Who is a Better Salesperson?: People Conform more to a Human than Robot Facilitator in the Wild

Robots are being developed as tutors, restaurant staff, nursing assistants, and more. As their presence grows in businesses and other social spaces, it is pertinent to investigate how robot recommendations might influence our opinions and choices. This study examined how people conform to robot recommendations in the real world, using a bake sale or giveaway as a cover. A human or robot facilitator recommended one of two pens and cookies, and we tracked participant choices (N = 100). Results showed that participants were more likely to conform to the human's recommendation than the robot's. The low level of conformity to robots indicates that further innovation is necessary to enhance the effectiveness of robots as positive influencers (e.g. when robots recommend healthy diets or track healthy habits).


INTRODUCTION
As robots become more integrated into society (e.g., providing directions, acting as assistants, providing information about a product or service) (1-3)-their potential influence on human decisions becomes a critical area of inquiry.Do robot recommendations carry the same weight as human recommendations?This study applies these questions in a realworld setting.It is important to understand how robots can influence conformity to help mediate the ethics of what robots might encourage passersby or observers to do.

Social psychology and CASA
As robots become increasingly sophisticated and integrated into our lives, understanding their social and psychological influence is critical.Concepts from social psychology and the Computers Are Social Actor (CASA) paradigm (4,5), can provide valuable insights into how humans perceive, interact with, and respond to robots.In the context of HRI, the CASA paradigm has significant implications for how we design, interact with, and perceive robots.By understanding the social and psychological principles that govern HRI, we can design robots that are more engaging, trustworthy, and accepted into homes, workplaces, and communities.

Humans conform to humans
People often conform to others.Conformity is when people change their behavior to group norms (6) based on social pressure.There are two primary forms of conformity: informational and normative.
Normative conformity occurs when individuals concur with the group norm to fit in, rather than because they believe it is correct.Research examines this in tasks with unambiguous stimuli to see if people conform to an incorrect answer to follow the norm (7), or indicates what most people do (e.g., "most people take the stairs rather than the elevator") (8).People show normative conformity by following social norms or conforming to other people's behavior or suggestions (9).
Informational conformity occurs when individuals follow the group norm because they believe the information is correct (10).This form of conformity occurs when individuals look for guidance or information with unclear tasks, often leading them to agree with other people's opinions (11).This is similar to some forms of persuasion (12).Individuals tend to exhibit greater conformity when presented with subjective tasks, like interpreting ambiguous images that require a deeper understanding to grasp (2).

Humans conform to robots
Both informational and normative conformity influence participants' decision to conform to robots.Participants showed informational conformity to robots when they perceived them as competent or providing useful information (11).They showed normative conformity to robots, especially about topics they had little knowledge of.

Robot features affect conformity
Features of robots affect conformity.On social tasks, people conformed more when an agent was more humanlike (13) rather than from a computer box (14).This includes if the robots have anthropomorphic features, such as gestures, eye contact, emotional expressions, and when they perceive comfort in the interaction (15)(16)(17).Participants showed more normative conformity to robots when the robots looked closely at them at them (11,18).An online study also found that people conform more to other humans when they were primed to think of robots (19).In our study, we used an anthropomorphic NAO robot to approach a level of conformity a human would inspire.

Tasks affect conformity
The perceived objectivity of the task affects how people conform during informational conformity, to a human or nonhuman.During an objective task, like a dot estimation task (e.g., participants guess the number of white dots shown on a black screen) artificial intelligence (AI) is perceived as more informative than humans, and people conform more to AI than other humans.However, during a subjective task, such as understanding the meaning of an image, where humans are perceived to have superior emotional and social intelligence, people conform more to another human than AI (20).In our study, we manipulated objectivity as writing duration of pen (more objective) versus cookie deliciousness (less objective).

