Investigating Objective and Subjective Social Exclusion in a Group Conversation Led by a Robot

Abstract

This study investigates ostracism-based social exclusion in multi-person interactions with robots. To examine this phenomenon, we will conduct a laboratory study in which participants engage in a simulated job interview with the robot Ameca acting as the interviewer.
The study compares objective exclusion (measured by the proportion of the robot's attention directed toward each participant) and subjective exclusion (participants' self-reported feelings of being ignored or excluded). We aim to identify the point at which objective exclusion leads to subjective feelings of exclusion and how this impact need fulfillment. After the interview, participants are allowed to stand somewhere else and are asked why they chose the same or a different standing position. Exploratory analyses will examine whether factors such as gender, height, or physical position (angle) relative to the robot influence the actual or assumed likelihood of being excluded.

Full text

1
Invesgang Objecve and Subjecve
Social Exclusion in a Group Conversaon
Led by a Robot
Authors:
Clarissa Sabrina Arlinghaus
Ashita Ashok
Karsten Berns
Günter W. Maier
Published: 03rd February 2025 DOI: hps://doi.org/10.17605/OSF.IO/HYB2S
Descripon:
(What is this study about?)
This study invesgates ostracism-based social exclusion in mul-person interacons with
robots. To examine this phenomenon, we will conduct a laboratory study in which
parcipants engage in a simulated job interview with the robot Ameca acng as the
interviewer.
The study compares objecve exclusion (measured by the proporon of the robot's aenon
directed toward each parcipant) and subjecve exclusion (parcipants' self-reported
feelings of being ignored or excluded). We aim to idenfy the point at which objecve
exclusion leads to subjecve feelings of exclusion and how this impact need fulllment. Aer
the interview, parcipants are allowed to stand somewhere else and are asked why they
chose the same or a dierent standing posion. Exploratory analyses will examine whether
factors such as gender, height, or physical posion (angle) relave to the robot inuence the
actual or assumed likelihood of being excluded.
Data collecon:
(Have any data been collected for this study already?)
No, no data have been collected for this study already.

2
Hypothesis:
(What’s the main queson being asked, or hypothesis being tested in this
study?)
Hypothesis:
1. Higher levels of objecve exclusion (measured by the robot's distribuon of aenon)
will lead to worse mood, and this relaonship will be mediated by subjecve
exclusion (parcipants’ self-reported feelings of being ignored or excluded). The more
people subjecvely feel excluded, the worse their mood is.
2. Higher levels of objecve exclusion (measured by the robot's distribuon of aenon)
will lead to lower need fulllment, and this relaonship will be mediated by subjecve
exclusion (parcipants’ self-reported feelings of being ignored or excluded). The more
people subjecvely feel excluded, the worse their need fulllment is.
Exploratory Quesons:
1. At what level of objecve exclusion do individuals begin to feel subjecvely excluded?
2. How do parcipants explain their experience of objecve exclusion by the robot?
3. Are there specic factors (e.g., age, gender, height, physical posion/angle) that
increase the likelihood of being excluded by the robot?
Dependent variables:
(Describe the key dependent variable(s) specifying how they will be measured.)
Objecve Exclusion:
Objecve exclusion will be measured by analyzing the robot's distribuon of aenon during
the simulated job interview. Precisely, we will calculate the proporon of interacons the
robot directed toward each parcipant. To improve interpretability, we will reverse the scale,
so that higher values represent higher levels of exclusion. The measure will be
operaonalized as follows:
𝑂𝑏𝑗𝑒𝑐𝑡𝑖𝑣𝑒 𝐸𝑥𝑐𝑙𝑢𝑠𝑖𝑜𝑛
= 1 − 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑞𝑢𝑒𝑠𝑡𝑖𝑜𝑛𝑠 𝑎𝑛𝑑 𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒𝑠 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑡𝑜 𝑎 𝑝𝑎𝑟𝑡𝑖𝑐𝑖𝑝𝑎𝑛𝑡
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑞𝑢𝑒𝑠𝑡𝑖𝑜𝑛𝑠 𝑎𝑛𝑑 𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒𝑠 𝑡𝑜 𝑎𝑙𝑙 𝑝𝑎𝑟𝑡𝑖𝑐𝑖𝑝𝑎𝑛𝑡𝑠
Subjecve Exclusion:
Subjecve exclusion will be assessed using two self-report items from the manipulaon
check by Williams (2009):
For each queson, please select the number that best represents the feelings you were
experiencing during the interview.
• "I was ignored."
• "I was excluded."
Parcipants will rate these items on a 5-point Likert scale (1 = Strongly disagree, 5 = Strongly
agree) to capture their perceived exclusion during the interacon with the robot.
We adjusted the instrucon to reect parcipants' feelings during the job interview rather
than during a game.

