Social Psychology of Persuasion Applied to Human Agent Interaction

An Interdisciplinary Journal on Humans in ICT Environments ISSN: 1795-6889 www.humantechnology.jyu.fi Volume 4(2), November 2008, 123–143 123 SOCIAL PSYCHOLOGY OF PERSUASION APPLIED TO HUMAN–AGENT INTERACTION Abstract: This paper discusses and evaluates the application of a social psychologically enriched, user-centered approach to agent architecture design. The major aim is to facilitate human–agent interaction ( HAI ) by making agents not only algorithmically more intelligent but also socially more skillful in communicating with the user. A decision-making model and communicative argumentation strategies have been incorporated into the agent architecture. In the presented content resource management experiments, enhancement of human task performance is demonstrated for users that are supported by a persuasive agent. This superior performance seems to be rooted in a more trusting collaborative relationship between the user and the agent, rather than in the appropriateness of the agent’s decision-making suggestions alone. In particular, the second experiment demonstrated that interface interaction design should follow the principles of taskorientation and implicitness. Making the influence of the agent too salient can trigger counterintentional effects, such as users’ discomfort and psychological reactance. Keywords : human-agent interaction, user-centered design, decision making, persuasion. INTRODUCTION Wide employment of agents in human–computer interaction ( HCI ) design has proven to be an effective way to construct robust yet flexible software architecture, in which information communication between the user and the technical system is mediated by many kinds of agents. The new interaction paradigm, evolved from traditional HCI, can be called human–agent interaction ( HAI ). In HAI , users are provided with a novel social collaborator during their tasks: the software agent (Wooldridge & Jennings, 1995). Obviously, this new interaction element opens a series of design considerations. At its core, HAI invites a more consequent evaluation and application of social psychological concepts to guide the agent’s behaviors during interaction. © 2008 Shenghua Liu, Sacha Helfenstein, and Ari Wahlstedt, and the Agora Center, University of Jyväskylä URN:NBN:fi:jyu-200810245833 Sacha Helfenstein Agora Center University of Jyväskylä Finland Ari Wahlstedt Department of Computer Science & Information Systems University of Jyväskylä Finland Shenghua Liu Department of Computer Science & Information Systems University of Jyväskylä Finland Liu, Helfenstein & Wahlstedt 124 Hence, a key question is what we can learn from social interaction research in the human context in order to design user-friendly, adaptive, and effective HAI (e.g., Nass & Moon 2000; Reeves & Nass, 1996). This exploitation of social psychological concepts in interaction design is a logical extension of the user psychological approach to human–technology research (Moran, 1981; Oulasvirta & Saariluoma, 2004)—a paradigm approach that is especially effective in projects where the product or technology is new or where the audience characteristics and habits are not yet well defined (Goschnick & Sterling, 2002; Murray, Schell, & Willis, 1997). Thus, it is ideal for contemporary HAI research pursuing psychologically-based, integral agent architectures (Pasquier, Rahwan, Dignum, & Sonenberg, 2006; Rahwan, 2005). In this vein, it is essential to evaluate core issues such as interpersonal communication, influence, persuasion, and decision making in interaction (e.g., Cialdini, 1984; Eagly & Chaiken, 1984; McGuire, 1969, 1985; Petty & Cacioppo, 1981; Sewell, 1989; Zimbardo & Leippe, 1991). For instance, one may argue that the way by which agents interact with people should be considerate and comfortable. This applies naturally to the presentation or communication of information to the user in general. Another criterion is to design for effective, that is, influential agent support of user deeds. Because one of agent’s crucial roles is that it can enhance or substitute human user decision making when encountering points of judgment during system tasks, the exertion of persuasive influence is quite central. Therefore, communication skills, including argumentative rhetoric, dialogue strategies, and verbal proficiency, are highly relevant for the agent to effectively argue for its decisions, and to achieve the user’s trust. The present research concerned agent communication skills and the influence of such in the buildup and sustainment of a trusting collaborative relationship between the user and the agent. The core interest was the agent’s ability to effectively persuade users during decisionmaking tasks in system interaction (e.g., Fogg, 2003; Parise, Kiesler, Sproull, & Waters, 1999; Stiff & Mongeau, 2002; Stock, Guerini, & Zancanaro, 2006). Persuasive design is an important complement to the traditional usability concept because it specifically addresses socioemotional dimensions of interaction. One problem with the traditional usability perspective is that it presupposes user need or motivation to utilize a tool. However, with the mushrooming of technological solutions, and ever-increasing functionalities built into them, it becomes of growing importance not just to allow users to do in a simple and effective manner what they essentially want to do, but to go above and beyond that, to influence their desire and inclination regarding what they want to do or use. Agents, as sophisticated extensions to the interaction interface, are of core concern in this context. Agents are conceptualized as supports to user tasks in various ways, but often their use is based on a freedom-of-choice model. As a result, legitimate concerns are not so much whether users