key: cord-0592832-fwflr117 authors: Maggioni, Mario A.; Rossignoli, Domenico title: If it Looks like a Human and Speaks like a Human ... Dialogue and cooperation in human-robot interactions date: 2021-04-23 journal: nan DOI: nan sha: 48c55b1b45be463251c13f03de2aa1cc105aa825 doc_id: 592832 cord_uid: fwflr117 The paper presents the results of a behavioral experiment conducted between February 2020 and March 2021 at Universit`a Cattolica del Sacro Cuore, Milan Campus in which students were matched to either a human or a humanoid robotic partner to play an iterated Prisoner's Dilemma. The results of a Logit estimation procedure show that subjects are more likely to cooperate with human rather robotic partners; that are more likely to cooperate after receiving a dialogic verbal reaction following the realization of a sub-obtimal social outcome; that the effect of the verbal reaction is independent on the nature of the partner. Our findings provide new evidence on the effect of verbal communication in strategic frameworks. Results are robust to the exclusion of students of Economics related subjects, to the inclusion of a set of psychological and behavioral controls, to the way subjects perceive robots' behavior and to potential gender biases in human-human interactions. For decades human-robot interactions have played an important part in the production of cultural products (stories, novels, movies or, more recently, videogames), populating the imagination of people across countries and cultures. More recently, we have witnessed anthropomorphic robots becoming more widespread beyond movie sets or research labs, paving the way for them to play an increasing role in economic and social interactions. These robots move and, most importantly, talk somehow like human beings; this last feature is particularly relevant, since the ability to communicate through words constitutes a distinctive feature of human interactions. A long tradition in Philosophy, dating back to , has stressed the ontological foundations of dialogic relationships. Martin Buber (1958) claimed that full understanding of one's own identity is strictly dependent on dialogue with another presence. More recently, Dumouchel and Damiano (2017) and Damiano and Dumouchel (2018) have argued that dialogue is the fundamental structure and basic pattern of humans' actions and thinking. Paleoanthropologists such as Dunbar and Bickerton (1996), Lieberman (2006) and Tattersall (2008) as well as paleoeconomists such as Horan et al. (2005 Horan et al. ( , 2008 have formulated the hypothesis that verbal interactions and "verbal grooming" were the crucial factors in the emergence of the evolutionary advantage of Homo Sapiens over Homo Neanderthalensis, (thus explaining the survival of the former) some 150 thousand years ago; thus dialogue has been somehow embedded in the evolutionary process of human nature. 1 Further, communication has been empirically proven to increase trust and cooperation (Balliet, 2010) , which in turn are pivotal in fostering positive social interactions. In particular, face-to-face communication has been shown to both promote and sustain cooperation between subjects even in strategic settings such as social dilemmas (Ostrom and Walker, 1997; Ostrom, 2000; Bicchieri, 2002) . In recent decades, the dialogic framework has been enlarged in scope by the increase in the use of anthropomorphic robots and by the development of the sub-fields of Social Robotics (SR) and Human-Robot Interactions (HRI), which study mechanical objects that are able to communicate, both verbally and non-verbally, in similar ways to human beings and to play the role of artificial subjects, acting as "social partners" (Fong et al., 2003; Sung et al., 2010; Dumouchel and Damiano, 2017) . Recent empirical research in HRI has implemented experimental frameworks in which human subjects have been partnered with humanoid robots in social dilemmas (see, among others, Krach et al., 2008; de Melo et al., 2011; DeSteno et al., 2012; Paeng et al., 2016; de Melo and Terada, 2020) . This process has almost seamlessly led on to the question of whether communication also promotes cooperation in these new settings where people interact with robotic agents. To address this important issue in social science, in this paper, we devised a randomized experiment in which human subjects are randomly matched with either a human or an anthropomorphic robot partner (NAO, produced by Softbank Robotics) and asked to perform an iterated Prisoner's Dilemma (PD). In each of these two sub-samples, after the first round of the game, and prior to a second round being proposed, half of the subjects are randomly assigned to treatment, which consisted of the partner providing a Dialogic Verbal Reaction (henceforth, DVR) if a sub-optimal social outcome had occurred in the first round of the PD. The aim of our experiment is threefold: first, to investigate whether subjects behave differently depending on the nature of their partner (whether human or robot); second, to analyze whether a DVR, that implicitly refers to cooperation as a socially desirable strategy, influences the subsequent choice of the subject; third, and most importantly, to check whether the effect of this verbal reaction depends on the (human or robotic) nature of the partner. 2 Our main result shows that being exposed to a DVR, following sub-optimal social outcomes, positively affects the cooperation rate of subjects at the next stage and that this effect is not dependent on the partner's type. In other words, despite the fact that no DVRs affect the players' payoffs, as long as the partner subtly evokes cooperation -either in the form of an apology or a reprimand or disappointment -the subject's subsequent choice tends towards cooperation, irrespective of whether their partner is a fellow human being or an anthropomorphic robot. This paper contributes to the existing economic literature in two ways. Firstly, we provide new evidence on the effectiveness of communication in affecting decision-making, 2 In a sense, our third research question can be thought of as a sort of modified Turing Test. In the original three-person "imitation game" (Turing, 1950) an interrogator chats with two respondents, located in separate rooms, asking questions to detect which one of the two is a machine. If the interrogator cannot reliably tell the machine from the human, the machine is said to have passed the test. In our