When evaluating anothers behavior is the tendency to overestimate the influence of their personality and underestimate the influence of the situation they are in?

The correspondence bias (FAE) has been thoroughly tested with people, but not with HRI. In general, people tend to overemphasize dispositional explanations for behaviors seen in others and, at the same time, under-emphasize features of the situation (Pak et al., 2020). Because a social robot’s behavior is completely determined by its design, programming, and humans behind the scenes, it is essential to know if people will still commit the correspondence bias for robot behavior. These findings have implications for assigning credit or blame to a social robot’s behaviors. In this section, we will summarize the results, discuss implications, and offer limitations and directions for future research.

Research question 1 asked if participants would attribute the cause of an agent’s (social robot or human) behavior to disposition or to situational factors. Participants exhibited the correspondence bias (FAE) toward both human and robot agents by assuming their behavior corresponded to their underlying attitudes (a dispositional attribution) even when their behavior was clearly assigned (a situational cause). However, their dispositional correspondent inferences were stronger for the robot than for the human. In other words, judges of the robot drew a stronger unit relation between the actor and its actions, as evidenced by the larger effect size of popular or unpopular behavior on attributed attitudes for the robot. With unpopular behavior, specifically, judges held the robot more dispositionally culpable than the human. Judges also felt greater confidence in their judgments of the robot’s true attitudes compared to the human.

Research question 2 asked if the nature of the agent’s behavior as popular or unpopular would influence causal attributions and global impressions. The relatively stronger correspondence bias toward robots was driven by the greater dispositional culpability attributed to robots committing unpopular behavior, whether freely chosen or coerced. Participants generally formed more favorable impressions of human versus robot agents and popular behavior versus unpopular behavior. Humans were rated more favorably for popular behavior than for unpopular behavior, regardless of whether they chose or were assigned the behavior.

When forming impressions of robots, there were some differences. For robots committing popular behavior, the same attitudes were attributed to them whether they chose or were assigned. However, judges were more generous in their impressions of the robot when its unpopular behavior was coerced rather than chosen, a tendency not displayed when forming impressions of the human agent. Although judges formed different impressions of the robot that chose to commit popular or unpopular behavior, coerced behavior type had no marked influence on impressions. Paradoxically, people held the robot more dispositionally responsible for its forced unpopular behavior than its chosen unpopular behavior, but were also more generous in their global impressions of the robot when its unpopular behavior was forced. Although judges formed different and valence-congruent impressions of the robots that chose popular or unpopular behavior, the impressions they formed of robots coerced to commit popular or unpopular behavior did not differ.

First, there were similarities in how participants made causal attributions about robot and human behavior. They made correspondent inferences for both, attributing the cause of behavior to the agent’s disposition even when the agent was coerced to do it. This may support the CASA Paradigm (Reeves and Nass, 1996) by showing similarities in how we treat social robots and people, consistent with prior research drawing parallels in terms of robot mind ascription, intention, goals, and so on. However, the differences between how participants judged humans and robots are perhaps more interesting and important. At the broadest level, these differences in how a classic social psychology finding applied to robots versus humans adds to a small set of studies challenging the notion that people necessarily interpret and react to social robots as they do to other humans. For instance, in an HRI replication of The Milgram Shock Experiment (Bartneck et al., 2005), found that every participant was willing to administer to a robot the highest voltage shock, whereas 60% of participants in the original study refused to use the maximum setting on another human. Furthermore, there are documented differences in the expectations for interaction people hold of social robots versus humans (Spence et al., 2014; Edwards et al., 2016; Edwards et al., 2019) and in their ontological understandings of these agents (Kahn et al., 2011; Edwards, 2018). Results of this experiment are also consistent with the idea that people view robots as unique from humans on dimensions including social presence, agency, free will, status, and capacity for suffering, which may lead them to develop and apply media-centric scripts developed specifically for cognition and behavior toward social robots (Gambino et al., 2020). Although both computer-based technologies and humans may be social actors (CASA), they are not necessarily seen as the same type of social actor.

