Adaptive Agents and Personality Change: Complementarity versus similarity as forms of adaptation

Youngme Moon & Clifford I. Nass

Dept. of Communication
Stanford University
Stanford, CA 94305-2050

ABSTRACT

The idea that computer agents should be adaptive is a well-accepted tenet in the software industry. The concept of adaptivity is rarely defined in explicit terms, however. On the one hand, adaptivity could mean change in the direction of similarity; on the other hand, an agent could adapt in the direction of complementarity. The question for software developers is, Which type of adaptivity -- similarity or complementarity -- does the user prefer? To investigate this question, a laboratory experiment was conducted (N=88). Results indicate that, consistent with the gain-loss literature in the field of social psychology, subjects preferred interacting with a computer that became similar to themselves over time.

KEYWORDS:

Adaptivity, Agents, Complementarity, Personality, Similarity, Social Psychology

INTRODUCTION

People prefer to interact with others who share their personality type. This social psychological rule has been empirically established to the extent that it is often called the "law of attraction" [2]. Over the past forty years, it has been borne out in studies on everything from marital satisfaction to roommate relationships. In a recent study, it has been shown that this same social rule applies to computers: People prefer to interact with computers that share their personality type [3]. More specifically, dominant people prefer to interact with dominant computers; submissive people prefer to interact with submissive computers.
The present study expands upon this finding by predicting the effects of changes in personality behavior over time. By definition, personality is generally conceived of as a stable attribute and thus, the effects of personality change in people have often been overlooked. With respect to computers, however, the question is of great consequence given the ongoing interest in adaptivity in computers. Indeed, the idea that computer agents should be adaptive is a well-accepted tenet in the software industry.
However, the concept of adaptivity is rarely defined in specific terms. What does it mean to be adaptive? On the one hand, adaptivity could mean change in the direction of complementarity; that is, the agent's personality complements the user's personality over time. On the other hand, adaptivity could also mean change in the direction of similarity, such that the agent's personality becomes more similar to the user's personality over time. The question under investigation in the present study is, "Which type of adaptivity -- complementarity or similarity -- does the user prefer?"
The gain-loss theory of attraction [1] in the field of social psychology predicts that changes in rewarding behavior from another have a greater impact on an individual than consistent, invariant reward. With respect to personality change, this leads to the prediction that people will be more attracted to individuals who change from dissimilar personality behavior to similar personality behavior, compared to individuals who are consistently similar in personality behavior over time. In addition, this theory predicts that a change from dissimilar personality behavior to similar personality behavior is preferred over change in the opposite direction.
Will the same social rule apply to interactions with computers? If so, then users will be more attracted to computers that change from being dissimilar to being similar in personality type, compared to computers that are either consistently similar, or become dissimilar, over time.
The present study thus investigates whether gain-loss theory can be used to successfully predict the effects of personality-based behavioral changes in computers. Note that the application of social psychological theory to the study of human-computer interaction is not unprecedented. Previous research has provided evidence that computers are treated as social actors even when the users know that the machines do not actually possess feelings, "selves," or human motivations [5].

METHOD

44 dominant subjects and 44 submissive subjects were randomly assigned to one of four conditions in a 2 x 4 balanced, between-subjects design. (Dominant and submissive subjects were categorized by the use of a standard personality survey, conducted several weeks prior to the experiment.) Upon arrival, the subject was told that he or she would work with a text-based computer to complete a task. The subject was told there would be two rounds of interaction. Depending on condition, subjects were randomly matched with a computer that was either dominant or submissive in the two rounds. Thus, there were four possible combinations of computer personality type: dominant-dominant, dominant-submissive, submissive-submissive, or submissive-dominant computer. All conditions were balanced for gender.
The manipulation of dominance and submissiveness was identical to that explained in previous publications [4], consisting solely of language-based cues and expressed confidence levels.
Following both rounds of interaction, the subject was given a questionnaire to fill out. The questionnaire asked the subject for an assessment of both the computer and the interaction itself. The dependent variables were four indices: Social Attraction, Intellectual Attraction, Utility, and Emotional Satisfaction.

RESULTS

The gain-loss prediction was that subjects would be more attracted to a computer that changed from dissimilar personality behavior to similar personality behavior ("gain"), compared to a computer that remained similar in personality behavior over time. Consistent with this prediction, for all four dependent variables, orthogonal comparisons among means showed that attraction ratings were significantly higher in the "gain" condition than in the consistently-similar (CS) condition (See Table 2).

Table 2. Means and orthogonal comparisons, Gain condition versus Consistently-Similar condition.

		  	    Conditions		   Difference
			_____________		   _________
	
Dependent Variable		Gain		 CS		    Gain-CS


Social attraction		6.06		5.12			 .94*

Intellectual attraction		7.03		6.30			 .73*

Utility				7.20		5.98			1.22*

Emotional satisfaction		6.92		6.07			 .85*


* p<.05.

In addition, for three of the four dependent variables, subjects were more attracted to a computer that changed from dissimilar personality behavior to similar personality behavior ("gain"), compared to a computer that changed in the opposite direction over time ("loss"). (See Table 3.) Again, these results are consistent with gain-loss predictions.

Table 3. Means and orthogonal comparisons, Gain condition versus Loss condition.

		   	   Conditions		   Difference
			_____________		   _________
	
Dependent Variable		Gain		Loss		    Gain-Loss


Social attraction		6.06		5.91			 .15

Intellectual attraction		7.03		5.96			1.07*

Utility				7.20		5.88			1.32*

Emotional satisfaction		6.92		5.81			1.11*


* p<.05.

DESIGN IMPLICATIONS

This study has several important implications for design. First, users prefer computers that share their personality type; thus, products should be segmented by personality. Second, users prefer adaptive computers over computers that remain consistently similar to them over time. And finally, the question of how best to define adaptivity -- in terms of similarity or complementarity? -- is resolved. Clearly, adaptivity is best-defined in terms of change in the direction of similarity to the user's personality type, rather than change in the direction of complementarity.

REFERENCES

Aronson, E., & Linder, D. (1965). Gain and loss of esteem as determinants of interpersonal attractiveness. Journal of Experimental Social Psychology, 1, 156-171.
Byrne, D., & Nelson, D. (1965). Attraction as a linear function of proportion of positive reinforcements. Journal of Personality and Social Psychology Bulletin, 4, 240-243.
Nass, C. I., Moon, Y., Fogg, B. J., Reeves, B., & Dryer, C. (in press). Can computer personalities be human personalities? International Journal of Human-Computer Studies.
Nass, C. I., Moon, Y., Fogg, B. J., Reeves, B., & Dryer, C. (1995). Human-computer interaction as interpersonal interaction. Proceedings of the CHI Conference, Denver, CO..
Nass, C. I., Steuer, J. S., & Tauber, E. (1994). Computers are social actors. Proceedings of the CHI Conference, Boston, MA.