3
Change in Mood:
Mood will be assessed using a scale from Williams (2009) that covers four posive aects
and four negave aects:
For each queson, please select the number that best represents the feelings you were
experiencing during the interview.
• "Good"
• "Bad"
• "Friendly"
• "Unfriendly"
• "Angry"
• "Pleasant"
• "Happy"
• "Sad"
Parcipants will rate these items on a 5-point Likert scale (1 = Not at all, 5 = Extremely)
before and aer their interacon with the robot. The dierence between the two measured
values is included in further analyses as a change in mood.
We adjusted the instrucon to reect parcipants' feelings during the job interview rather
than during a game.
Need Fulllment:
Need fulllment will be measured using a scale from Rudert and Greifeneder (2016). This
scale assesses parcipants' sasfacon with fundamental psychological needs following the
interacon. Four Items will be rated on a 9-point Likert scale to measure overall Need
Fulllment:
Please indicate how you felt during the interview. During the interview, I felt…
• "invisible ……… recognized"
• "devalued ……… valued"
• "rejected ……… accepted"
• "powerless ……… powerful"
We adjusted the instrucon to reect parcipants' feelings during the job interview rather
than during a game.
Standing posion:
Aer the interview, the parcipants are allowed to choose a new posion where they would
prefer to stand during a new round of interviews with the robot. We ask them "Where do
you want to stand for a second round of interviews?" and they can choose one out of ve
standing posions that are marked on the ground. We have also documented their standing
posion from the rst round so we can see if they want to change their posion of if they
decide to remain standing in the same posion.
Aer their decision, we ask parcipants the following two open-ended quesons to
understand their moves beer:
• "Why did you choose this posion?"
• "What do you think has a greater inuence on the conversaon with the robot: your
standing posion or the content of your answers?"
These responses will be analyzed qualitavely with a large language model using the LLM-
Assisted Inducve Categorizaon (LAIC) method (Arlinghaus et al., 2024).

4
Feedback:
At the end, parcipants have the chance to provide feedback though an open-ended
queson:
• "Is there anything else you would like to tell me about your job interview experiences
or how they can be improved?"
These responses will be analyzed qualitavely with a large language model using the LLM-
Assisted Inducve Categorizaon (LAIC) method (Arlinghaus et al., 2024).
Condions:
(How many and which condions will parcipants be assigned to?)
There are no predened experimental groups in this study. All parcipants will engage in a
simulated group job interview with the humanoid robot Ameca, which operates in its
standard mode without any behavioral manipulaon.
During the interview, Ameca will naturally select which parcipants to ask quesons and
respond to. This interacon will be video recorded to accurately capture and later analyze
the robot's distribuon of aenon among parcipants.
Parcipants will not be assigned to specic condions. Instead, their experience of objecve
exclusion will be quaned based on how frequently Ameca directed quesons and
responses to them during the interacon.
All parcipants stand during the interview so that height dierences are more visible.
Analyses:
(Specify exactly which analyses you will conduct to examine the main
queson/hypothesis.)
1. Mediaon Analysis:
To test whether objecve exclusion indirectly aects change in mood or need
fulllment through subjecve exclusion, two simple mediaon analyses will be conducted.
• Predictor: Objecve Exclusion
• Mediator: Subjecve Exclusion
• Outcome: Change in Mood or Need Fulllment
A (bootstrapped) mediaon analysis (e.g., with PROCESS Model 4) will be used to test the
indirect eect. 95% condence intervals will determine signicance.
2. Threshold Analysis (Objecve to Subjecve Exclusion):
To idenfy the point at which objecve exclusion leads to subjecve exclusion, a piecewise
(segmented) regression analysis will be conducted.
• Predictor: Objecve Exclusion
• Outcome: Subjecve Exclusion
This analysis will detect whether a threshold eect exists. If the relaonship is linear, a
standard regression will be used. For non-linear paerns, a piecewise regression will idenfy
crical turning points. Alternavely, a Johnson-Neyman analysis may explore the exact point
where the relaonship becomes signicant.