experiment the subject knows that the robot is a machine but he/she is surprised by the DVR and reacts as if the robot were a human. extending this result to HRI; secondly, and most importantly, we show that the effect of a DVR is independent of the human vs artificial nature of the agent, thus contributing to an area in the study of economic interactions between human and non-human subjects that remains under-researched. Several empirical works (e.g. Dawes et al., 1977; Braver and Wilson, 1986; Ostrom and Walker, 1991; Bicchieri, 2002) have tested the importance of communication in influencing the outcome of strategic interactions. Sally (1995) published the first meta-analysis of this stream of literature and concluded that communication exerts the strongest effect, relative to other variables known to influence cooperation, such as group size, the magnitude of the reward for choosing not to cooperate, and group identity. Subsequently, Balliet (2010) addressed the same issue through an improved meta-analysis, by adopting mediationanalysis techniques, and confirmed that communication has a strong positive effect on cooperation within social dilemmas. In a different epistemological tradition, "critical social theory", Habermas relies on the notion of "strategic competence" in his account of the development of moral consciousness, defined as "the ability to make use of interactive competence for consciously processing morally relevant conflicts of action" (Habermas, 1979, p.88) . He also refers to the ability of social actors to examine the validity and legitimacy of established norms discursively. Habermas thus "identifies a mechanism that might compellingly account for the binding force of language in strategic interaction" (Johnson, 1993, p. 81) . The nature of inter-subject interactions has lately been extended to different types of agents by the more widespread use of robots and, specifically, of humanoid social robots. The original goal of robotics was to create "artificial workers" engaged in a broad range of activities. 3 However, computer scientists and engineers soon had to acknowledge that, to be able to operate in many of these fields, robots needed to exhibit a variety of social behaviors and, in particular, to evince a believable "social presence", defined as a robot's capability to give the user the "sense of being with another" (Biocca et al., 2003) , or the "feeling of being in the company of someone" (Heerink et al., 2008) . For this reason, social robots have been developed to express and/or perceive emotions; communicate using high-level dialogue; and learn/recognize models of other agents; establish/maintain social relationships; use natural cues (gaze, gestures, etc.); exhibit distinctive personality and character; learn/develop social competencies. 4 The growth in the use of humanoid robots and the emergence of the scientific sub-fields of SR and HRI, has spurred interest in analyzing the possible consequences of repeated interactions between experimental subjects and "social robots" i.e. artificial agents that exhibit one or more of the previously defined "social" characteristics (Fong et al., 2003; Dumouchel and Damiano, 2017; Gaggioli et al., 2021) . In economic literature, the decision to trust another agent to behave as a trustworthy partner in a transaction or to cooperate in a social dilemma 5 is generally seen as not consistent with the pursuit of individual self-interest. For this reason, these decisions of trust are usually explained either in terms of repeated interactions within a finite uncertain time horizon, or by assuming: agents endowed with "self-regarding preferences", an appropriate time discount rate, a certain degree of uncertainty over the type of the opponent, a "warm glow" effect" (Andreoni, 1990) , a "gift-giving" behavior (Akerlof, 1982) ; or, finally, by referring to concepts such as equity and fairness (Fehr and Schmidt, 1999; Rabin, 1993) in case of agents endowed with "other-regarding preferences". Relatively few papers explicitly consider the real relational dimension of agents' interactions in social dilemmas. In those papers, cooperative and trustful behaviors are explained as being motivated by an acknowledgment of the other party's attitudes and intentions. Similarly, the literature on so-called psychological games addresses the role of subjects' intentions, by making payoffs belief-dependent (Geanakoplos et al., 1989; Rabin, 1993; Dufwenberg and Kirchsteiger, 2004; DeAngelo and McCannon, 2020) . Finally, support for the importance of attitudes, intentions and verbal and nonverbal cues in communication emerges from behavioral economics, as well as psychology 4 Indeed, a number of recent studies have investigated the potential effects of robots on human behavior. For example, on the one hand there is increasing concern that the complex relationship between humans and machines may have detrimental effects on the mental health of workers (Robelski and Wischniewski, 2018) and might function as an additional stressor in the workplace (Körner et al., 2019) . On the other hand, in some cases interactions with robots have been shown to be perceived as skill enhancing and capable of increasing job satisfaction. (Compagni et al., 2015) and have shown that the social perception of robots is influenced by physical appearance and behavior (Marchetti et al., 2018; Manzi et al., 2021) . 5 Commonly operationalized within a game-theoretical framework through a Prisoner's Dilemma, a Trust Game, a Centipede Game, or a "lost letter" experiment. See, among others, Dasgupta (1988) , Kreps (1990) Yezer et al. (1996) and Skeath (1999) . and neuroeconomics experiments, where players show different behaviors and neurological activations when playing incentivized tasks and games with human counterparts as opposed to automata (ranging from PCs to robots with various degrees of humanization), despite facing identical material payoffs (Kiesler et al., 1996; McCabe et al., 2001; Rilling et al., 2002 Rilling et al., , 2004 Bicchieri and Lev-On, 2007; Krach et al., 2008; Miwa and Terai, 2012; de Melo et al., 2011; Nouri and Traum, 2013; Paeng et al., 2016; Wu et al., 2016; Terada and Takeuchi, 2017; Crandall et al., 2018; Ishowo-Oloko et al., 2019) . Among studies involving humanoid robots, the use of NAO has become increasingly popular due to the robot's features and capabilities, that make it appropriate for experimental (especially clinical) researches. Robaczewski et al. (2020) documents 70 experimental studies in which subjects are involved in a HRI with NAO. In particular, 26 studies are specifically designed to study social interactions between human and robots, and only 5 of them especially focus on communication as the main moderating feature. To the best of our knowledge, only one of these studies (i.e. Sandoval et al., 2016) ask participants to play a Prisoner's Dilemma -and an Ultimatum Game -and none of them is designed to infer the effects of verbal messages evoking social norms to subjects' behavior. 