can in principal profit from agent use or in what way agent support is beneficial but, rather, whether users are willing to make use of the agent and how this is expressed in human–agent collaborative decision making. Hence, HAI is a suitable and interesting subject in persuasive interaction design research, especially in the context of the spreading relevance of agent technology in industrial applications (e.g., Luck, McBurney, Shehory, & Willmot, 2005; Wooldridge & Jennings, 1995). Previous research has, for instance, investigated the potential of recommendation agents for electronic shopping to influence the human decision making by shaping user preferences (Häubl & Murray, 2001). Other research projects, such as those pertaining to the RPD - Persuasive Human-Agent Interaction 125 enabled (recognition-primed decision) agent, focus on supporting decision-making teams by anticipating information relevant to their decisions based on a shared mental model (Fan & Yen, 2004). The results indicate that human teams, when supported by agents, can perform better in highly time-sensitive situations. Pasquier et al. (2006) developed an argumentation framework for an agent that is best suited to persuade other agents in a particular situation with a given standpoint. Social psychological insight is hereby applied to help in the exploration of belief/decision formation within a single agent and “social” interaction among many agents (Rahwan, 2005), yet not agent–user interaction. Finally, Katagiri, Takahashi, and Takeuchi (2001) reported on two preliminary experimental studies focusing on the nature and the effectiveness of social persuasion in HCI environments. In these types of studies, social factors, such as affiliation, authority and conformity, have been taken into account in interface agent design. Nguyen, Masthoff, and Edwards’ (2007) experiment also suggests that dialog-based systems with the visual appearance of a conversational agent are preferred over systems that use text only. The former are perceived to be more personal and caring, less boring, and, to some extent, easier to follow. However, in spite of these valuable efforts, more research on the issues of collaboration and persuasion in HAI is needed. CONCEPTUALIZING A PERSUASIVE AGENT IN A DECISION-MAKING TASK ENVIRONMENT The major paradigm for the distribution of interaction roles between users and machines is that humans make decisions while machines carry out automated processes and routines. However, with user task complexities increasing and technology becoming more sophisticated, machines progressively enter the domain of decision making. Agents are the premier example for this latter development, geared at supporting user decision making on different authority levels, from merely offering useful information, to serving in advisory functions (e.g., decision support systems), to making decision in place of the user. In the type of suggestive agent architectures explored here, users should maintain the dominant role during interaction; that is, the agent leaves the final decision to the user but tries to persuade the user to accept its decision. As noted above, this also means that HAI confronts users with a new type of decision, in addition to the ones concerning the actual use task: whether and how to utilize the agent and its suggestions during interaction. Whereas agent development traditionally focused on the technical elements, such as highly sensitive and thorough algorithms, somewhat less effort has gone into understanding how users can be convinced to utilize an agent’s suggestions and support offers. Logically, however, the latter must precede the former concern. And thus, we need to be concerned regarding how to make agents not only algorithmically intelligent, but also socioemotionally so. Ultimately, agent participation in system interaction tasks should award users with processing capacity and accuracy benefits. However, gaining people’s trust and will to collaborate are vital requirements for smooth and effective HAI . It is expected that people may doubt or feel reluctant to accept an agent’s decision or the information the agent provides, especially when they are not convinced and cannot validate the trustworthiness of the agent. On the other hand, due to the limited information-processing capacity of the human working memory, users may also feel tempted to overrely on agent recommendations in order to reduce Liu, Helfenstein & Wahlstedt 126 the cognitive load of decision making (Häubl & Murray, 2001). This tendency may backfire on well-functioning HAI within the context of system disturbances or other disagreements between user and agent assessments. In any case, acceptance or refutation of a given agent’s standpoint may critically depend on the availability and form of presentation of the information. Generally, we believe that the way agents influence people’s thinking and behavior must exploit evidence from human–human interaction skills. This belief is based on the fact that people regard HAI as having similar social dynamics as human–human interaction (Katagiri et al., 2001; Nass & Moon, 2000; Reeves & Nass, 1996). Consequently, HAI design needs to consider the social dimension and significance of such interface agent traits as appearance, voice, and communication style (e.g., Guadagno, Blascovich, Bailenson, & Mccall, 2007; Hargie, 1997). Architecturally, the present design of a persuasive agent was based on the BDI (beliefs, desires and intentions) agent model (Rao & Georgeff, 1992), which calls upon mental notions to encapsulate the hidden complexity of the inner functioning of an individual agent, and was further developed into a Procedure Reasoning System ( PRS ) by Ingrand, Georgeff, and Rao (1992). Although the cognitive BDI agent and its PRS applications have been