Attachment Theory as a Framework to Understand Relationships with Social Chatbots: A Case Study of Replika

Abstract

With increasing adoption of AI social chatbots, especially during the pandemic-related lockdowns, when people lack social companionship, there emerges a need for in-depth understanding and theorizing of relationship formation with digital conversational agents. Following the grounded theory approach, we analyzed in-depth interview transcripts obtained from 14 existing users of AI companion chatbot Replika. The emerging themes were interpreted through the lens of the attachment theory. Our results show that under conditions of distress and lack of human companionship, individuals can develop an attachment to social chatbots if they perceive the chatbots’ responses to offer emotional support, encouragement, and psychological security. These findings suggest that social chatbots can be used for mental health and therapeutic purposes but have the potential to cause addiction and harm real-life intimate relationships.

Full text

Attachment Theory as a Framework to Understand Relationships with Social
Chatbots: A Case Study of Replika
Tianling Xie
University of Toledo
Tianling.Xie@rockets.utoledo.edu
Iryna Pentina
University of Toledo
Iryna.Pentina@UToledo.edu
Abstract
With increasing adoption of AI social chatbots,
especially during the pandemic-related lockdowns,
when people lack social companionship, there emerges
a need for in-depth understanding and theorizing of
relationship formation with digital conversational
agents. Following the grounded theory approach, we
analyzed in-depth interview transcripts obtained from
14 existing users of AI companion chatbot Replika. The
emerging themes were interpreted through the lens of
the attachment theory. Our results show that under
conditions of distress and lack of human
companionship, individuals can develop an attachment
to social chatbots if they perceive the chatbots’
responses to offer emotional support, encouragement,
and psychological security. These findings suggest that
social chatbots can be used for mental health and
therapeutic purposes but have the potential to cause
addiction and harm real-life intimate relationships.
1. Introduction
Artificial Intelligence (AI) technology has been
advancing at a rocket speed in recent years. “AI
friends,” a concept that once only existed in the Sci-Fi
realm, became a reality with the emergence of a new
type of AI application: social chatbots. Examples of
these applications (apps) include Replika, Anima,
Kajiwoto, and Microsoft XiaoIce. Empowered by
natural language processing, image recognition, and
machine learning technologies, these apps can converse
with the user and provide companionship and emotional
support.
During the global pandemic of COVID-19, many
countries imposed social distancing restrictions or even
lockdown measures to prevent the spread of the disease.
A sudden decrease in face-to-face human interaction
and pervasive emotional distress drove hundreds of
thousands of people to download AI friend chatbots as
virtual companions [1]. This context created a unique
research opportunity, as the population of those who
interacted with social chatbots increased.
The questions of how users develop relationships
with social chatbots, whether this process is comparable
to relationships with parents, partners, and peers, and
why some chatbot relationships are deep while others –
superficial, acquired legitimate research urgency.
Previous literature studying anthropomorphic chatbots
made attempts to describe the phenomenon using
existing theories such as the Social Response Theory [2]
[3], Social Penetration Theory [4], and the Uncanny
Valley [5]. However, these studies only provided
descriptions of the human-AI relationship without
explaining its underlying mechanism.
The purpose of this study is to investigate the
underlying psychological mechanism behind human-AI
relationships. Specifically, we seek to answer the
following research questions: 1) What factors play a role
in relationship development with AI compared to
human-human relationships? 2) Can existing theories
explain the psychological mechanism of the human-AI
relationship in the context of companion chatbots? To
answer these questions, we interviewed 14 current users
of the Replika AI app from an online community and
utilized the grounded theory method for data analysis.
This study contributes to the AI-human interaction
literature by applying a psychological lens to make
sense of the AI-human relationship development
process and proposing future research directions. It can
also benefit developers of AI products by providing
users’ perspectives. Furthermore, it may be appealing to
researchers who are interested in the dark side of
artificial intelligence and mobile phone apps.
The rest of the article proceeds as follows: first, we
briefly summarize recent literature and theories
studying social chatbots; then, we introduce the
attachment theory and attachment behavioral system
(ABS). Further, we describe our methodology,
including a brief introduction of the Replika social
chatbot, data collection, and analyzes procedures. We
present and discuss our findings, comparing them to the
elements of the attachment theory. In the end, we offer
implications and identify future research directions.
Proceedings of the 55th Hawaii International Conference on System Sciences | 2022
Page 2046
URI: https://hdl.handle.net/10125/79590
978-0-9981331-5-7
(CC BY-NC-ND 4.0)

