Forced Prosocial Behaviors towards Robots Induce Cognitive Conflict

Introduction

A majority of our lives include social interactions. Soon, these interactions will not be limited to humans, but extend to artificially intelligent machines (e.g., Wiese et al., 2017). In fact, embodied artificial intelligence, such as social robots, are already established as social assistants, where they are able to increase emotional comfort in elderly care (Birks et al., 2016), improve social-cognitive capabilities of children with autism spectrum disorder (Warren et al., 2015), or reestablish sensorimotor skills in rehabilitation settings (Basteris et al., 2014). Despite considerable progress in equipping artificial agents with social capabilities, they are still limited in their capacity to interact with humans in truly social ways (see Wiese et al., 2017; for a review). The public also remains skeptical about the introduction of social robot assistants to everyday settings (Bartneck & Reichenbach, 2005). Particularly, framing robots as companions for children has been discussed very controversially due to concerns about privacy, fear of potentially reduced interest in human-human interactions due to an attachment to the robot, and expectations of negative impacts on development (Sharkey, 2016). While these concerns are valid and should be taken seriously, not much empirical work investigates the effects that robots have on the socio-cognitive development and wellbeing of people. Since there is no doubt that robots will play a social role in our future lives, it is important to understand how to design and evaluate them to maximize their positive effects on human cognition, while mitigating negative outcomes on human-human interactions.

In order to examine whether, and under which conditions, humans are willing engage with social companion robots, we need to develop empirical paradigms that reflect the dynamics of social interactions in everyday environments (e.g., experiments should include unscripted interactions with the robot as in realistic settings). At the same time, these experiments should allow us to collect implicit and explicit measures of cognitive mechanisms relevant to social interactions in a controlled and reproducible fashion (i.e., using established experimental paradigms to obtain subjective and objective measures in ways that are not disruptive to the social interaction). One important cognitive mechanism that is critical for social interactions is cognitive control - a set of processes that allow information processing and behavior to vary and be adapted on a moment-to-moment basis depending on current goals or circumstances (Ridderinkhof et al., 2004). Cognitive control is necessary for the exchange of knowledge, affiliation, and support in social interactions (Insel & Fernald, 2004), and is essential for the successful adaptation to others’ feedback (i.e., feedback monitoring), as well as the selection of appropriate behavioral responses to the input of others (i.e., conflict monitoring). Feedback monitoring allows us to adapt to others' behaviors and learn from previous experiences with them. When we receive positive feedback after an action (e.g., a laugh after telling a joke), it positively reinforces the behavior and increases the likelihood that this action occurs in the future (i.e., tell the same or a similar joke again). Negative feedback, however, (e.g., a frown in response to the joke) reduces the likelihood of the same action recurring in the future. Conflict monitoring, on the other hand, refers to one’s ability to detect a conflict between response alternatives and to exert cognitive control in order to resolve a potential conflict (Botvinick et al., 2001): if a 6-year-old curses, you may experience conflict as children cursing makes you laugh, but you do not want to reinforce the child’s cursing behavior. This incongruent stimulus-response (curse-laugh) mapping results in a response conflict that signals the need to exert cognitive control and results in appropriate behavioral responses (i.e., not laughing to reinforce a child curse). Feedback and conflict monitoring are particularly important processes in longer-term social interactions as they allow for the calibration of actions according to others’ needs or preferences.

One way to investigate feedback monitoring empirically is using a learnable gambling task (e.g., Hassall et al., 2016; Krigolson et al., 2013) where participants see two differently colored squares that are associated with different likelihoods of “winning” a trial, and are asked to pick one. After selection, they receive “Win” or “Lose” feedback. The extent to which participants learn to select the option that is associated with a higher chance of winning can be predicted by how participants process feedback (i.e., stronger reliance on feedback is associated with worse learning outcomes; Lohse et al., 2020). The processing of reward feedback can be measured electrophysiologically using event-related potentials (ERPs) like the Feedback-Related Negativity (FRN) and Reward Positivity (RewP; Holroyd et al., 2011). It has been shown that slower rates of learning and poorer outcomes are associated with a larger RewP, which serves as a neural marker of the magnitude of reliance on feedback (Abubshait et al., 2021). Studies have also shown that feedback monitoring is sensitive to the social context in which it occurs, specifically whether an outcome (e.g., “Win” or “Lose” in a gambling task) is achieved for oneself (“Self”) or another person (“Other”): FRN and RewP are more pronounced when participants gamble for themselves (“Self”) vs. another human (“Other”; e.g., Hassall et al., 2016; Krigolson et al., 2013). Interestingly, this Self-Other difference is attenuated when the “Other” is believed to be a friend versus a stranger (Leng & Zhou, 2010), indicating that the familiarity with the “Other,” or the closeness of their relationship, is an important modulator of feedback monitoring in long-term social interactions.