well studied within the recent decade (Brazier, Jonker & Treur, 2002; Georgeff, Pell, Pollack, Tambe, & Wooldridge, 1998; Huhns & Singh 1998; Ingrand, Chatila, Alami, & Robert, 1996; Maes, 1994), current HAI research still continues in the design of the social BDI agent that has the ability to model human behaviors (Guzzoni, Cheyer, & Baur, 2007; Lokuge & Alahakoon, 2005; Pasquier et al., 2006; Peebles & Cox, 2006). In a broad sense, our research aims to enhance the agent’s ability by incorporating social communication skills into the BDI agent decision model structure (see Figure 1). In this structure, the plan library holds the rules that ultimately govern decision making based on input and in accord with its beliefs and goals components, while the intentions component formulates the agent’s decisions. Here, the architecture is essentially improved by Figure 1. The architecture of the persuasive agent (adapted from Ingrand et al., 1992). Persuasive Human-Agent Interaction 127 adding a decision model and an argumentation model to the intentions component so that the agent can argue for its decisions in a persuasive manner. In operationalizing the decision and argumentation model outputs, we relied on various insights gained from original social psychology research. Considering that HAI is situated in the wider context of a user’s task-oriented interaction with a system, time and processing resources are two obvious constraints to persuasion. This argues for the use of the heuristic model of persuasion (Chaiken, 1980) and Petty and Ciacoppo’s (1981) “central versus peripheral framework.” According to these theories, people’s compliance with appeals often follow simple decision-making rules that are based on little-evaluated (contextual) persuasive cues, such as the likability of the message source, the connotations of expertise, or social (e.g., majority) reference. Other design-relevant findings concerning social collaboration are the positive correlation of people’s willingness to cooperate with the frequency and richness of communication: For example, a visible collaboration partner is preferred over an invisible one (e.g., Deutsch, 1958; Wichman, 1972). In addition, friendliness, social liking, and request justification, among many others, have clearly shown to positively affect cooperation motives (e.g., Cialdini, 1984; Langer, 1978; Swingle & Gillis, 1968). In contemplating such findings, we propose a design model for a persuasive agent that comprises five communication skill-relevant dimensions: agreeableness, anthropomorphism, informativity, persuasiveness, and adaptivity. They are explained more fully here. 1. Agreeableness. The agent should be agreeable, which is in line with most conventional criteria in interface design, requiring a friendly appearance and eloquent communication style. As an effect, users should feel willing and comfortable in interacting with the interface agent. 2. Anthropomorphism. Agents with human traits may be more attractive to users than machine-like ones. The anthropomorphic representation allows for a rich set of easily identifiable behavioral cues for social interaction (Hargie, 1997; King & Ohya, 1996; Takama, Dohi, & Ishizuka, 1998). Especially at initial exposure to the agent, these representations may make the agent seem more intelligent, capable of a higher level of agency, and more trustworthy. 3. Informativity. The agent’s advice or decisions should be useful (necessary and sufficient) to the user and justified. Thus, when there is need for more information, the agent should provide more assistance. It is also valuable to accompany suggestions with some explanation or rationale. However, exhaustive information should be avoided since it can cause people to become impatient or overloaded. 4. Persuasiveness. Although all dimensions of communication skills affect the agent’s persuasiveness, this dimension is concerned with more specific elements inducing social influence. Foremost, these include persuasive cues pertaining to the influence schemes of request justification, reciprocation, commitment and consistency, social proof, liking, authority, and scarcity (Cialdini, 1984). Persuasive behaviors should nevertheless be subtle enough so that people will not feel that they are being directed and apparent enough that people can understand it. 5. Adaptivity. Collaborative style and skill are usually not predefinable in absolute terms but must evolve and adapt to the HAI setting, especially the style and preferences of the user. Therefore, all of the above-mentioned dimensions outlining Liu, Helfenstein & Wahlstedt 128 the communication skills of the agent must be modifiable to take into account user goals and actions. In order to apply this model to interface agent interaction design, and test the effect of agent communication skills on user persuasion, user–agent collaboration, and task performance, we constructed two experimental scenarios. Within such, we varied different HAI parameters according to the communication skill-relevant dimensions presented above. EXPERIMENTATION Investigating psychological dimensions of HCI is always tricky, especially when they pertain to socioemotional issues. A key challenge concerns the apt calibration of the task nature (e.g., difficulty) and setting in order to relate the elicited user actions and reactions in reliable ways to the experimentally chosen variations in stimuli, and not to peripheral (artifactual) or even external influences. Our research primarily focused on the HAI element of persuasion. This emphasis on user motive and intention, rather than mere necessity and sophistication of implemented agent support, demands a task setting that favors the aspect of user preference over that of task operation requirements. For instance, it is a rather intuitive result that people