5
3. Qualitave Analysis of Open Responses:
Parcipants’ open-ended responses will be analyzed qualitavely. The LLM-Assisted Inducve
Categorizaon (LAIC) method will be used so that a large language model will inducvely
generate categories and assign parcipant statements accordingly (Arlinghaus et al., 2024).
4. Exploratory Analysis of Addional Factors:
To explore whether individual characteriscs inuence the likelihood of being excluded, the
following exploratory analyses will be conducted:
• Predictors: Age, gender, internaonal student, language prociency, height, and
physical posion (angle relave to the robot)
• Outcome: Objecve Exclusion
A mulple regression analysis will be performed to determine if these factors predict how
oen parcipants were engaged by the robot. If the data structure allows, interacon
terms may be included to examine potenal moderang eects.
5. Assumpon Checks:
• Normality, linearity, homoscedascity, and mulcollinearity of variables will be
checked.
• If assumpons are violated, non-parametric tests or data transformaons will be
applied.
Outliers and Exclusions:
(Describe exactly how outliers will be dened and handled, and your precise
rule(s) for excluding observaons.)
Outliers:
• For connuous variables, outliers will be idened using the 3 × IQR rule (values
falling three mes the interquarle range below the rst quarle or above the third
quarle).
• If extreme outliers (+- 3 x IQR) are detected, analyses will be conducted with and
without these cases to assess their impact on the results. Outliers will only be
excluded if they substanally distort the ndings.
Exclusion Criteria:
Parcipants will be excluded from the analysis if any of the following condions apply:
• Technical Issues: Signicant technical problems during the interacon (e.g., recording
failures, robot malfuncons) that prevent accurate measurement of objecve
exclusion.
• Non-compliance: Parcipants who do not follow the task instrucons (e.g., refusing
to engage with the robot or disrupng the interview).
• Early Terminaon: Parcipants who do not complete the experiment and drop out
before the session is nished.
• Incomplete Data: Parcipants who fail to complete the post-interacon
quesonnaire assessing subjecve exclusion and need fulllment.
All exclusions and the nal sample size will be transparently reported.

6
Sample Size:
(How many observaons will be collected or what will determine the sample
size?)
We plan to recruit approximately 35 parcipants for this study to account for potenal
exclusions due to technical issues or non-compliance.
This sample size was determined using a G*Power 3.1 (Faul et al., 2009) analysis for a linear
mulple regression (xed model, R² deviaon from zero) with:
• A large eect size of f² = 0.35 (based on Hartgerink et al., 2015)
• An α level of 0.05
• A desired power of 0.80
• Two predictors in the regression model.
The analysis indicated a required sample size of 31 parcipants. To ensure sucient power
and account for possible exclusions, we aim to recruit approximately 35 parcipants.
Recruitment will connue unl this number is reached or unl praccal constraints (e.g.,
me, resources) limit further data collecon. Any deviaons from the planned sample size
will be transparently reported.
Other:
(Anything else you would like to pre-register? E.g., data exclusions, variables
collected for exploratory purposes, unusual analyses planned)
The robot-mediated job interview scenario was inspired by prior research (Kumazaki et al.,
2017; Nørskov et al., 2020; Zafar et al., 2021). Even though this is a simulaon, we are
following the high-quality standards that apply to real interviews. We simulate a structured
job interview with pre-prepared queson since structured interviews are superior to
unstructured interviews (Levashina et al., 2014; Sacke et al., 2022). Apart from a short
round of introducons, which contains biographical elements, the job interview consists of
situaonal quesons and quesons about past behavior, as these have a parcularly high
validity (Taylor, 2002). The eight behavioral or situaonal quesons are presented randomly
and stem from the Career Development Services of the North Central State College (2016):
• "Tell me about a challenge or conict you’ve faced at work (or university), and how
you dealt with it."
• "How do you deal with pressure or stressful situaons?"
• "Tell me about a me you failed. What did you learn from that experience?"
• "How do you handle working with people who annoy you?"
• "Tell me about your proudest achievement. What are you most proud of?"
• "If I were your supervisor and asked you to do something that you disagreed with,
what would you do?"
• "What if you were given an assignment that was too dicult for you? How would you
resolve the issue? Has this happened?"
• "Describe how you would handle a situaon if you were required to nish mulple
tasks by the end of the day, and there was no conceivable way that you could nish
them."