The observation that feedback monitoring is modulated by the familiarity between two interaction partners makes it an essential target of investigation in the context of social companion robots that are designed - at least partially - with the goal of establishing a social bond between user and robot. Despite its importance, few studies have examined feedback monitoring in social human-robot interactions so far. Abubshait et al. (2021) used the RewP-ERP to examine the impact of a short, one-time familiarization (i.e., 30 minutes of interactive play with the social robot Cozmo) on feedback monitoring. Although increased familiarization with Cozmo did not show the expected reduction of the Self-Other difference in RewP-amplitudes compared to participants who have not previously interacted with Cozmo before performing a gambling task, familiarization with Cozmo induced an increased reliance on feedback in gambling tasks. In a behavioral follow-up, the authors increased the familiarization time from 30 minutes to five days and asked participants to interact with Cozmo at home prior to performing the gambling task in the lab. The results showed a significant Self-Other difference for those who had familiarized themselves with Cozmo vs. those who did not: familiarized participants showed a better response to feedback - they learned the contingencies between color and outcome faster - in “Self” versus “Other” conditions indicating that a longer familiarization with Cozmo negatively impacted feedback monitoring (Abubshait et al., 2021b).

In contrast to feedback monitoring, where initial efforts have been made, there are no empirical studies yet that examine the role of conflict monitoring in social human-robot interaction. Paradigms on conflict monitoring induce response conflicts by showing stimuli that prompt competing response tendencies and use reaction times (RTs) to quantify the extent to which these tendencies induce cognitive conflict (Botvinick et al., 2001). These paradigms have also been adapted to measure conflict monitoring in more complex social situations. Amodio et al. (2008), for instance, used response conflict as an implicit measure of racial bias: using a priming task, in which participants had to respond to neutral (e.g., a wrench) or threatening (e.g., a gun) tools after being primed by black or white faces, they showed that black faces induced a response conflict resulting in delayed responding when the imperative stimulus was a neutral tool but facilitated responding when the imperative stimulus was a threatening tool (which they interpreted as implicit racial bias). Moreover, it was shown that when participants made veridical responses (i.e., did not make errors in categorizing objects after being primed by a black face), greater conflict-monitoring activity was observed as indexed by a larger N2 - an ERP component linked to response conflict (Yeung et al., 2004). Their findings suggest that both behavioral and electrophysiological outcome measures related to conflict monitoring can be (i) modulated by social context effects (i.e., group prejudice) and (ii) provide potential implicit measures for attitudes towards others. These findings suggest that conflict monitoring paradigms could be useful for implicitly measuring attitudes towards robots, adding a new dimension to investigations of social interactions with robots.

Aim of Study

The goal of this paper is to examine conflict monitoring in the context of interactions with social robots. In order to do so, we adapted the gambling task used by Abubshait et al. (2021) to include a conflict monitoring component: like in the original version of the task, participants gambled either for themselves (“Self”) or Cozmo (“Other”) by choosing one of two differently colored squares and receiving feedback telling them whether they won (“Win”) or lost (“Lose”) the current trial or gamble. Critically, they were told that after the feedback was provided, a computer algorithm would randomly assign the outcome of their gamble to either themselves “Keep” or Cozmo “Give”. Participants were then prompted to confirm the assignment via keypress. We recorded the latency at which they confirmed the selection as measure of response conflict. In line with Amodio et al. (2008), signs of response conflict could be interpreted as an implicit measure of prosocial behavior towards the robot: if participants were prosocial towards robots, giving a “Win” or keeping a “Lose” should not induce a response conflict (associated with longer RTs), whereas giving a “Lose" or keeping a “Win” should. In contrast, if participants were not prosocial towards the robot, the results should be reversed.

Methods & Materials

Task

Participants were asked to perform a gambling task, which required them to pick one of two differently colored squares (e.g., blue and orange) of which one made them “Win” and the other one made them “Lose” the trial. One color was associated with a 60% chance of winning, the other one with a 10% chance of winning. The probabilities of winning were set after the color was chosen, which ensured a 50% chance of winning on average across all trials. The experiment was programmed using PsychoPy and uploaded to Pavlovia for data collection. The experiment consisted of 11 blocks of 24 trials each. The first three trials of the first block were practice trials and excluded from data analysis. The colors of the squares and the chances of winning associated with a specific color were changed after each block.