would rely on agent support for a very complex task, where agent information processing capacities are evidently superior. However, when users notice (or have) almost no alternative to system or agent reliance, the value of persuasive design is obviously greatly undermined. The real challenge is to persuade users in the context of relatively uncomplicated task requirements, because then users are given a true choice between working autonomously or collaboratively. On the other hand, it is worth noting that seemingly easy, repetitive tasks may pose their own performance challenges, due to impending lack of concentration or users’ underestimation of the task demands. The task chosen for the current study was easy enough (in terms of reasoning demands) to make the choice realistic, yet sufficiently challenging (in terms of operational demands, such as speed) in order to substantially afford the allocating of the user’s attention to either the agent’s suggestions or autonomous performance. This dilemma in deciding whether and how to employ the agent opens up an ideal influence space for persuasive agent design measures. Experiment 1 Participants in our experiment needed to manage materials in a learning content management system (LCMS), with and without the help of various communicatively skillful interface agents. In order to make the task more natural, we also simulated agent performance failures. Our general expectation was that users would benefit from agent assistance and, beyond this, that a persuasive agent design will be superior in its enhancement of HAI collaboration and task performances compared to a nonpersuasive agent design. Specifically, we assumed a communicatively skillful agent design to promote trusting HAI , which potentially jeopardizes user performance when the agent does not perform at optimal level, yet would enable quick collaboration restoration after such disruptions. Persuasive Human-Agent Interaction 129 Method Participants. The experiments were conducted in English by competent non-native speakers of English. We employed three experimental groups ( H : no agent support, HAI : nonpersuasive agent support; and HA2 : persuasive agent support), each containing 10 volunteer participants with university education backgrounds. There were 19 males and 11 females, aged 21 to 35, balanced across the experimental groups. Materials. The task scenario was described to the participants as follows: A resource manager (i.e., human user) maintains a LCMS database that holds a number of local content files that frequently need to be updated. The user’s task is to evaluate incoming files, that is, updates to existing files, and decide whether the current local copy should be replaced by the update or not (i.e., confirm or not confirm the update). A new update announcement was received every 10 seconds, 72 in total. Each update was visible in a list for 30 seconds, after which it expired. During these 30 seconds, the participant needed to evaluate the incoming update, by comparing its attributes to those of the related local copy, and reach a confirmation decision. The unique relation of an incoming update to a file in the local database was expressed by a common ID . Further, the files had four attributes: date (the publication date, e.g., 13/06/2007 ), size (the file size, e.g., 1200 units), rating (i.e., the recommendation rank of this update, from 1 as the lowest to 5), and cost (the total cost of managing the file, e.g., 400 units). The ideal update for confirmation had four positive attribute values: It was newer, smaller, higher ranked, and cheaper than the local copy of the corresponding file. A clearly undesirable update had four negative attribute values; it was older, bigger, lower ranked, and more expensive. As a rule of thumb, participants were instructed to confirm an update when there were more positive attribute values than negative ones, compared to the local copy. Otherwise it was wise to not confirm the update. The interface of the experimental group receiving the persuasive agent is depicted in Figure 2, using a female avatar (Elina) to symbolize the agent. No avatar was used in the nonpersuasive agent condition. The general interface further contained a table displaying the active incoming updates, interaction buttons and, in the agent-supported conditions, there was a text field containing the agent’s messages. All user actions were performed with a mouse. For each update, the two main operations participants could do were to confirm the update, by clicking the CONFIRM button, or to not confirm the update, by clicking NOT CONFIRM . In the agent-supported conditions, participants could also click the DO AS AGENT SAYS button to adopt the agent’s suggestion. The psychological aim of this button was to attract and underscore the momentum of direct user-agent collaboration. Design. Participants were naive to the purpose of the experiment, as well as to which experimental group they belonged. We employed a design with agent-support conditions as a between-subject factor (i.e., H , HA1 , HA2 ) and three distinct task phases as within-subject factor applied to HA1 and HA2 . As illustrated in Figure 3, group H performed all 72 update decision tasks under unchanged conditions. In HA1 and HA2 , however, three task phases were differentiated: Phase 1, updates 1-24 (normal agent support), Phase 2, updates 25-48 (disrupted agent support, including misleading suggestions), and Phase 3, updates 49-72 (normal agent support). From the point of view of collaborative and trusting HAI , we may denote Phase 1 as Liu, Helfenstein & Wahlstedt 130 Figure 2. Interface in group HA2 (persuasive agent) for Experiment 1. Figure 3. Experimental design: 3 experimental groups H (no agent), HA1 (nonpersuasive agent), and HA2 (persuasive agent), across experimental