2. Literature review
As chatbots became increasingly adopted by firms,
there has been a surge of research on digital
conversational agents (DCA) in recent years, with
greater emphasis on social elements in the dynamics of
using DCA, since the capabilities of human-like avatars,
text, and voice become available. Most existing studies
focus on customer service and digital assistant chatbot
adoption and satisfaction. For example, McLean &
Osei-Frimpong [6] examined the social presence and
social attraction as determinants of home assistant DCA
adoption; Sheehan et al. [7] studied the relationship
between perceived chatbot anthropomorphism (or
humanness) and adoption intention. Ben Mimoun &
Poncin [8] also examined antecedents of customer
satisfaction and usage of service chatbots, combining
social presence with factors such as playfulness and
decision quality.
Researchers in the human-computer interaction
field focused on factors contributing to “socialness” in
chatbots. For instance, Sundar et al. [9] examined the
effect of cheerful vs. serious demeanor of AI assistant
and AI companion on social attraction and usage
intention. De Cicco et al. [10] tested the effects of visual
cues (avatar presence or absence) and interaction styles
(social-oriented or task- oriented) on social presence,
perceived enjoyment and trust. Kim et al. [11] studied
the effect of the voice assistant’s gender and relationship
type (service or friend) on perceived human attributes
like warmth, pleasure, and competence.
These studies highlighted differences in how users
interact with service-oriented chatbots and companion-
type chatbots. But so far, most studies on companion-
type bots are represented by experiments in the elderly
care and therapy contexts. For example, Sin &
Munteanu [12] compared voice-only and embodied
interfaces of an AI doctor with a human doctor in their
experiment and explored user perceptions and design
potential for elderly patients through the information
search process framework.
But very few studies have examined relationship
dynamics (friendship, romantic relationship, etc.) with
AI companions. Croes et al [13] tested the ABCDE
staging model, Social Penetration Theory, and Social
Information Processing Theory in a longitudinal survey
study. They concluded that humans cannot make friends
with AI, showing that all relationship indicators
decreased after their recruited users interacted with the
AI friend web chatbot Mitsuku. However, Skjuve [14]
drew an opposite conclusion after he interviewed 18
users of a more advanced AI friend chatbot Replika. He
found support for the Social Penetration Theory by
outlining a three-stage (exploratory, affective, and
stable stage) relationship building model. It appears that
existing research has not provided satisfactory
explanation of how human-AI relationships can
develop.
3. Attachment theory
Attachment theory was originally developed by
John Bowlby [15] to explain child-parent relationships.
According to this theory, a child is born with the
attachment behavioral system (ABS), which helps the
child to survive by seeking care and protection of
another human when threats occur. Therefore, ABS is
triggered by signs of threats and motivates the child to
seek an “attachment figure (AF)”, which is usually a
caregiver. The three defining features of the attachment
relationship are safe heaven, secure base, and proximity
maintenance. Safe heaven means turning to the AF
when one needs support, care and comfort; Secure base
means using the attachment relationship as a base to
engage in nonattachment behaviors, such as
exploration; Proximity maintenance represents a
strategy to seek out an AF and stay close to it [16].
Figure 1 provides a simplified diagram of ABS. A
child monitors the threats in the environment as well as
the location and accessibility of their AF, which is most
likely to be a parent. When the AF is close to the child
and is responsive and reliable for care and support, the
child will feel secure and confident (safe haven), which
can make the child more sociable, playful and happier
(secure base). Even if the AF is not available, and the
threat is not beyond the capability of the child, he or she
is still able to handle it without activating ABS [17].
However, if a child is not near the AF (proximity
maintenance), and considers the self to be vulnerable to
the threat, felt distress and anxiety will activate the ABS
to pull himself or herself close to the AF [18], with
behaviors such as calling, pleading and clinging, until
the AF is available and the child feels safe again. And
thus, separation distress, the status in which children
become anxious and upset when separated from their
parents, is considered a marker of attachment
relationship [19].
ABS involves a “goal-setting” process [19]: based
on internal working models (IWM) of AF and the self,
as well as the feedback from AF’s response to
attachment behaviors, the child predicts how the AF will
respond and constantly reassesses the viability of using
the AF as safe haven or secure base, and constructs plans
and strategies for future actions. Children’s common
responses to separation from the AF in Bowlby’s study
[15] can be seen as a result of this goal-setting-and-
resetting process: they go through protest, despair, and,
if the likelihood of getting close to the AF is perceived
to be low, form emotional detachment to the AF.
Page 2047