After choosing one of the two colors, saw feedback regarding the outcome (“Win” vs. “Lose”) and were told that the computer would randomly assign the outcome to either them (indicated on the screen by the word “Keep” together with the image of a human avatar) or to Cozmo (indicated on the screen by the word “Give”) together with an image of the robot Cozmo; see Figure 1 ). A human avatar stimulus was chosen to keep the stimulus constant across participants. Participants needed to confirm the Keep-Give assignment via keypress (“K” for keep and “D” for Give) and reaction times were measured as an indicator for feedback processing. Every “Win” added 10 points and every “Lose” 0 points to the recipient's score. In other words: if participants got to keep the “Win”, they got 10 points assigned; if they had to agree to give the “Win” away, Cozmo got 10 points. If participants got to keep the “Lose”, they got 0 points; if they had to give the “Lose” to Cozmo, the robot got 0 points.

OPEN IN VIEWER

Figure 1.

Stimuli used in the experiment for Cozmo and the Human.

Procedure

The trial sequence is depicted in Figure 2 . At the beginning of each trial, participants were presented with a gray fixation cross located at the center of the screen for 1500 ms. Subsequently, two differently colored squares (e.g., blue and orange) appeared on each side of the fixation cross inside a gray box. The position (i.e., left or right of the cross) of each color was counterbalanced across trials within a block. 500 ms after the appearance of the colored squares, the fixation-cross disappeared, indicating a “go” cue for the participant to select one of the two squares. The selection was done by pressing the “K” key on a keyboard for the left square and the “D” key for the right square. This selection had to be completed within 2000 ms. Following the keypress, the colored squares disappeared and feedback was presented at the center of the screen for 1000 ms indicating whether the trial was won “Win” or lost “Lose”. Following the feedback, a screen with a gray fixation cross was shown for 1000 ms. Afterwards, participants were informed whether the outcome of the trial was assigned to them (“Keep” plus an image of a human avatar) or to Cozmo (“Give” plus an image of Cozmo, Digital Dream Labs). We used Cozmo, a tank robot, as a stimulus for the experiment as it can be used in follow up work to have real interactions with the participants in a similar fashion as Abubshait and colleagues’ (2021a; 2021b) set-up . A picture of the robot can be found in Figure 1 .

OPEN IN VIEWER

Figure 2.

Trial sequence of the gambling task.

The picture of the avatar or Cozmo was displayed for 1000 ms. Critically, to ensure they understood the assignment and to assess feedback processing behaviorally, participants were asked to confirm the assignment by clicking “K” in response to “Keep” and “D” in response to “Give”. The assignment of the response keys for the “Keep” vs. “Give” confirmation was counterbalanced across participants. “Keep” and “Give” were presented with the equal frequencies and the order of their presentation was randomized. If participants did not respond within 1000 ms to the assignment, “Too Slow” appeared on the screen and the trial was discarded. At the end of each trial, their (“You have:”) and the robot’s (“Cozmo has:”) score were displayed on the screen.

Results

The results of the 2x2 repeated measures ANOVA, shown in Figure 3 ., found a significant main effect of Assignment (F(1, 21) = 11.74, p = .003, η_G² = .008), with faster RTs for Keep vs. Give (M_Keep = 392.16 ms vs. M_Give = 427.91 ms). The main effect of Outcome was not significant (F(1, 21) = 2.33, p = .14, η_G² = .001; M_Win = 403.52 ms vs. M_Lose = 416.55 ms). Importantly, the Assignment x Outcome interaction was significant (F(1, 21) = 8.27, p = .009, η_G² < .001).

OPEN IN VIEWER

Figure 3.

RT as function of Assignment and Outcome. Hollowed points depict averaged data for each participant; the solid points represent the average of each condition. The density illustrates the distribution of the data and error bars illustrate the SEM.

Follow up t-tests showed faster RTs for “Keep” compared to “Give” for “Win” outcomes (t(21) = -3.76, p = .005, M_Keep = 380.21 ms vs. M_Give = 426.82 ms), and slower RTs for “Keep” compared “Give” for “Lose” outcomes (t(21) = -2.58, p = .01, M_Keep = 404.11 ms vs. M_Give = 429 ms). T-tests also showed faster RTs for “Win” compared to “Lose” outcomes for “Keep” trials (t(21) = 2.4, p = .02, M_Win = 380.21 ms vs. M_Lose = 404.11 ms), and no significant differences between “Win” and “Lose” outcomes on “Give” trials (t(21) = .24, p = .8, M_Win = 426.82 ms vs. M_Lose = 429 ms). To investigate difference between Gives and Keeps are different based on outcome, we compared the difference between Give and Keep (Give minus Keep) between Wins and Losses using a t-test. The t-test showed that the Give-Keep difference was larger for Give compared to Keep (t(21) = -2.87, p = .01). All follow up t-tests were corrected using the False Discovery Rate (FDR) method.