Phases 1-3. the collaboration and trust build-up phase, Phase 2 as the collaboration and trust disruption phase, and Phase 3 as the collaboration and trust restoration phase. In order to implement distinct agent behavior, the phases in HA1 and HA2 differed with regard to the rules the agent used in deriving its decisions. We defined two main rules: R0 (i.e., confirm the update when there are more positive attribute values than negative ones, otherwise not confirm) and R1 (i.e., confirm the update when a selected attribute compares positively to the local copy). In Phases 1 and 3, the agent always made decision by using default R0, and its suggestions can thus be considered as correct throughout. In Phase 2, the agent made decision by using the adaptive R1 . This meant that the agent’s suggestion were at times incorrect and therefore misleading (e.g., one attribute was positive, but the three others were negative). Which of the four attributes the agent selected depended on the most frequently Persuasive Human-Agent Interaction 131 featured positive update attribute in the confirmatory decisions the participant made during Phase 1. For example, if the users had made 10 confirmatory decisions and 7 of these updates featured a newer date, which was more than for any other attribute, then the date attribute was selected to express R1. The rational behind this R1 design was to keep agent failures as credible and suggestive as possible. In sum, the experimental task for group H was straightforward and did not include any agent support. Agents in HA1 and HA2 had the same adaptive decision-making algorithm but differed in the communicatively skillful presentation of arguments used to persuade users. Compared to the pragmatic agent in HA1 the agent in HA2 was more human-like and more eloquent in “promoting” its intentions to participants. Communication skills varied in line with four of the five dimensions presented earlier (see Table 1), the exception being adaptivity, to maintain equality between the agent-supported persuasion groups in this dimension. . In HA2 the agent presented its suggestions by elaborating reasons in logical argumentation statements, set forth in a persuasive, friendly, personalized, and convincing way. Elina (the agent) also introduced herself before the task started, with the intention to obtain a sympathetic impression. Persuasive cues, like “certain,” “most safe,” “ideal choice,” “should,” “best,” and expert reference, were used to make the statement more impactful. We sought to keep the argumentative message structure consistent throughout the experiment because of its effect on increasing the processing speed and thereby enhancing persuasiveness. Feedback was implemented as the final element in the communication skill and persuasion factor. In HA1 , a brief verbal feedback was given only when participants instantiated their decision by clicking CONFIRM or NOT CONFIRM . The feedback simply restated the user’s decision. The agent in HA2 reacted specifically when the participant clicked the DO AS AGENT SAYS button by saying, “Good. Well done!” but did not give feedback when participants made their decision by clicking CONFIRM or NOT CONFIRM . The intention of this was to further promote direct compliance with the agent through a kind of positive, reciprocal reinforcement. Procedure: Experimentation was implemented on a PC in a quiet laboratory. Participants first received written instructions of the task and the decision rules. The actual test session lasted 12 minutes, after which the posttest questionnaire was completed and participants were debriefed. The questionnaire items inquired about general reactions to the experiment, such as task complexity and time pressure, as well as four main aspects of their HAI experience (agent-supported conditions only): agreeableness, anthropomorphism, persuasiveness, and usefulness. Finally, HA2 participants also answered questions concerning the agent’s appearance. Table 1. Agent’s Social Communication Skills in the Two Experimental Groups. Agent in Group Dimension of Communication Skills HA1 : Non-persuasive HA2 : Persuasive Agreeableness neutral friendly, supportive feedback Anthropomorphism no appearance, robotic human face, personalized (Elina) Informativity low (suggestion-only) high (decision justification) Persuasiveness no (suggestion-only) high (persuasive cues, facial cues) Adaptivity minimal minimal Liu, Helfenstein & Wahlstedt 138 nonsignificant interaction effect, there were strong indications that participants in the agentoriented interface condition suffered from persuasive agent messages, as compared to nonpersuasive ones, while this relation was contrary in the task-oriented condition (see Figure 6). Message persuasiveness showed no main effect on decision-making correctness. Concerning task speed for all updates, we found a strong overall tendency for persuasive messages to increase task time, F( 1,22 ) = 4.11 , p = . 06 . This seemed easily explainable based on the lengthier message texts involved, and affected both groups in a comparable manner (i.e., no group or interaction effect was found). However, from a psychological perspective, it was interesting to look at the correct and incorrect decision-making outcomes separately, since only the latter would trigger cognitive dissonance due to the implication that the user had to disagree with the agent’s suggestion. As apparent from Figure 6 (right side), time-cost increase for persuasive messages was limited to and strongest for participants in the agent-oriented interface interaction design, who had to explicitly disagree with the agent’s suggestion. During correct decision making, persuasive messages, although richer in syntax, actually decreased interaction costs (F( 1,22 ) = 9.62 , p < . 