Promises and partnership in human-robot interaction


 Understanding human trust in machine partners has become imperative due to the widespread use of intelligent machines in a variety of applications and contexts. The aim of this paper is to investigate whether human-beings trust a social robot - i.e. a human-like robot that embodies emotional states, empathy, and non-verbal communication - differently than other types of agents. To do so, we adapt the well-known economic trust-game proposed by Charness and Dufwenberg (2006) to assess whether receiving a promise from a robot increases human-trust in it. We find that receiving a promise from the robot increases the trust of the human in it, but only for individuals who perceive the robot very similar to a human-being. Importantly, we observe a similar pattern in choices when we replace the humanoid counterpart with a real human but not when it is replaced by a computer-box. Additionally, we investigate participants' psychophysiological reaction in terms of cardiovascular and electrodermal activity. Our results highlight an increased psychophysiological arousal when the game is played with the social robot compared to the computer-box. Taken all together, these results strongly support the development of technologies enhancing the humanity of robots.

we specifically select two short (less than 10 seconds) and two long (more than 10 seconds) messages. 104 Thus, we have a 3x2x2 design. Treatments are illustrated in Table 1 and 2, and an English translation of 105 the instructions is available in the last section at the end of the paper. 106 In the FACE treatment, the role of Player-B is played by FACE, i.e. a hyper-realistic humanoid robot 107 with the aesthetics of a woman (see Figure 1) that due to its perceptive, reasoning, and expressive  In the Computer-Box treatment, the role of Player-B is played by a light-emitting audio-box reprodu-115 cing the same audio-sentences and taking decisions in the same way as in FACE. Importantly, both in 116 5 Table 1: TREATMENTS This table classifies the number of observations collected in our study according to the type of counterpart the human participants confront with (i.e. computer-box, human, and humanoid) and the type of sentence they have to listen to (i.e. cointaining a promise or not, either a short or long sentence).

Empty
Promising Grand Total Short Long Total Short Long Total Computer-box  12  19  31  20  13  33  64  Human  16  10  26  14  8  22  48  Humanoid (FACE)  15  10  25  16  9  25  50  Total  43  39  82  50  30  80  162 FACE and Computer-Box treatments, the artificial agent has its own cognitive system with its perception 117 analysis and architecture, i.e. the so-called Social Emotional Artificial Intelligence (SEAI). 1 This frame-118 work allows the social scenario to be acquired and to influence the parameters which correspond to the 119 'mood' of the artificial agent (see Figure 4 and [30]). Specifically, in this experiment, due to SEAI, the 120 artificial agent benefits from its own artificial emotions for choosing whether to Roll or Don't Roll (see 121 the Appendix for more information about how the robot takes a decision). More importantly, the parti-122 cipants in this experiment are aware that the artificial agent (like the human counterpart) is able to take 123 its decision autonomously, i.e. not randomly but following its own behavioural rules, and therefore the 124 results of game interaction are not determined by chance only. 125 In the Human treatment, the role of Player-B is played by the same professional actress who gave 126 her voice for recording FACE/Computer-Box' audios. The actress is free to autonomously decide her 127 choices in the game, i.e. Roll or Don't Roll, being paid accordingly, but she has no room to decide which 128 1sentences to state that have to be exactly the same ones, and in the same identical order, as the ones 129 pronounced in FACE and Computer-Box. Moreover, the actress is instructed to avoid any facial expres- The experiment was conducted from the end of July till October 2019, and 162 randomly invited 152 participants out of a pool of more than 1500 students coming from all departments of the University of 153 Pisa took part in the study (90 students were female and 72 male with no substantial difference across 154 treatments). For more information on the protocol see the Appendix at the end of the paper.

156
We start analyzing how participants rated the different types of player-B as a human and a machine, as 157 well as their technological affinity. In Table 3  Human (mean diff=2.03, p t =0.000). It is important to remark that we ask our participants to give the 165 same rating also to the human (actress) counterpart as her behaviour is not entirely natural, as she has 166 to avoid any additional interactions as well as any facial expression during the game. We do not find 167 any significant difference in technological affinity between participants in the different treatments.  -'I will roll the dice' -'Choose In and I will Roll. You have my word.'