6 Our research design addresses three specific questions that relates to the way humans interact with humanoid robots in a repeated Prisoner's Dilemma. RQ1: Do subjects' cooperation rates differ according to partner types? Following recent developments in empirical research in economics, psychology and social robotics we aim to understand whether subjects with a robotic as opposed to a human partner choose to cooperate with different or indistinguishable probabilities. RQ2: Does a DVR affects the subjects subsequent behavior (irrespective of partner type) in terms of cooperation rates? Empirical evidence in the literature on social dilemmas shows that communication tends to promote cooperation. We aim to 6 On the contrary, in Sandoval et al. (2016) human partners were asked to be neutral, to interact as little as possible with participants and to avoid conversations. They were even instructed to nod at participants in return for their greetings at the beginning of the experiment. Further, an experimenter (called "referee") was always present during the interactions in the lab room, thus likely introducing a strong bias in participants' behavior. investigate whether subjects are more likely to cooperate after the partner has activated a DVR, after observing a sub-optimal outcome in the previous round of the game. RQ3: Are subjects' reactions to a DVR dependent on the partner's nature (human vs robot)? This question originates directly from the former two. Since our experiment is designed as a 2×2 matrix (see Table 1 below) we may expect subjects' decisions to be affected affected by either the nature of the partner or a DVR (or both). In other words, we want to explore whether any of these effects is characterized by heterogeneity. To investigate how subjects behave when faced with a robotic partner we needed to exclude their initial (online) choice in Phase 1 to cooperate or not to cooperate to be dependent on the partner's type. At the same time, we could not entirely rely on anonymized interactions, since the purpose of the investigation is strictly related to the possible effects arising from the interactions with different types of partner. For this reason, we designed an experimental procedure with two distinct phases. In Phase 1, we asked the subjects (university students) to answer an online questionnaire and to play an incentivized task against an unknown anonymous partner. At the end of Phase 1 we asked subjects whether they wanted to come to the University Lab, proceeding to Phase 2, in order to: (i) learn the result of the interaction; (ii) be rewarded; and (iii) possibly have other interactions with the partner. The treatment consisted of a "Dialogic Verbal Reaction" that the partner delivers after observing a sub-optimal aggregate outcome of the interaction. Different stimuli were administered depending on the observed outcomes in the Prisoner's Dilemma, 7 as summarized in Table C2 in Appendix C. The layout of our experiment is summarized by the flow-chart in Figure 1 . We sent invitation emails to all freshmen and sophomore students of Università Cattolica del Sacro Cuore, who were attending lectures at the Milan campus. 8 The email included a link to an online survey containing the first phase of the experiment. In addition to a set of psychological, attitudinal and socio-demographic questions, the survey included the first round of a repeated Prisoner's Dilemma (as in Kreps et al., 1982) . At the beginning of each round, each player is given 3 euros -as initial endowment -and can choose between two actions: if he/she chooses to "Cooperate", their sum is transferred to their Partner who receives double the original sum. If he/she chooses "Not Cooperate", he/she keeps their original sum. Thus, the game proposed to subjects in this experiment consists of N repetitions of a two-person, two-strategy, game (as summarized in Table 2 ). 7 No DVR is activated when the aggregate Pareto optimal outcome is obtained, i.e when both subject and partner cooperate. 8 Università Cattolica del Sacro Cuore is a multi-campus university mainly based in Milan but with other locations in Rome, Brescia, Piacenza and Cremona. Students who agreed to take part in the incentivized game were informed that they had been randomly paired with an unknown anonymous (human or artificial) partner who was playing the same game. 10 If they agreed to participate, they were asked to make their choice (either Cooperate, C, or Not cooperate, N). The outcome of the experimental session (and the monetary reward, including show-up fee) would only be revealed on participants' agreeing to attend and take part (in person) in the second phase of the experiment at the university Lab. Only at that point would respondents discover whether the partner they had faced in the online interaction was a human or a robot, the outcome of the interaction would be revealed and they might receive an invitation to play one (or more) rounds of the game again. Students who agreed to take part in Phase 2 had to schedule an appointment through an online third-party application. Phase 2, the proper Lab Experiment, consisted of three sequential stages. In Stage 1, the subject was taken to the lab room by the experimenter, where he/she could meet and face his/her partner, whether Human or Robot, 11 with whom he/she was told he/she had played the PD online in Phase 1. When the subject 9 It seems reasonable to assume that IDR is not relevant for our experimental setting given that: (i) when subjects play their first round they are unaware of the delay between the first and second rounds; (ii) this delay is short (on average about 7 days); (iii) all subsequent rounds are played immediately one after the other in the experiment room. However, to address potential biases driven by this issue, in Section 6 we also control for a measure of intertemporal discount rate. 10 The partners' cooperation rate was set at 50%, to be consistent with the average cooperation rate observed in 8 experimental studies, collected in 5 papers, involving Italian subjects, namely Pepitone et al. (1967 Pepitone et al. ( , 1970 ; Gallucci and Perugini (2000) ; Ciardo et al. (2015) ; Meier et al. (2016) . The weighted average of the cooperation rate in these studies is 49.23%. Data retrieved from https://app.cooperationdatabank.org/. 