The question becomes, what is the significance of the specific differences observed in this experiment: 1) that there were stronger dispositional correspondent inferences (stronger actor/agent conflation) for robots than for humans, 2) that people were more certain about a robot’s “true disposition” than a human’s, and 3) that people uncoupled attributed attitudes from global impressions to a greater degree for robots? Satisfying answers will depend upon why people (appeared to) not only commit the fundamental attribution error with robots–which are machines logically understood to operate without interior “dispositions” like personality, attitudes, beliefs, and feeling–but also to commit it to a greater degree and with greater certainty then they did with humans. At first glance, these causal inferences of robot behavior may appear to be a mistake or error akin to the one people make in judging one another.

However, there are three problems with calling the observed results an instance of Fundamental Attribution Error (FAE). The first two arise from cross-application of criticism surrounding human FAE studies using attributed attitude paradigms: 1) the judge never really knows whether the coerced actor actually agrees or disagrees with the direction of their forced action, which means a dispositional attribution is not necessarily incorrect/erroneous and 2) correspondent inferences in which an actor is presumed to possess action-congruent attitudes do not necessarily mean that the central underlying premise of FAE—that people routinely overemphasize dispositional and underemphasize situational causes of behavior—has been supported. These critiques have resulted in a preference for the terms “correspondence bias” or “dispositional correspondent inferences” over FAE when there is no direct test of Situation Theory (S-Theory) awareness and its role in attribution processes (Gawronski, 2004). In the case of robots, there is a third and obvious reason to hesitate to apply the term “error” to a tendency to infer that a robot’s behavior corresponds to its disposition: Logically, it does not seem possible that robots, as programmed machines, hold true dispositions, beliefs, or attitudes that are incongruent with their actions. This is because beliefs and attitudes are widely understood to require inward experiential aspects or subjectivity of thought that does not characterize present robots.

Therefore, viewing the results through the lens of the correspondence bias is more fruitful because it removes both the evaluative aspect of whether people are “right” or “wrong” to conflate a robot agent and its actions and the necessity of linking observed effects to a broad and pervasive underestimation of situational influence, to center only on whether people bend toward disposition-situational convergence. Now the issue remains of how to interpret the relatively stronger and more confident correspondence bias people exhibited toward social robots. As discussed by Gawronski (2004), the correspondence bias may arise from a number of different processes involving how people apply causal theories about the role of situation on behavior (S-Theory). These include 1) lack of S-Theory (when there is no awareness of or there is disagreement with the premise that situational factors constrain behavior), 2) failed application of S-Theory (when there is knowledge of and belief in S-Theory adequacy, but people are unmotivated, lack cognitive capacity, or have inferential goals which result in failure to correct dispositional attribution bias), 3) deliberate neglect of S-Theory (when S-Theory is deemed irrelevant because observed behavior seems highly diagnostic irrespective of situational forces, as in cases of morality and performance ability), and 4) biasing application of S-Theory (when S-Theory is applied in a manner that amplifies rather than attenuates correspondent dispositional inferences) (Gawronski, 2004).

This fourth and final cause of correspondence bias—biasing application of S-Theory—seems especially relevant to understanding why people may make stronger correspondent dispositional inferences for robots than for other humans. The “over-” or biasing application of S-Theory (where “over” implies an attributional effect and not a normative or judgmental inadequacy) may occur in cases in which people understand that behavior is constrained by situational factors, are aware of present situational factors (e.g., whether the behavior was freely chosen or assigned, and the nature of the agent), have the capacity and motivation to apply S-theory, then do so to such a high degree that it appears as if they have totally ignored the causal role of situational factors (Gawronski et al., 2002). For example, people may disambiguate ambiguous human behavior by defining disposition completely in terms of the situation; Ambiguous behavior has been attributed to dispositional anxiety because the situation was perceived as anxiety-inducing (Snyder and Frankel, 1976).