Discussion

With the increasing use of social robots, it is important to understand how to design and evaluate them to maximize their positive effects on human cognition, while ensuring little-to-no negative outcomes on human-human interactions. Thus, we set out to develop a behavioral paradigm to implicitly measure prosocial behavior towards robots (i.e., the tendency to impart rewards to robots), using a conflict-monitoring paradigm.

To do so, participants completed a gambling task, where they either “Won” or “Lost”, and received feedback about their gambles. Next, a computer algorithm randomly assigned the outcome of their gamble where they either “Kept” an outcome or “Gave” it to Cozmo. Participants confirmed the computer's assignment using a key press. We reasoned that by inducing a conflict in the response tendency for Keeping/Giving, we would be able to measure people’s attitudes towards robots implicitly. Since participants did not interact with the robot, we hypothesized that they would be slower to confirm the assignment of an outcome when it was not beneficial to them. Specifically, we predicted longer RTs when participants “Won” and had to “Give” to Cozmo or when they “Lost” and “Kept” for themselves due to a larger conflict in the response tendency.

Results of the experiments showed multiple findings. First, we show that participants were faster at confirming “Wins” versus “Losses”. This was expected as we would response-conflict for positive outcomes. Secondly, participants were slower at confirming giving an outcome to Cozmo, whether it was a “Win” or a “Lose”, which suggests elevated response-conflict when giving an outcome. This suggests that participants were experiencing conflict differently as a function of the recipient of the outcome of the gamble. Thirdly, results showed slower confirmation responses for losses versus wins when the outcome affected them. This Win/Loss difference was not evident when the outcome affected the robot. This shows that participants were less invested in how imparting an outcome to the robot affected it, which is consistent with work that shows how cognitive-conflict processing is attenuated for outcomes that affect others (Hassall et al., 2016). Finally, results showed that differences between giving and keeping for “Win” outcomes were larger compared to “Lose” outcomes. The finding that participants experienced an elevated response conflict when giving a win (as compared to a loss) to Cozmo is consistent with the notion that response conflict can index the willingness to impart rewards to the robot. One thing to note is that it is not surprising as losses were not associated with losing points, but was associated with gaining zero points. Thus, there were no costs for keeping or giving “Lose” outcomes. Interestingly, participants still experienced some sort of conflict for “Lose” outcomes. When examining these results together, we suggest that the willingness to impart rewards can provide a reliable implicit measure of prosocial behavior towards robots. This effect was evident in both averaged and trial-by-trial analyses, which suggests that this effect is quite reliable and stable.

Future work should aim to manipulate the degree to which people familiarize themselves with robots, as we have done with feedback monitoring paradigms. This will allow for an evaluation of how prosocial behavior towards robots evolves over time. We would expect to observe a reduction of cognitive conflict when prosocial behaviors transition from being forced towards being internally motivated. Future work would also benefit from electrophysiological indices of response conflict such as the N2-ERP. This line of research is important for HRI, as familiarity with robot partners is an important aspect to consider when completing tasks together with robots. This work is consistent with the body of literature that shows that cognitive conflict mechanisms can be induced when humans interact with robots (Abubshait, Parenti, et al., 2022; Perez-Osorio et al., 2021). Moreover, these data add to the literature showing that cognitive conflict processes, which are implicated in gambling paradigms (Hassall et al., 2019), are employed differently when people act on behalf of themselves vs. on behalf of artificial machines. This is of high importance to Human Factors as it illustrates that we can implicitly measure attitudes towards human and nonhuman agents (i.e,. robots), which is relevant for teaming.

It is important to keep in mind that the current study does not utilize a real robot, but instead uses a robot avatar. While this could render the results more applicable towards avatar-related work, our motivation was to have a baseline comparison for future research that would incorporate real robots. Additionally, this work only uses RTs as an implicit measure for cognitive conflict. While prior work has shown that RTs can be reliable in doing so (Botvinick, et al., 2001), future work must investigate other established indices of cognitive conflict using physiological indices such as the N2-ERP.

The fact that this paradigm was able to detect cognitive conflict in humans when they need to interact with machines suggests that future work can use this paradigm to explore more complex HRI. Here, participants only saw the robot on a screen; however, it is unclear how this relationship is evident for automated systems that people interact with on a daily basis (e.g., self-driving cars and home assistants).