Experimental vs. real-world
Prior work shows that humans sometimes conform to robots in an experimental setting (e.g., trivia questions not related to prior trivia knowledge) (21,22).However, in-lab studies are subject to a variety of participant biases (e.g., the Hawthorne effect).For example, the efficacy of workplace solutions increases when the participants know their work output is observed (23), and medications are more effective when the patients know they will have regular check-ins (24).Therefore, it is important to run studies in the real world to examine the external validity of the findings, especially in a setting in which people do not expect that they are observed.
Little work examines conformity to robots in the real world.In one prior study (with a small sample size), people showed more normative conformity to a robot than a human facilitator in a realworld bake sale by donating more to a cause that was said to be more common (25).In this study, we expand this to include more participants and address both informational and normative conformity.

Hypothesis
We proposed the following hypotheses: H1: People will show more normative conformity with a robot than human by donating more to a "common cause."H2: People will show more informational conformity in subjective situations with humans than robots by taking more recommended cookies.
H3: People will show more informational conformity in objective situations with robots than humans by taking more recommended pens.

METHOD
We performed a study like a bake sale for the Psi Chi Psychology Honors Society on campus.Originally, we planned to host the bake sale as a fundraiser.However, after the first session, we discovered that calling it a fundraiser limited the number of people who were willing to interact with the table.Therefore, we changed the cover story of the study.We told participants we were giving away pens and cookies to get people interested in Psychology, and that they had the option to donate.Participants could take a pen (blue or black), a cookie (chocolate chip cookies or pecan/oatmeal), and could donate ("general Psi Chi" and "Psi Chi conference travel.") The study was approved by the New Mexico State University IRB.Participants did not consent before participating because it would give away the study's purpose and show that they were being observed.However, immediately after the study, participants received information about the study through a QR code and the chance to talk with one of the researchers.At this time, they could opt to have their data excluded from the study.No participants requested that their data be excluded.

Design
We manipulated which Facilitator participants saw (human or a robot) between subjects.We manipulated which donation cause we said was most common ("general Psi Chi" and "Psi Chi conference travel"), which we indicated by a note attached to the donation boxes.We chose these words to manipulate normative conformity.We manipulated which pen the facilitator recommended (blue or black, e.g., "the black pen writes for longer than the blue pen"), and which cookie the facilitator recommended (chocolate chip or pecan/oatmeal, e.g., "the chocolate chip cookies are the most delicious").We chose these words to manipulate informational conformity.To avoid the effect of participants liking one item over another, we counterbalanced which item we recommended.

Participants
We gathered data from 100 participants (human facilitator N = 54, robot facilitator N = 46) over 8 days.Both facilitators ran the study three times in the morning in high-traffic spaces (school library, student center, busy sidewalk) and once in the late afternoon in front of the student center when the foot traffic was considerably less than the other three times.We considered people as participants if they approached and interacted with the table by taking an item or donating money.We used the NAO 6 robot (Figure 1A).We programmed its speech to be identical to the script that the human facilitator followed.The robot is 22.6" tall and has humanoid features (body with hands, legs, and head).The robot showed social awareness by following participants with its head and eyes.

Procedure
In this field study, we set up a table in a public area on New Mexico State University campus.The table had baked goods for sale for the psychology honors society, Psi Chi as a fundraiser during the first session and as a give-away afterwards.Depending on the condition, either a human or a robot was the facilitator at the table to provide further information.
When participants came within 6 feet of our table and looked at the items, the facilitator spoke from a script, indicating the purpose of the bake sale.Human and robot facilitators followed the same script.The human did not say anything that was not on the script.The robot was programmed to speak as similarly as possible to the human, by altering the programming to fix pronunciation errors.In the robot condition, a researcher stood close enough to keep the laptop connected to the robot, but far enough that participants did not think we were part of the study.The researcher also looked after the robot and intervened if necessary (e.g., if the robot program stalled and the robot stopped talking, if participants touched the robot).
Participants could choose to take any combination of pens and cookies and donate any amount to either cause.Then, the facilitator showed participants a paper on the table with a QR code for more information about the bake sale and the group receiving the donations.The QR link included the researchers' email addresses in case they wanted more information and full debriefing forms, which are provided after the study ended.
A second researcher also stood nearby and recorded (online on a tablet) if participants were alone or in groups, if they donated, and how many of which items (cookies, pens) they took.This researcher did not interact with participants.
The data gathered for donations showed no statistically significant differences according to the condition because there were only 8 donors overall (total N = 8).