01 ), here in comparable manner for both groups. Finally, looking at the posttest questionnaire results, we found that the task, as intended, was judged overall as challenging, yet not too difficult. Participants did not indicate that they sensed the flawless functioning of the agent, which supported the natural validity of the experiment. Two thirds said that they could keep up their concentration until the end. We further found that the majority of participants experienced the agent’s assistance as helpful and rather comfortable, and thus not annoying or distracting, or strongly seductive as such. Concerning group differences, and consistent with our expectations, we found nevertheless a significantly lower level of interaction comfort in the group with agent-oriented interface interaction design, U = 29.58 , Z = 2.71 , p < . 05 . No other significant findings emerged. Figure 6. The comparison of correctness and time cost for incorrect decision making between the agentoriented and task-oriented groups for different agent message types (1: persuasive; 2: nonpersuasive). Persuasive Human-Agent Interaction 139 Discussion Both the number of updates handled and the decision-making correctness diminished in the second experiment, as compared to the first, which reflected the increased complexity of the task (i.e., nearly double the number of update attributes to consider). The agent’s message type (persuasive vs. nonpersuasive), which alternated randomly for each participant during the test session, did not seem to affect user decision-making correctness. This would confirm our view that the positive performance effects of HAI collaboration depends on the relatively slow process of trust buildup based on consistent persuasive agent behaviors. Persuasiveness is not visibly effective when message types vary arbitrarily. However, although both experimental groups received the same updates and agent messages, participants in the agent-oriented interface interaction design chose more frequently to disagree with the agent, resulting in inferior task performance. This was very interesting, considering that this group had the opportunity to click simply the AGREE WITH AGENT button every single time in order to score a 100% correct result. We interpret this finding in terms of psychological reactance, that is, the natural human reluctance to accept external authority and the need to retain a sufficient level of autonomy during decision making (cf. Brehm’s theory of psychological reactance, 1966). The finding was also in line with these participants’ more critical evaluation of HAI comfort. Further, the results suggested that reactance is especially fueled by persuasive agent messages, and thereby also triggering interaction time costs. Rejecting the agent’s decision by clicking the DISAGREE WITH AGENT button, of course, raised the salience of the contradictory action and may have been experienced as psychologically more distressing. The length of the persuasive messages had no negative effect on decision-making speed when the user’s decision conformed to that of the agent. Indeed, we found here that persuasive messaging actually increased interaction speed. The findings give rise to an important HAI -design criterion, according to which the nature of collaboration with agents should be subtle rather than explicit or overly salient. We should avoid situations in which users experience the agent as a central actor, and thereby retain a human-centric feel to the interaction. GENERAL DISCUSSION AND CONCLUSIONS The presented empirical facts substantiated our claim that HAI could be enhanced through the use of elements of persuasive agent design, for example, persuasive message cues familiar from research on social influence. The results overall demonstrated that a communicatively more skillful agent can boost user task performance without the legacy of slower interaction time. The most appropriate interpretation of these effects hinted at an augmentation of users’ trust and collaborative attitude regarding the employment of the agent. The findings of Experiments 1 and 2 suggested that absolute performance superiority induced by the persuasive agent grew steadily from update to update and phase to phase, and that beneficial task performance effects were not tied to single persuasive messages alone. This means, as in social reality, collaboration buildup is not instant, but evolves gradually. Hence, it is a Liu, Helfenstein & Wahlstedt 140 persuasive agent using consistent, social psychologically sophisticated communicative cues, not the persuasive messages per se, which achieves the best effects. The findings also revealed two valuable constraints or downsides to persuasive agent design. The first pertains to agent inaccuracies and system failures or disruptions, which are always a real threat in HCI settings. In such circumstances, trusting HAI may, as was shown, have negative performance implications. Nevertheless, the findings of Experiment 1 also showed that such disruptions, even if the agent’s suggestions are outright false, do not need to be catastrophic. Indeed, positive collaboration can be restored well by a persuasive agent. The second challenge concerns establishing a social psychologically apt level of persuasive influence. Experiment 2 in particular suggests that the interaction design should refrain from making the persuasive nature and collaborative demands too salient or agentcentered. Human users want to retain a healthy degree of autonomy; influence exertion easily can go overboard and trigger user discomfort and reactance. In sum, the effectiveness of HAI is often questioned on the level of trust that people would grant to agents. This means that research on agent design should focus not only on the agent’s algorithmic sophistication in solving a problem, but also on its ability to communicate in an