11 The Robot was NAO, a humanoid robot produced by Softbank Robotics (see Figure D4 in Appendix D for a picture, Gelin (2018) and Robaczewski et al. (2020) for references); human partners were Ph.D. students recruited and trained for this task. Having given the subject instructions about the experimental procedures, the experimenter would tell the players to wait and listen to the game director's instructions which were available both as a recorded voice message as well as being displayed on the screen in front of each player; then the experimenter left the room. The subject would sit at a table, then the partner (either a Ph.D. student or a humanoid robot) would greet the subject and verbally introduce him/her/itself. 12 Finally, the screen would reveal the outcome of the Prisoner's Dilemma played online in Phase 1. Based on the game outcome as shown in Table C2 in Appendix C, a random algorithm determined whether to activate the DVR with a 50% probability, triggering the partner, whether human or robot, to deliver the appropriate verbal stimulus. The subject was then asked whether he/she would like to continue to play another 12 In all sessions, the Ph.D. students introduced themselves by their first name, greeted the subjects and asked the subject's first name; the robot, introduced itself stating the following in all experimental sessions: "Hi, I am ToM, a humanoid robot developed by Softbank Robotics able to perform complex interactions with human beings. What's your name?" round of the game, according to a random algorithm with a probability of 95%. If the subject accepted, a second round was then implemented, without any further assignment of treatment.We checked that the probability of both proposal and agreement to proceed to Stage 2 was not conditional on the outcome observed in Stage 1. 13 After the results of the second round were revealed, the subject could ask to continue with a further round of the game, according to a random algorithm with a 15% probability. If the subject agreed to play again, a third round of a PD was implemented, again without any treatment assigned. 14 After the last incentivized interaction, the game director communicated the total amount gained and asked the subject whether he/she would like to donate 1 euro to a charity of his/her choice. 15 Once this last choice was made, the subject had his/her receipt automatically printed, he/she left the lab room to enter the check-out area where he/she was paid the amount in cash. On leaving the Lab the student signed a confidentiality agreement requiring him/her not to share any information about the experiment with fellow students. We planned to run the experiment from 13 th February, to 13 th March, 2020. However, due to the outbreak of the COVID-19 pandemic, we had to stop all activities in the lab on 21 th February, which proved to be our last day of data collection in 2020. In February 2021, universities in Italy opened again for in-person lectures. We seized the moment 16 and intended to run another wave of the entire experiment (both Phases 1 and 2) between 22 nd February and 31 th March. Again, due to the resurgence of the pandemic, we had to stop all activities in the lab on 4 th March. At the end of the experimental session we fully disclosed all experimental procedures to all participants. We sent a total of 23,552 emails, and received 2,205 individual answers and 1,893 valid and completed questionnaires (Phase 1). 1,829 respondents (96.6% of the valid entries) agreed to participate in the Second Phase of the experiment. Out of a total of 490 subjects taking part in Phase 2 (see table C4 . 17 An initial inspection shows that subjects taking part in the research were overall more likely to cooperate than not to cooperate. 18 Table 3 Interestingly, also the baseline measure of our outcome variable (i.e. the choice to Cooperate in Stage 1) is not statistically different across treatment groups. Conversely, a t-test shows that in both samples subjects assigned to the treatment group are more likely to cooperate than those in the control group, suggesting a potential effect of the DVR on subjects' choices in Stage 2. Finally, the bottom panel of Table 3 provides a summary of the differences in DVR 17 As illustrated in Table C2 in Appendix C a DVR stimulus could only be applied in 3 out of 4 possible outcomes. 18 Please note that most of the cooperative choices, i.e. those yielding an outcome of "CC" in Stage 1, were not included in Stage 2 in the Experimental sample: therefore, by design, the Experimental sample is "less cooperative" than the Full sample. 19 In this case, Freshman is statistically significant, and greater in the case of Robot rather than Human partner. In Figure B3 we also provide a balance test for the two experimental waves, 2020 and 2021, showing that only Freshmen are more represented in 2021. effects by subsamples, i.e. it shows whether the increase in cooperation in the treatment group is heterogeneous across Human/Robot condition. Interestingly, being assigned to the Human condition is likely to increase the probability of cooperation in Stage 2 in the control group, where no DVR is performed; however, this statistical difference disappears when the comparison is made in the treatment group, suggesting that the effect of DVR is not heterogeneous across partner types. Notes: Summary statistics refer to subjects eligible for treatment, i.e. excluding those yielding "CC" as the outcome of Stage 1. We addressed the aforementioned set of research questions by implementing a Logit model, in which the probability that respondent i makes a Cooperative choice is conditional on a set of control variables and experimental conditions. Formally: where π i is the probability that Y i equals 1; Y is the dependent variable measured at rounds 1 and 2; β 0 , β 1 , β j , · · · , β k and δ are the parameters to be estimated, with δ 1 , δ 2 and δ 3 being respectively the main coefficients of interest for RQ1, RQ2 and RQ3; Choice