Theoretically, people’s ideas about what robots are, how they work, and how they compare to humans could also lead to a biasing application of S-Theory. To the degree that robots are understood as programmed and controlled by humans, the situation may become salient to the degree it is considered completely determinative of and the same thing as disposition (they are programmed, hence their behavior literally is their personality/attitude/disposition). Ironically, this strong or complete application of S-Theory would appear in the data as heightened dispositional inference because participants would presume alignment between the robot’s behavior and its true attitudes or personality. In reality, this pattern of findings may simply reflect participants’ tendency to conflate an agent whose nature is to lack independent, interior life with its situationally determined actions.

Perhaps most significantly for theorizing HRI, this possible explanation prompts serious consideration of the idea that people may use different causal attribution processes to display a correspondence bias with robots than they do with other humans, even under the same circumstances. Both the stronger and more certain unit relation participants drew between a robot actor and its actions and the looser relationship they displayed between attributed attitudes and general impressions of the robot (i.e., the greater impression-related generosity for robots coerced to do unpopular things compared to humans) compel further investigation into whether unique perceptual patterns and theoretical mechanisms underlie causal inferences of robot behavior. Naturally, people’s causal theories about the role of situation on behavior (S-Theory) may be different for robots and human beings because people perceive them to be ontologically distinct (Kahn et al., 2011; Edwards, 2018).

The FAE, from which correspondence bias research derived, has been called the conceptual bedrock of the field of social psychology, which rests on the assumption that we tend to see others as internally motivated and responsible for their own behavior (Ross, 1977). Drawing a distinction between personality and situation is meaningful when making sense of other humans, and it appears to factor prominently in the dispositional correspondent inferences we tend to make of one another. But with robots, the similar-appearing, but stronger correspondence bias demonstrated by participants could arise from a different psychology altogether, and one more akin to the analytical/logical behaviorism which equates behavioral and mental tendency. Viewed from this lens, much of our descriptive vocabulary for human beings—mind, personality, intention, disposition, attitude, belief—may still be productively transferred to robots, but meant in a different sense [see, e.g., (De Graaf and Malle, 2019)]. Thellman et al. (2017) suggest a similar explanation of their finding that when asked explicitly, people rated goals and dispositions as a more plausible cause of behavior when the actor was human: “This raises the question whether people think of robots as less likely to have dispositions in the human sense, or as having less stable dispositions as humans, or whether people see robot dispositions as less efficacious in causing behavior than human dispositions” (Thellman et al., 2017, p. 11).

Our participants readily attributed to the robot a “true” or “real” attitude and they inferred the nature of that attitude heavily from observed behavior. However, is a robot attitude the same thing as a human attitude (see Nilsson, 2014, on robot “belief”)? Or, is the latter understood to be held (and therefore possibly concealed or subordinated), while the former is purely beheld (manifest, observed, perceived through sight or apprehension), rendering the causal distinction between an agent and its action unhelpful or illogical in the case of robots?

In other words, might people be social psychologists when it comes to other humans and behavioral psychologists when it comes to robots? For commentary on the application of behaviorist principles to robots, see: Sætra (2021); Danaher (2019).

Naturally, working out the fruitfulness of the paths of inquiry suggested above will require asking people what they think about the meaning of attitudes, beliefs, or personality (and situation) in the context of robots, and observing their language and behavior both in situ and in experiments designed specifically to test alternative explanations for a correspondence bias (or “agent-action conflation bias”) in HRI and to chart the boundaries of when, where, why, and how it may converge or diverge from human-centric causal inference processes.