apt manner with human users. The present work explicates an improvement to the conventional BDI agent structure by incorporating two important models into the intentions component (see Figure 1): decision making and argumentation. The decision-making model helps the agent formulate its intention according to the input from the environment and the reasoner’s actions. The argumentation model handles the presentation of the intention to the user, applying social psychologically based communication skills in order to make the agent’s arguments more persuasive. This kind of persuasive agent design, if applied with consideration, is user-task effective and best suited for building trusting, long-term HAI relationships. We challenge future research to generate more insight into these issues. In particular, we encourage extension of our work by attending to matters such as user personality effects, including user-versus-agent-avatar gender interaction. ENDNOTES 1. Outliers, defined as observations departing more than 1.5 inter-quartile ranges from the first and third quartiles, were excluded from analysis. 2. In groups HA1 and HA2 , the agent also provided misleading suggestions, depending on the decision rule extracted from participant’s prior decisions. Interestingly, we found that all participants followed without exception the date rule, meaning that the date attribute featured most frequently as a positive attribute in those updates the user confirmed. It also meant that all participants received the same number (8) of misleading suggestions for the identical updates during Phase 2 (see Figure 3). This conformity was despite that other update attributes (i.e., size, rating, and cost) were more likely to be featured in the confirmatory updates. We explain this in terms of a primacy effect (Asch, 1946): Because the date attribute was displayed as the first (furthest left) attribute in the interface, it might have been the first to be compared between the file’s local copy and its update. 3. In the posttest questionnaire, only the agent-supported group participants were asked about the appearance of an agent. HA2 participants were asked about the human appearance of the agent in their testing, while HA1 participants were asked what they would have thought about an agent with human appearance. Persuasive Human-Agent Interaction 141 REFERENCES Asch, S. E. (1946). Forming impressions of personality. Journal of Abnormal and Social Psychology, 41, 258–290. Brazier, M. F., Jonker, M. C., & Treur, J. (2002). Dynamics and control in component-based agent models. International Journal of Intelligent Systems, 17, 1007–1047. Brehm, J. W. (1966). A theory of psychological reactance. New York: Academic Press. Cialdini, R. (1984). Influence: The new psychology of modern persuasion. New York: Quill Publishing. Chaiken, S. (1980). Heuristic versus systematic information processing and the use of source versus message cues in persuasion. Journal of Personality and Social Psychology, 39, 752–766. Deutsch, M. (1958). Trust and suspicion. Conflict Resolution, 2, 265–279. Eagly, A. H., & Chaiken, S. (1984). Cognitive theories of persuasion. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 17; pp. 268–361). New York: Academic Press. Fan, X., & Yen, J. (2004). Modeling and simulating human teamwork behaviors: Using intelligent agents. Journal of Physics of Life Reviews, 1, 173–201. Festinger, L. (1957). A theory of cognitive dissonance. Evanston, IL, USA: Row, Peterson & Company. Fogg, B. J. (2003). Prominence-interpretation theory: Explaining how people assess credibility online. In Proceedings of CHI’03: Extended Abstracts on Human Factors in Computing Systems (pp. 722–723). Ft. Lauderdale, Florida, USA: ACM. Georgeff, P. M., Pell, B., Pollack, E. M., Tambe, M., & Wooldridge, M. (1998). The belief-desire-intention model of agency. In J. Muller, M. P. Singh, & A. S. Rao (Eds.), Proceedings of the 5th International Workshop on Intelligent Agents V: Agent Theories, Architectures, and Languages (pp. 1–10). Heidelberg, Germany: Springer-Verlag. Goschnick, S., & Sterling, L. (2002). Psychology-based agent architecture for whole-of-user interface to the Web. Retrieved October 10, 2008, from http://www.solidsoftware.com.au/Information/Paper/Hf2002/Poster-SteveGoschnick.pdf Guadagno, R., Blascovich, J., Bailenson, N. J., & Mccall C. (2007). Virtual humans and persuasion: The effects of agency and behavioral realism. Media Psychology, 10, 1–22. Guzzoni, D., Cheyer, A., & Baur, C. (2007). Modeling human-agent interaction with active ontologies. In N. Yorke-Smith (Chair), proceedings from the Interaction Challenges for Intelligent Assistants, Spring Symposium (Series Technical Reports, 1; pp. 52–59). Menlo Park, CA: AAAI Press. (Also available at http://infoscience.epfl.ch/record/102324) Hargie, O. (1997). Communication as skilled performance. In O. Hargie (Ed.), The handbook of communication skills (2nd ed., pp. 10–11). London: Routledge. Häubl, G., & Murray, K. (2001). Recommending or persuading? The impact of a shopping agent’s algorithm on user behavior. In M. Wellman & Y. Shoham (Eds.), Proceedings of the ACM conference on Electronic Commerce (pp. 163–170). New York: Association for Computing Machinery. Huhns, N. M., & Singh, P. M. (1998). Cognitive agents. IEEE Internet Computing, 2(6), 87–89. Ingrand, F., Chatila, R., Alami, R., & Robert, F. (1996). PRS: A high level supervision and control language for autonomous mobile robots. In Proceedings of the Thirteenth IEEE International Conference on Robotics and Automation (pp. 43–49). Minneapolis, Minnesota, USA: IEEE. Ingrand, F., Georgeff, M., & Rao, A. S. (1992). An architecture for real-time reasoning and system control. IEEE Expert, 7(6), 33–44. Katagiri, Y., Takahashi, T., & Takeuchi, Y. (2001). Social persuasion in human–agent