i , r−1 is the choice made by respondent i in the previous experimental round, r−1; and X j · · · X k are a set of k experiment-related and control variables, illustrated below. To increase the precision of the estimates and to account for potential confounding factors, all models are estimated using different specifications. Control variables include experiment-related controls and background characteristics of the subjects. In the former set of controls we include Instruction order, a categorical variable that takes into account which of the potential outcomes of the Prisoner's Dilemma is shown first to the subject during the introductory instructions for the game, to control for potential priming; 20 Wave, to control for the year of the experiment (either 2020 or 2021); and Experiment day which relates to the number of days the experiment has been running (e.g. Day 1, 2, 3, ...) to control for potential spillover effects. In the latter group of controls we include Female, to control for subjects' gender; Freshman, to control for subjects undertaking the first year of their BA; Econ, to control for subjects enrolled in BA degrees in Economics, Management, Finance and Banking; and Fail test, to control for subjects failing the pregame test designed to verify their full understanding of the game's procedures, rules and payoffs. Moreover, all models are also estimated by excluding this latter group of subjects. The description of all variables is summarized in Table C3 in Appendix C. Finally, since our experiment design consists of 2 × 2 treatment conditions, in Section 6 we test the first two hypotheses by estimating separately the effects of Robot and DV R using sample splits. On average, all subjects eligible for treatment who proceeded to Stage 2 chose to Cooperate 58% of the times, as shown in the top panel in Table C4 . The breakdown by treatment and experimental condition is illustrated in Figure 3 and summarized in Table C1 in Appendix C. The estimated coefficient of Robot in Table 4 provides evidence that subjects display different behavior depending on the type of partner: being assigned to a Robot significantly decreases the cooperation rate, with an estimated probability between 16.5% and 18%, depending on specifications, as shown in the "Margins" columns. This result holds once control variables are included, as shown in Table C6 in Appendix C, and the magnitude of the coefficient is substantially the same, implying a lower probability of cooperating, at between 15.6% and 18.3%. Overall, these results show that, ceteris paribus, cooperation is strongly and significantly affected by partner type. This result is aligned with the evidence provided by a systematic review of experiments involving computer players performed by March (2019) who finds that players' behavior differs across human vs.computer opponents; and that subjects generically behave more selfish and more rational when interacting with computers. The coefficient of DVR provides evidence to support the assertion that a DVR positively affects cooperation rates: in all models, subjects cooperate more when a DVR is introduced, with a probability of between 20.4% and 22.3%. In this case also, the results hold true once control variables are included, as shown in Table C6 in Appendix C. All types of DVR exert a positive influence on the cooperation rate, as summarized by Figure D3 . However, the lack of statistical power, due to the limited sub-sample size, hinders a disaggregated analysis. In aggregate, this result provides new evidence of the positive effect of communication in fostering cooperation in strategic decision-making, strengthening and confirming previous similar findings (such as Sally, 1995; Kollock, 1998) . In the previous section we showed that the effect of a DVR is strong and significant under both (Human and Robot) experimental conditions. In this section we test whether these effects are heterogeneous across partner types, i.e. whether being faced with a Human or Robot partner affects the way verbal interactions promote cooperative choices. In Table 4 , the coefficient of the interaction variable DV R × Robot provides the outcome of this test. 21 As the table shows, the coefficient is never significantly different from 0. Note that the coefficients of both Robot and DV R are significantly different from 0, as shown in the previous sections. Therefore, while the effect of the DVR is large and significant, this effect is not heterogeneous across partner types. In other words, once the respondent is randomly assigned to the DVR treatment group, he/she is more likely to cooperate with his/her partner, irrespective of whether the partner is a Human or a Robot. The lack of heterogeneity provides a striking result: as long as the partner in the PD provides a DVR, subjects respond by increasing, on average, their probability of cooperating. Therefore, the ability to provide a verbal dialogic interaction -which indirectly evokes a commonly shared ethical norm in terms of the social desirability of cooperation -eliminates any previous differential cooperative attitudes with respect to human versus robots. These results are robust to the inclusion of the usual set of control variables, (as shown in Table C6 , in Appendix C). The results shown in Table 4 may depend on the belief that a robot which is able to implement a sensible contextual dialogical interaction 22 is "more human", thus it elicits a behavior that a subject normally reserves for fellow human partners. Alternatively, the DVR may produce the effect by simply acting as a mere soft reminder of the social desirability of cooperation. 23 Our experimental framework is unable to disentangle between these two competing and/or complementary explanations. However, it is worth noting that NAO, the robot we used in the experiment, despite being able to display appropriate gaze and body gesture cues to increase its appearance of "socialness", is still far away to reproduce the pitch, accent and expressiveness of a natural human voice. 