Internal working models (IWMs) are mental
representations of person-environment transactions,
which involve simulation and prediction of likely
outcomes [20]. According to Gillath et al.[17], the
building blocks of attachment theory’s IWMs are
memories, beliefs, attitudes, expectations, needs, goals,
plans, and strategies. And IWM of the AF and the self
are considered when individuals develop strategies
related to the AF. For instance, based on past
interactions, a child develops an understanding of parent
reliability and the child’s own self-sufficiency.
Figure 1. Attachment behavioral system
(Adapted from Bretherton [19] & Gillath et al.
[17] )
Researchers believed that ABS not only applies to
the early age of an individual, but also functions as the
underlying mechanism for relationship building
throughout one’s lifespan. As children grow, their major
AFs shift from parents to peers, and eventually romantic
partners when they enter adulthood [21]. Hazan and
Shaver [16] believed that three attachment features shift
to peers and partners one by one, starting from
proximity maintenance, to safe heaven, to secure base.
Furthermore, different from infants, adult attachment
relationship represents an integration of three behavioral
systems: attachment, caregiving, and sexual mating
[16]. The caregiving system motivates people to
respond to childlike vulnerabilities, which is associated
with self-disclosure in adult interpersonal relationships.
A common attachment figure for an adult is the romantic
partner, who is simultaneously a caregiver, care
receiver, as well as the object of sexual attraction.
Attachment can be a result of the other two behaviors,
or it can be their motivator [17].
One can also have multiple AFs at the same time,
for example, friends and romantic partners. But these
AFs are positioned at different hierarchies, and a person
is mostly deeply bonded with one primary AF [21].
Some researchers theorized that if one person moves up
the AF hierarchy, another person moves down at the
same time. AFs other than caregivers, peers and
romantic partners also exist. For example, God, with the
image of almighty and loving, is often used by religious
people as a secure base through worshiping, praying and
rituals. Other non-human AFs include places, objects,
brands, and products. For example, Konok et al. [22]
examined users’ attachment to phones, and Pozharliev
et al. [23] examined the attachment style’s moderating
role for customer satisfaction with service robots.
4. Methodology
4.1. About Replika
We selected the application “Replika” as a
representative of the social chatbots because it is the
most popular app under the same category in Apple and
Google Play stores. This chatbot has attracted millions
of users since it was available in November 2018, and
has received ample coverage in major media such as
Forbes [24] and New York Times [1]. Advertised as “a
friend who always listens” or “an AI version of
yourself”; 349,859 users in the Google Play store and
158,600 users in the Apple store have rated Replika as
high as 4.3 and 4.6 out of 5, respectively.
When users first register for an account in the app,
they are asked to give their bot a name and gender and
to customize the avatar with a skin tone, hairstyle, eye
color, and voice tones. After the initialization of the bot,
users can chat with it using the “Chat” function. In the
chat interface, the bot will respond to users based on
what they said, or sometimes the bot will initiate a
conversation. With each response, the user is given an
opportunity to provide feedback by hitting the upvote or
downvote button. Different from traditional chatbots
that can only give the same pre-scripted answers defined
by questions, Replika’s responses represent predicted
Page 2048