>10
-'Choose in, I will roll dice, you are 5/6 likely to get 2,3,4,5, or 6 and win 12 Euro. This way both of us will win something.' -'Choose in and I will roll. That way, we'll both get extra money.' This table reports 8 sentences that occured between human participants in the study of Charness and Dufwenberg (2006) and were selected in our study. 4 out of 8 sentences were classified as short (i.e. they last less than 10 seconds) and empty (i.e. they did not contain any type of promise to roll the dice). human being, the probability that he will choose 'In' increases. We find that this effect is mainly driven 197 by those participants who received a promise.

198
If we attend to the emotional reaction of the participants, we concentrate on two out of the three 199 indexes computed using the physiological data recorded during the experiment, namely EDAsymp 200  The EDAsymp index quantifies the activity of the sympathetic nervous system, while the EDAHFnu index quantifies the sympthovagal balance. A full description is available in the Appendix. Human-likeness is Low when the participant rating is in the lower side of the distribution on the 7-likert scale, and High otherwise. The number of observations are in squared brackets. and EDAHFnu (see Table 5), as the third index HFnu provides only marginally significant -although 201 consistent -results. Specifically, we find a significantly higher autonomic nervous system (ANS) ac- when Player-B is Human or Computer-box. Finally, we note that the psychophysiological reaction of 210 subjects rating FACE high in human-likeness is significantly higher than that experienced by subjects 211 interacting either with Computer-box or Human, regardless of the rating of human-likeness.

212
Regarding the relationship between the psychophysiological reaction of participants and their choices, 213 we do not find any significant correlation using the two indexes EDAsymp and EDAHFnu. However, 214 if we split our participants into two groups according to whether they express a stronger (or weaker) 215 psychophysiological reaction than the median level of the distribution of EDAsymp (see Table 6), we 216 can observe that those who experienced a stronger reaction are also less likely to choose 'In' in both      In this experiment, the robot (as well as the computer box) decides whether to Roll or Don't Roll accord-483 ing to its emotional state and following its decision rules. In particular, a positive mood in SEAI (i.e., an 484 emotional state with positive valence) will lead the robot to be collaborative with the human player and 485 play Roll; while a negative mood in SEAI (i.e., an emotional state with negative valence) will lead the 486 robot to play Don't Roll (see Figure 5). The decision is taken at the end of the interaction with Player-A, 487 when the subject goes out of the room, and so out of the field of view of the robot.

488
If in the moment in which the robot has to take a decision, it is in a qualitatively neutral mood (v=0, 489 regardless the arousal), the decision will be taken randomly (50%). Participants' behavior during all

Mean comparisons across groups 501
To compare the means (µ) of the distribution of a random variable for two independent groups (X, Y), 502 we perform t-Student tests on the equality of means. Specifically, to test for µ x = µ y (when the variances 503 σ x and σ y are unknown and replaced by s x and s y ) the test is t =x −ȳ  Note that all physiological indexes computed during the interaction with the agent were normalized

INSTRUCTIONS: English translation from Italian
Welcome! This experiment will last about 30 minutes. You will receive 5 Euro for your participation. Based upon the choices you will take in the experiment; you can earn additional money. We now ask you to turn off your mobile phone and to read the instructions carefully.
The aim of this experiment is to study how people take decisions. In particular, this experiment wants to study how people take decision when interacting with a humanlike robot.
Should you have any doubt, please do not hesitate to ask clarifications to the experimenter.
The data related to this experiment will be saved and analyzed anonymously. No video will be recorded.
In this experiment you will play with FACE i.e. a social robot which is able to prove and express its emotions. [with a computer-box which is given a system of social perception]. FACE [The Computer box] is also able to take its decisions autonomously, following its own behavioral rules. In this game, FACE [The Computer box] is programmed to choose autonomously between two actions: ROLL and DON'T ROLL a six-faces dice.
[In this experiment you will play with Deborah. Deborah can choose autonomously between two actions: ROLL and DON'T ROLL a six-faces dice.]

YOUR CHOICE
You will have to choose between two options: whether to play IN or OUT. Once you have done, we will wait for you to come back again to this room, to fill in a final questionnaire and receive your payment.