24 Further research is needed to disentangle these two possible explanation. One possible strategy would be to have the DVR provided by the Game Director (the pre-recorded voice over) commenting on the realization of socially sub-optimal results. In this case, the "ethical reminder" would be separated from the partner (and, in particular, from the robot). Nonetheless, our results seem consistent with some recent empirical findings support the importance of communication in HRI using the same social robots employed by our experiment (NAO). Pelikan and Broth (2016) find that subjects tend to use the same signals as in human-human interactions, such as adjusting word selection, turn length and prosody, thus adapting to the perceived limited capacity of the robot. Tahir et al. 22 Note that we can exclude that the effect is due to the mere fact of NAO being able to talk since it greets any subject it is interacting with, whether or not, in a later moment, it will perform a DVR. 23 Similarly to the role played by the mentioning of the Ten Commandments as in (Mazar et al., 2008) , or the honor code as in (McCabe and Trevino, 1993) in stimulating academic honesty. 24 NAO communicated with the subject via a "Wizard of Oz system" controlled by the laptop PC located in the control room. All moves and speech items were written and coded in the system. The robot followed a pre-programmed protocol where the experimenter did not need to speak or type anything during the interaction but only press a button to start the interaction as in Laban et al. (2021) . (2018) asses two modalities to deliver the feedback: audio only and audio combined with gestures and show that when audio and gesture are combined subjects better understand the feedback delivered by the robot. In this section we provide robustness checks for the main result shown in this paper: while being matched with a robot partner decreases, ceteris paribus, cooperation rates and receiving a DVR from a partner (of any type) increases them, the effect of a DVR is not heterogeneous across partner types, i.e. it does not depend on the partner being a robot or a human. We address potential sources of bias in our results that may be driven by the presence of Econ students (as defined below); the subject's psychological and behavioral traits; the choice made by the partner and observed by the subject in stage 1; the subject's perception of robots' behavior; and the occurrence of gender biases. We address each one of these concerns in separate subsections and finally also check for possible differential outcomes when using sample splits. 25 To assess the robustness of our results, we carry out our econometric analysis on different sub-samples. First of all, we analyze the potential bias induced by students studying for BAs in Economics, Management, Finance and Banking (henceforth: Econ students). Empirical evidence shows that students in this population are more self-interested than their non-Econ peers either because of a self-selection process (Frey and Meier, 2005; Carter and Irons, 1991) or because of an indoctrination effect (Frank et al., 1993; Haucap and Müller, 2014) . 26 Analysis of the causes of this differential behaviors lies beyond the scope of this project; nevertheless, we tested whether our results are driven by the inclusion of a substantial number of Econ students, who accounts for 36% of the sample. 27 Table 5 displays the results of the estimation of Eq. (2), as does Section 5.4 where Econ students are excluded from the analysis. The main outcome of the analysis is confirmed, since the coefficient of the interacted variables is still not significant, and both the size and significance of the DVR and Robot coefficient is consistent with the results shown in Section 5.4. To complete our analysis, in Table C7 in Appendix C we show results of the same models when control variables are included. Also in this case, the main outcome of our analysis still hold. * p < 0.10, * * p < 0.05, * * * p < 0.01 As a further robustness check we assess whether our results are affected by underlying unobserved psychological and behavioral traits of the subjects. The choice to cooperate entails the risk of being taken advantage of and may be driven by the level of general trust towards other people that is "embedded" in each individual subject. Furthermore, greater or lesser degree of impulsiveness may drive the cooperation/non cooperation choice. Finally, cooperation may be related to subjects' personal levels of pure altruism. While the empirical literature has not reached a consensus on the relation between risk aversion and cooperation (Dolbear Jr and Lave, 1966; Sabater-Grande and Georgantzis, 2002; Burks et al., 2009; Proto et al., 2019) , trust is generally associated with a positive influence on cooperation (Tedeschi et al., 1969; Cook and Cooper, 2003; Jung et al., 2012) , while IDR and impulsiveness have been shown to be negatively correlated to cooperation (Yi et al., 2005; Streich and Levy, 2007; Jones and Rachlin, 2009; Locey et al., 2013; Malesza, 2020) For this reason, Table 6 shows the outcome of the same models estimated in As Table 6 shows, only the coefficient of the Trust scale is weakly significant and consistent with the current empirical evidence. All the other psychological and behavioral controls are not statistically significant and the estimated coefficients relating to the main effects are unchanged with respect to the main model. We also assess the robustness of our results to the inclusion of partner's choice 28 in the online phase of the experiment. Table 7 summarizes the outcome of the analysis, showing that partners' cooperative behavior significantly increases the subjects' probability of cooperation at stage 2 only when previous subjects' choices are not included in the model (see columns 1 to 3 in Table 7 ). Therefore, the partner's behavior is not affecting per se the subsequent decision to cooperate, while partner type and occurrence of DVR are. These results are unchanged when control variables are included, as shown in Table C8 , Facing a robotic partner in a PD is currently an uncommon situation for human subjects. Since we do not know a priori how our subjects perceive a robot partner (especially in terms of preferences and/or beliefs), we decided to include a simple questionnaire during Phase 1 in which we asked subjects to identify their perception of robots. In this way we obtained information about whether subjects perceived a robot as an "Adaptive" device, or as a device simply controlled by a predetermined set of program lines. 