interaction. In The Second International Joint Conference on Artificial Intelligence Workshop on Knowledge and Reasoning in Practical Dialogue Systems (pp. 64–69). Menlo Park, CA: AAAI Press. Liu, Helfenstein & Wahlstedt 142 King. W. & Ohya, J. (1996). The representation of agents: Anthropomorphism, agency, and intelligence. In R. Bilger, S. Guest, & M. J. Tauber (Eds.), Human Factors in Computing Systems: CHI’96 Electronic Conference Proceedings (pp. 289–290). Vancouver, British Columbia, Canada: ACM. Langer, E. (1978). Rethinking the role of thought in social interaction. In J. Harvey, W. Ickes, & R. Kidd (Eds.), New directions in attribution research (Vol. 2; 35–58). Potomac, MD, USA: Lawrence Erlbaum Associates. Lokuge, P., & Alahakoon, D. (2005). Reinforcement learning in neuro BDI agents for achieving agent’s intentions in vessel berthing applications. In Proceedings of the 19th International Conference on Advanced Information Networking and Applications (pp. 681–686), Los Alamitos, CA, USA: IEEE Computer Society. Luck, M., McBurney, P., Shehory, O., & Willmott, S. (2005). Agent technology: Computing as interaction; A roadmap for agent based computing. Retrieved October. 10, 2008, from http://www.agentlink.org/roadmap/al3rm.pdf Maes, P. (1994). Agents that reduce work and information overload. ACM Communications, 37(7), 30–40. McGuire, W. J. (1969). The nature of attitudes and attitude change. In G. Lindzey & E. Aronson (Eds.), The handbook of social psychology (Vol. 3; 2nd ed., 136–314). Reading, MA, USA: Addison-Wesley. McGuire, W. J. (1985). Attitudes and attitude change. In G. Lindzey & E. Aronson (Eds.), The handbook of social psychology (Vol. 2; 3rd ed., 233–346). New York: Random House. Moran, T. P. (1981). An applied psychology of the user. ACM Computing Surveys, 13, 1–11. Murray, J., Schell, D., & Willis, C. (1997). User centered design in action: Developing an intelligent agent application. In Proceedings of the 15th Annual International Conference on Computer Documentation (pp. 181–188). Salt Lake City, UT, USA: The ACM Press. Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. Oulasvirta, A., & Saariluoma, P. (2004). Long-term working memory and interrupting messages in humancomputer interaction. Behaviour and Information Technology, 23, 53–64. Nguyen, H., Masthoff, J., & Edwards, P. (2007). Persuasive effects of embodied conversational agent teams. In J. Jacko (Ed.), Proceedings of the 12th International Conference on Human-Computer Interaction (pp. 176–185). Beijing, China: Springer-Verlag. Parise, S., Kiesler, S., Sproull, L., & Waters, K. (1999). Cooperating with life-like interface agents. Computers in Human Behavior, 15, 123–142. Pasquier, P., Rahwan, I., Dignum, F., & Sonenberg, L. (2006). Argumentation and persuasion in the cognitive coherence theory. In P. Dunne & T. Bench-Capon (Eds.), Proceedings of the 1st International Conference on Computational Models of Argument (COMMA; pp. 223–234), Liverpool, UK: IOS Press. Peebles, D. J., & Cox, A. L. (2006). Modelling interactive behaviour with a rational cognitive architecture. In P. Zaphiris & S. Kurniawan (Eds.), Human computer interaction research in Web design and evaluation (pp. 290–306). London: Idea Group Inc. Petty, R. E., & Cacioppo, T. J. (1981). Attitudes and persuasion: Classic and contemporary approaches. Dubuque, IA, USA: Brown. Rahwan, I. (2005). Guest editorial: Argumentation in multi-agent systems. Autonomous Agents and Multi-Agent Systems, 11, 115–125. Rao, A. S., & Georgeff, M. P. (1992). An abstract architecture for rational agents. In C. Rich, W. Swartout, & B. Nebel (Eds.), Proceedings of Knowledge Representation and Reasoning (pp. 439–449). San Mateo, CA, USA: Morgan Kaufmann Publishers Inc. Reeves, B., & Nass, C. (1996). The media equation. Cambridge, UK: Cambridge University Press. Sewell, W. H. (1989). Some reflections on the golden age of interdisciplinary social psychology. Annual Review of Sociology, 15, 1–16. Shiffrin, R. M., & Schneider, W. (1977). Controlled and automatic human information processing: II. Perceptual learning, automatic attending and a general theory. Psychological Review, 84, 127–190. Persuasive Human-Agent Interaction 143 Stiff, J. B., & Mongeau, P. A. (2002). Persuasive communication (2nd ed.). New York: The Guilford Press. Stock O., Guerini M., & Zancanaro, M. (2006). Interface design and persuasive intelligent user interfaces. In S. Bagnara & G. C. Smith (Eds.), The foundations of interaction design (pp. 193–210). Hillsdale, NJ, USA: Lawrence Erlbaum Publishing Co. Swingle, P. G., & Gillis, J. S. (1968). Effects of emotional relationship between protagonists in the prisoner’s dilemma. Journal of Personality and Social Psychology, 21, 121–137. Takama, Y., Dohi, H., & Ishizuka, M. (1998). A visual anthropomorphic agent with learning capability of cooperative answering strategy through speech dialog. In the proceedings of the Third Asian Pacific Computer and Human Interaction conference (pp. 260–265). Washington, DC, USA: IEEE Wichman, H. (1972). Effects of isolation and communication on cooperation in a two-person game. In L. S. Wrightsman, J. O’Connor, & N. J. Baker (Eds.), Cooperation and competition: Readings on mixed-motive games (pp. 197–206). Belmont, CA, USA: Brooks/Cole Publishing Company. Wooldridge, M., & Jennings, N. R. (1995). Intelligent agents: Theory and practice. Knowledge Engineering Review, 10, 115–152. Zimbardo, P., & Leippe, M. (1991). The psychology of attitude change and social influence. New York: McGraw-Hill. Authors’ Note All correspondence should be addressed to: Shenghua Liu Department of Computer Science & Information Systems P.O. Box 35 FIN-40014 University of Jyväskylä, Finland [email protected] Human Technology: An Interdisciplinary Journal on Humans in ICT Environments ISSN 1795-6889 www.humantechnology.jyu.fi