7
We will use the robot Ameca. The Ameca hardware specicaons are the following (cf.
Roboc Research Lab, 2022):
• Height 1870 mm
• Weight 49 kgs
• Binocular eye mounted cameras
• Binaural ear mounted microphones
• Speaker on chest
• 51 degrees of freedom
• Eye tracking based on face detecon
• External ZED2 camera mounted on chest
• External microphone
References:
Arlinghaus, C. S., Wul, C., & Maier, G. W. (2024). Inducve Coding with ChatGPT – An
evaluaon of dierent GPT models clustering qualitave Data into categories. OSF Preprints.
hps://doi.org/10.31219/osf.io/gpnye
Career Development Services of North Central State College (2016). Common Interview
Quesons – Pracce List.
hps://ncstatecollege.edu/documents/StudentSrvcs/CareerSvcs/Interview%20Quesons_Lis
t.pdf
Faul, F., Erdfelder, E., Buchner, A., & Lang, A.-G. (2009). Stascal power analyses using
G*Power 3.1: Tests for correlaon and regression analyses. Behavior Research Methods,
41(4), 1149–1160. hps://doi.org/10.3758/BRM.41.4.1149
Hartgerink, C. H. J., van Beest, I., Wicherts, J. M., & Williams, K. D. (2015). The ordinal eects
of ostracism: A meta-analysis of 120 Cyberball studies. PloS ONE, e0127002.
hps://doi.org/10.1371/journal.pone.0127002
Kumazaki, H., Warren, Z., Corbe, B. A., Yoshikawa, Y., Matsumoto, Y., Higashida, H., Yuhi, T.,
Ikeda, T., Ishiguro, H. & Kikuchi, M. (2017). Android Robot-Mediated Mock Job Interview
Sessions for Young Adults with Ausm Spectrum Disorder: A Pilot Study. Froners in
Psychiatry, 8, 169. hps://doi.org/10.3389/fpsyt.2017.00169
Levashina, J., Hartwell, C. J., Morgeson, F. P., & Campion, M. A. (2014). The Structured
Employment Interview: Narrave and Quantave Review of Research Literature. Personnel
Psychology, 67(1), 241-293. hps://doi.org/10.1111/peps.12052
Nørskov, S., Damholt, M. F., Uhlhøi, J. P., Jensen, M. B., Ess, C., & Seibt, J. (2020). Applicant
Fairness Percepons of a Robot-Mediated Job Interview: A Video Vignee-Based
Experimental Survey. Froners in Robocs and AI, 7, 586263.
hps://doi.org/10.3389/frobt.2020.586263

8
Roboc Research Lab (2022). Emah. Empathic Mechanized Anthropomorphic Humanoid
System. hps://rrlab.cs.rptu.de/en/robots/ameca
Rudert, S. C. & Greifeneder, R. (2016). When it’s okay that I don’t play: Social norms and the
situated construal of social exclusion. Personality and Social Psychology Bullen, 42(7), 955-
969. hps://doi.org/10.1177/0146167216649606
Sacke, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revising meta-analyc esmates
of validity in personnel selecon: Addressing systemac overcorrecon for restricon of
range. Journal of Applied Psychology, 107(11), 2040-2068.
hps://doi.org/10.1037/apl0000994
Taylor, P. J. (2002). Asking applicants what they would do versus what they did do: A meta-
analyc comparison of situaonal and past behaviour employment interview quesons.
Journal of Occupaonal and Organizaonal Psychology, 75(3), 277-294.
hps://doi.org/10.1348/096317902320369712
Williams, K. D. (2009). Ostracism: A temporal need-threat model. In M. P. Zanna (Ed.),
Advances in experimental social psychology (Vol. 41, pp. 275–314). Elsevier Academic Press.
hps://doi.org/10.1016/S0065-2601(08)00406-1
Zafar, Z., Ashok, A. and Berns, K. (2021). Personality Traits Assessment using P.A.D. Emoonal
Space in Human-robot Interacon. In Proceedings of the 16th Internaonal Joint Conference
on Computer Vision, Imaging and Computer Graphics Theory and Applicaons (VISIGRAPP
2021) (pp. 111-118). SciTePress. hps://doi.org/10.5220/0010161801110118