In terms of methodological implications for the study of HRI, this research demonstrates the value of including within HRI experiments a human-human condition. Classically, research undertaken within the CASA framework encourages choosing a social science finding (theory and method) that applies to human interaction, replicating the research while substituting a robot/computer for a human actor in the statement of theory and design, providing the robot with human-linked characteristics, and determining whether and to what degree the same social rule still applies (Nass et al., 1994). We argue that including a human-human comparison group offers three advantages to the traditional methodology: 1) it tests again the applicability of the theory to human behavior, which is important given recent replication and reproducibility difficulties in social fields (Maxwell et al., 2015), 2) allows for the identification of both similarities and differences in HHI and HRI (including effect magnitudes) without relying on comparisons between dissimilar datasets and samples, and 3) opens examination of the possibility that even patterns of similarity in HHI and HRI may manifest for a different reason than the mindless application of human social scripts to interactions with robots (Edwards et al., 2019; Gambino et al., 2020; Fortunati and Edwards, 2021). This latter point is especially crucial because the original procedure to conduct CASA research is not sensitive to the potential operation of different theoretical mechanisms responsible for similar observational endpoints. Had we not included a human condition in this experiment, the results would have appeared only to generally mirror a tendency found in human interaction (Forgas, 1998); to suggest people also overemphasize personality at the expense of situational consideration when explaining robot behavior) and left unaddressed questions including “Are we certain the correspondence bias would be replicated with humans today, in this historical and cultural context?” “Are there any differences in how our participants would have evaluated human beings performing the same actions in the same situation?” and “Do any differences, large or small, suggest the possibility that even observed similarities warrant interpretive scrutiny?”.

The current study demonstrates that the correspondence bias extends to human-robot interactions. We do not know what factors influence the situational and dispositional attributions people make about robots. Do people over-apply situational theory to robots? In other words, how can bias attenuation occur in an interaction? Identifying future research needs to examine, through experimental design, why exactly people appear to make stronger correspondent inferences for robots than humans and how that will translate to the assignment of credit, blame, moral agency, and moral patiency. Additionally, future research needs to examine what factors may enhance or attenuate correspondent inferences.

People have an anthropocentric bias about conversations in that they expect to speak with a human and not a machine partner. In these studies, people report lower liking for social robots and have greater uncertainty about the interaction (Spence et al., 2014; Edwards et al., 2016; Edwards et al., 2019). Do these findings impact potential attributional errors with social robots? And if so, what can be done to attenuate them? Does the greater uncertainty cause the over-application? Aspects of the robot, including morphology, scripting, interaction modality, and interaction history, should be explored for potential effects on causal attributions of its behavior. Future research needs to explore why people held the robot more dispositionally responsible than the human and why they felt greater confidence in their judgments of the robot’s attitudes than the human actor.

Third, how exactly is responsibility handled differently with robots than humans? Because participants were relatively more kind in their reported impressions of the robot when its bad behavior was coerced (not so for the human agent), we need future research to examine how responsibility for decision-making will occur. Previous research has demonstrated that even when participants are given transparent details about robot behaviors and drives, they thought the robot was thinking more (Wortham et al., 2017). Although it is possible that the meaning of robot “thinking” shifted following explanations of how the robot functioned. We suspect that interpersonal relationship dimensions will come into play. If we have a relationship with a social robot, do we offer more responsibility for decision-making to the robot? We certainly do with people, and it stands to reason that relationships will make a difference in HRI. In the current study, the exposure time was the same for each condition and yet the robot was held more dispositionally responsible. Future research needs to examine if relationship differences can attenuate these differences.

Finally, it is possible that the video stimulus was not as “real-world” as a study with face-to-face embodied presence with the robot. Furthermore, the scenario was hypothetical and pertained to a single, short speech. Potentially, attribution processes play out differently following longer-term, real-world observation of robot behavior, and could differ when evaluating message behavior versus other types. Future research should replicate this study in a live interaction. Being in the room with a social robot might cause a differing correspondence bias than simply watching one on a video. Issues such as social presence (Short et al., 1976) might impact these judgments.

Page 2

PMC full text:

Copyright/LicenseRequest permission to reuse

Copyright © 2022 Edwards and Edwards.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

Means and standard deviations for attributed attitudes (1–7; opposes it:supports it).