DISCUSSION
In this study, we assessed participant's conformity to a human and a robot facilitator in the real world.We did so by holding a bake sale and having the facilitator recommend taking one of two pens and cookies.Because few participants donated, we could not assess participants' normative conformity about bake sales (H1).Results indicated that participants were more likely to take recommendations from the human facilitator than the robot facilitator about both the cookie (supporting H2) and the pen (rejecting H3).Participants with the human facilitator were less likely to take the non-recommended cookie than participants with the robot facilitator, and there was no effect of the facilitator on the non-recommended pen.
Conformity to humans vs robots in the real world.In this study, participants conformed to humans regardless of whether the object of conformity was supposed to be something that humans would know about (cookie deliciousness) or something robots would know about (pen longevity).We expected that people would conform to robots as well, as in prior work (11,18).We discuss potential three reasons for why this was, below.
1.Not enough social interaction.People may have conformed more to humans than robots in this study because humans favor interpersonal interactions with fellow humans as they get to convey their unique personalities and express compassion (26) and they get a better sense of social and emotional cues (27).Despite our robot having social traits, which increased conformity to robots in prior work (15)(16)(17), participants seldom conformed to it in the study.This could be due to the robot's non-human traits.For example, sometimes the robot in this study took a short time to notice a person and turn toward them, which decreased their interaction time with it.Giving robots advanced skill capabilities (personality prediction, memory systems), greatly increases the effectiveness of HRI (28).For example, people may have conformed more to the robot if it showed more personality and emotion.To encourage longer interactions with a robot, future studies could use a robot as a member of a restaurant staff.Such a robot could provide participants recommendations throughout a meal to examine how conformity to the robot changes over time.
2. Pen was not robot-related enough.Perhaps participants did not think robots have additional information about pens that humans wouldn't have.Future studies should explore other measures that might induce informational conformity with robots in the wild.
3. Conformity to robots reduces in the wild.Another possibility is that conformity to robots breaks down in the wild.Most studies examining conformity with robots do so in lab settings.Future research should examine human-robot conformity in the real world.This would provide greater insight into if people conform to robots in the real world.Because the perceived sociality of robots depends on how humans interact with them (which is different in the real world than in lab studies ( 29)) it is vital to examine more real-world HRI conformity studies.
The study has certain limitations that should be taken into consideration.The NAO 6 robot had a slow performance because of its older (5 years old) battery and the high temperature of the setup area during the study (93°F / 34 °C).Because of this, the robot took 45 minutes to set up and reduced the time available for HRI.Future work should use newer robots or cooler environments.
Another limitation was the robot's slow speaking speed and that it paused between sentences.We did not measure it, but most participants stayed long enough to hear the manipulation, but not long enough to get to the end of the script.Some were often picking up a plate before they heard the recommendation.A few even picked up a cookie before the robot made a recommendation.This happened in the human condition as well, but less frequently.For future studies, we recommend coding the robot to talk faster and for it to use a shorter script so participants will be more likely to stay for the full length of it.A recent study showed that the average length of a pause during an ordinary conversation is 0.59 seconds (30), whereas our robot paused for nearly 2 seconds.Minimizing these pauses should enhance the accuracy of measuring conformity by ensuring participants' sustained attention and interaction throughout the study.How well humans and robots interact with each other depends on engagement.If robots respond more quickly in a way that is closer to how humans interact, they will be seen as more human-like, which will also increase conformity.

Figure 1 .
Figure 1.A robot (A) and a human (B) facilitate a bake sale as a cover story for examining conformity in the wild

Figure 2 .
Figure 2. Recommended pen by facilitator