29 * p < 0.10, * * p < 0.05, * * * p < 0.01 Table 8 provides a test of the potential heterogeneity driven by the perception of robots as an adaptive tool. Interestingly, also in this case, the main outcome of our analysis is unchanged, with both the DVR and Robot variable remaining statistically significant, while all the interaction terms, capturing potential heterogeneity, are not significantly different from 0. As a further check, we inspect whether the effect of a DVR is still significant in the Robot subsample, as shown in Table 9 . The outcome shown in the Table strengthens our main result, confirming that the DVR is the main driver of cooperative behavior in our experiment. Since we enrolled both male and female PhD students for the role of the Human partner, a possible concern may arise relating to subjects' biased responses depending on the partner's gender. For this reason, we perform a robustness check by including an interaction term between the partner's gender and the DVR, whose outcomes are presented in Table 10 . Obviously, this analysis is only performed in the Human sub-sample. As the results in the As the Table shows , the coefficient of the DVR is strongly positive and significant in both the Human and Robot experimental conditions. The upper panel in Table 12 shows that when the robot partner delivers a DVR, subjects are on average 27% more likely to choose Cooperate in the PD than in the control group. A similar result can be found in the Human sub-sample, in which the probability of cooperation is between 21% and 25% higher when the human partner delivers a DVR than in the case he/she is not. 31 It is worth noting that the effect is particularly strong when controlling for the outcome of Stage 1, i.e. the choice of respondents in the previous "blind" stage of the experiment. In other words, the result holds when taking account of the subject's cooperative attitude, which, as expected, is the most important predictor of his/her choices. 31 We also checked that this result is not affected by some sort of gender bias and found that no differential effect can be found depending on both partner's gender, subject's gender and the interaction between the two. A summary of this robustness check is reported in the Appendix, in Table 10 , showing substantially no significant coefficients in the interaction terms. Notes. Logit model, dependent variable: choice of Cooperation at Stage 2, Prisoner's Dilemma. Subjects yielding the outcome "CC" in Stage 1 are excluded from this sample. Standard errors are in parentheses. † This sub-sample excludes all subjects who failed the test assessing their full comprehension of the instructions. * p < 0.10, * * p < 0.05, * * * p < 0.01 Notes. Logit model, dependent variable: choice of Cooperation at Stage 2, Prisoner's Dilemma. Subjects yielding the outcome "CC" in Stage 1 are excluded from this sample. Standard errors are in parentheses. † This sub-sample excludes all subjects who failed the test assessing their full comprehension of the instructions. * p < 0.10, * * p < 0.05, * * * p < 0.01 In a seminal paper, Farrel and Rabin state: "People in reality do not seem to [...] question each other's statements as much, as game theory suggests they should" (Farrell and Rabin, 1996, p. 108) . Thus, people tend to respond even to cheap talks. Does what is true for Human-Human interactions, hold for Human-Robot interactions? In this paper, we devised a randomized experiment in which human subjects are randomly matched to either a human or an anthropomorphic robot partner and asked to perform a repeated Prisoner's Dilemma to investigate (i) whether subjects behave differently depending on the nature of their partner (human or robot); (ii) whether a Dialogic Verbal Reaction (DVR), which implicitly refers to cooperation as a socially desirable strategy, influenced the subject's subsequent choice; (iii) whether the effect caused by the DVR depended on the (human or robotic) nature of the partner. Our results suggest that subjects tend to act more cooperatively with fellow human beings, rather than with robots; are influenced by a DVR towards a more cooperative strategy; and finally, that the effect of a DVR is strong enough to make a difference in behavior, based on the partner's type, insignificant. We are aware of the possible limitation of extending lab experiment results to the "real world" -as evidenced by Rabin (1993) ; List (2007, 2008) ; however we are convinced that our results may suggest interesting implications in a number of cases where robots are used to interact with -often fragile -human beings (Nursing Homes, Care Facilities, Hospitals, etc.). The main result on the DVR may depend on the belief that a robot which is able to implement a dialogical interaction with the subject appears to be "more human", thus deserving a behavior similar to that reserved for fellow human partners. Further research is still needed to disentangle the effects of verbal interactions in HRI versus the reminder effect performed by the message they deliver. Apology "I realize I made a mistake in our online interaction. I meant to press C to cooperate. However I pressed N by mistake. I am really sorry! I will be more careful next time" (N, C) Reprimand "I am really upset. If you had chosen to cooperate we would have gained 6 euro each, a reasonable amount! On the contrary you exploited my goodwill and I got nothing" (N, N) Disappointment "What a pity! If we had chosen to cooperate we would have gained 6 euros each. Why not cooperate in the next round?" (C, C) None Other controls Adaptive Binary variable based on the answers given in the online questionnaire to the question: "When interacting with a human being how does a robot behave?". The variable is equal to 1 if the subject chose the following answer: "A robot adapts its behavior to the interaction with the human being'; the variable is equal to 0 if the subject chose one of the following answers: "A robot privileges its own interest over the interest of the human being", "A robot matches the interest of the human being to its own interest", "A robot exhibits a random behavior", "Robots execute a fixed list of commands and operations" Percentage of Stage 1's outcomes given acceptance to continue to Stage 2 Figure D2 : Percentage of respondents accepting to proceed to Stage 2 by observed outcome in Stage 1. Figure D3 : Choice of Cooperation in Stage 2 by experimental condition and type of DVR Figure D4 : NAO robot, produced by Softbank Robotics Figure D5 : Experiment room Interaction terms in logit and probit models Labor contracts as partial gift exchange Impure altruism and donations to public goods: A theory of warmglow giving Communication