on a person’s behalf. These features
-
natur al
langua ge (text), contingent behavior (intelligence), and social role
(autonomous assistance)
-
are the three factors that predispose people to
respond socially (Reeves & Nass 1996).
Various tests o f social interface th eory have provided evidence that
people apply apparently irrelevant knowledge about social interactions to
their i nt eraction s with machines. For example, people are polite to agents
(Nass et al., 1997b), use social rules to explain their behavior (Nass et al.,
1994), apply social stereotypes to them (Nass et al., 1997a), treat them as
teammat es (Nass et al., 1996a), and look for social mo ti ves when agents
¯
atter or criticize (Fogg & Nass, 1997). Given the importance of personality
in people’s social behav ior, w e wondered whether perceiving a personality is
anoth er social response th at people have when interacting with an a gent.
This q uestion was
®
rst addressed with a laboratory study concerning
social responses to computers (Dryer et al., 1993a). In this study, 44
comp uter
-
literate participants interacted with p erson al computers through a
traditional
-
style graphical user interface. A computer
®
rst tutored partici
-
pants on a series of topics (social customs, mass media, and computers). The
partic ipants were tested with a number of questions on the topics, ostensibly
to determine how successful th e tutoring was.
The participants then had a
®
nal interaction with a computer. In this
interacti on, a computer reviewed each of the test questions. For each ques
-
tion, the computer
®
rst indicated whether the pa rticipant had answered the