and cooperation in social dilemmas: A meta-analytic review Covenants without swords: Group identity, norms, and communication in social dilemmas Computer-mediated communication and cooperation in social dilemmas: an experimental analysis. politics Toward a more robust theory and measure of social presence: Review and suggested criteria Choices in social dilemmas: Effects of communication within subgroups Cognitive skills affect economic preferences, strategic behavior, and job attachment Are economists different, and if so, why? Eyes keep watch over you! competition enhances joint attention in females How early implementations influence later adoptions of innovation: Social positioning and skill reproduction in the diffusion of robotic surgery Neanderthals and homo sapiens had similar auditory and speech capacities Experimental studies of cooperation, trust, and social exchange Cooperating with machines Anthropomorphism in human-robot co-evolution. Frontiers in psychology 9 Trust as a commodity Behavior, communication, and assumptions about other people's behavior in a commons dilemma situation The effect of expression of anger and happiness in computer agents on negotiations with humans The interplay of emotion expressions and strategy in promoting cooperation in the iterated prisoner's dilemma Psychological game theory in public choice Detecting the trustworthiness of novel partners in economic exchange Individual risk attitudes: Measurement, determinants, and behavioral consequences Risk orientation as a predictor in the prisoner's dilemma A theory of sequential reciprocity Living with robots Grooming, gossip and the evolution of language Cheap talk A theory of fairness, competition, and cooperation A survey of socially interactive robots Does studying economics inhibit cooperation Selfish and indoctrinated economists? Relationship development with humanoid social robots: Applying interpersonal theories to human/robot interaction An experimental test of a game-theoretical model of reciprocity Psychological games and sequential rationality NAO Communication and the Evolution of Society Why are economists so different? nature, nurture and gender effects in a simple trust game The influence of social presence on acceptance of a companion robot by older people How trade saved humanity from biological exclusion: an economic theory of neanderthal extinction Coevolution of human speech and trade Altruism versus egoism in human behavior of mixed motives: An experimental study Election Study Behavioural evidence for a transparency-efficiency tradeoff in human-machine cooperation Is talk really cheap? prompting conversation between critical theory and rational choice Delay, probability, and social discounting in a public goods game Trust and equity theory in prisoner's dilemma A prisoner's dilemma experiment on cooperation with people and human-like computers Social dilemmas: The anatomy of cooperation Perceived stress in human-machine interaction in modern manufacturing environments-results of a qualitative interview study Can machines think? interaction and perspective taking with robots investigated via fmri A Course on Microeconomic Theory Rational cooperation in the finitely repeated prisoners' dilemma Tell me more! assessing interactions with social robots from speech. Paladyn What do laboratory experiments measuring social preferences reveal about the real world Homo economicus evolves Toward an evolutionary biology of language Social discounting and the prisoner's dilemma game The effects of the dark triad traits in prisoner's dilemma game Can you activate me? from robots to human brain The behavioral economics of artificial intelligence: Lessons from experiments with computer players. Number 154 Theory of mind and humanoid robots from a lifespan perspective The dishonesty of honest people: A theory of self-concept maintenance Academic dishonesty: Honor codes and other contextual influences A functional imaging study of cooperation in two-person reciprocal exchange Trust and in-group favoritism in a culture of crime Impact of two types of partner, perceived or actual, in human-human and human-agent interaction A cross-cultural study of playing simple economic games online with humans and virtual humans Collective action and the evolution of social norms Communication in a commons: cooperation without external enforcement. Laboratory research in political economy Neither markets nor states: Linking transformation process in collective action arenas Human-robot trust and cooperation through a game theoretic framework Why that nao? how humans adapt to a conventional humanoid robot in taking turns-at-talk The role of self-esteem in competitive choice behavior Justice in choice behavior: A cross-cultural analysis Intelligence, personality, and gains from cooperation in repeated interactions Incorporating fairness into game theory and economics. The American economic review A neural basis for social cooperation The neural correlates of theory of mind within interpersonal interactions Socially assistive robots: The specific case of the nao Human-machine interaction and health at work: a scoping review Accounting for risk aversion in repeated prisoners' dilemma games: An experimental test Conversation and cooperation in social dilemmas: A meta-analysis of experiments from 1958 to 1992 Reciprocity in human-robot interaction: a quantitative approach through the prisoner's dilemma and the ultimatum game Games of Strategy Time horizons, discounting, and intertemporal choice Domestic robot ecology. International A user study of a humanoid robot as a social mediator for two-person conversations An evolutionary framework for the acquisition of symbolic cognition by homo sapiens Trust and the prisoner's dilemma game Emotional expression in simple line drawings of a robot's face leads to higher offers in the ultimatum game Some empirical evidence on dynamic inconsistency Computing machinery and intelligence Trust and cooperation in human-robot decision making Does studying economics discourage cooperation? watch what we do, not what we say or how we play Relationship between cooperation in an iterated prisoner's dilemma game and the discounting of hypothetical outcomes This section presents a replication of the main results shown in Section 5 by adopting a Linear Probability Model, estimated through Ordinary Least Squares (OLS), in place of Logit models adopted in the main texts.