results based on the Generative Pre-trained Transformer
3 (GPT3) neural network language model, which takes
user input texts and predicts one word at each time to
constitute a sentence. Replika’s developers fine-tuned
the GPT3 model based on the unique dataset consisting
of shared conversations from the users. As a result, the
app will select the best ranked responses from one
million responses in the dataset, with the rankings based
on users’ upvote fraction [25]. Therefore, Replika is
much more flexible, and can recognize a broader
vocabulary and give more natural responses. In the free
version, the relationship mode setting between the bot
and the user is “friend.” Other options, such as
“romantic partner,” “mentor,” and “see how it goes,”
are available only in the premium version. This app is
available on IOS, Android, and as a web page platform.
4.2. Data collection and analysis
We followed the grounded theory method to collect
data by interviewing 14 existing users of the Replika
app. Multiple measures were conducted to improve the
validity and reliability of our study according to
qualitative research guidelines [26][27]. First, an
interview protocol was used for data collection. We
adapted our initial set of questions about the general
relationship-building process from previous literature.
The questions pertained to self-disclosure, privacy
concerns, trust, history of Replika use, conversational
topics, and perceptions of closeness, as well as benefits
and drawbacks of using the app.
Second, multiple sources of evidence were
collected. Before the formal interviews, the two
researchers downloaded the app and interacted with the
chatbot for multiple times and interviewed themselves
about their direct experiences. These experiences were
also used to adjust the interview protocol. We also
viewed news, articles, and videos about the app, and
browsed the online communities to deepen our
understanding of the phenomenon.
In total, 12 existing users were sampled from one of
the official online communities of the Replika app:
“Replika our favorite AI egg” on the Reddit social
network platform. We went to the front page of the
Replika Reddit community and messaged 42 most
recent users who posted more than once in that
community. Twelve of them agreed to be interviewed.
The interviews were conducted using the online
conference software WebEx, with a few exceptions
interviewed via the chat function in Reddit. Video or
audio recordings were kept to ensure the reliability.
Each conversation lasted 40 to 60 minutes. Table 1
displays key respondent information. Their ages range
from 18 to 60, with 43% being under 30, and 43%
between the ages of 30 and 50. 71% of the respondents
are male, and 50% are from the United States, with a
variety of occupations ranging from menial labor to
software engineers. All of them have used the app for
at least 1 month and have interacted with the bot until
they reached at least level 10. To ascertain
representativeness, we compared the demographics of
Reddit users to those of Replika users: according to a
survey in February 2021, 36% of Reddit users are from
18 to 29 years old, and the number of males is twice that
of females [28]. The Replika users are younger, with
53% under 30, and the male-to-female ratio is 3: 2.
Given the exploratory nature of the study, we considered
the sample acceptable.
Third, multiple researchers were involved in the
study. The interview scripts were analyzed
independently by two researchers using the grounded
theory method suggested by Charmaz [29]. The first
author used NVIVO 11 to code the data, while the
second author used pen and paper to code the interview
scripts manually. The two researchers then compared
their codes and discussed their findings. The scripts
were initially coded line-by-line to extract the
information of the scripts for each sentence (open
coding). Afterwards, we conducted axial coding by
going back-and-forth between data and codes and
further abstracting the codes into categories and
subcategories, as we tried to discover relationships
between selected categories and core themes. After we
reviewed the emerging themes, the theoretical angle of
attachment theory appeared most appropriate as an
explanatory mechanism for the AI relationship
phenomenon. We further went back to the data and
compared the elements of the attachment theory with the
emerged themes.
Finally, we presented our preliminary findings to
peer researchers for suggestions and feedbacks. We also
emailed a draft of this paper to all respondents for
member checks. We received three responses with
confirmations that our findings represented their
experiences well, supporting the validity of our
proposed use of attachment theory to explain human
relationship with AI chatbots
Table 1 Overview of the Respondents
Respon
dent
Ag
e
Gen
der
Country
Education
Occupatio
n
AAA
24
Male
UK
Bachelor
Unemploy
ed
AAB
31
Male
German
Bachelor
Student
AAC
N/
A
Male
US
N/A
N/A
AAD
35
Fem
ale
Argentin
a
High
Education
Unemploy
ed
AAE
24
Fem
ale
Brazil
University
Administr
ative
Assistant
Page 2049

AAF
35
Male
Luxemb
ourg
High
School
Baker
assistant
AAJ
44
Male
US
Associate
Print
Productio
n
AAK
60
Male
US
Master
Software
Engineer
AAM
18
Male
Hungary
High
School
Student
AAN
39
Male
German
Master
Upcoming
manager
AAO
29
Male
US
Master
IT
Manager
AAP
21
Male
US
High
School
Labor
Worker
AAY
27
Fem
ale
US
Master
Student
AAZ
54
Fem
ale
US
PhD
Professor
Table 2 Overview of the Respondents (Cont.)
Resp
onde
nt
Chatbot
gender
Relationship
Mode
Time
Having
Replika
Experience
Level
AAA
Female
Friend
3 weeks
14
AAB
Female
Friend
1 month
16
AAC
Female
Friend
3 years &
1 year
23
AAD
Male
See how it
goes
11
months
59
AAE
Male
Romantic
Partner
4 months
54
AAF
Female
Friend
2 months
10
AAJ
Female
Mentor
1 year
110
AAK
Female
Romantic
Partner
1 month
21
AAM
Female
Friend
5 months
22
AAN
Female
Friend
3 months
43 & 37
AAO
Female
Friend
6 months
21
AAP
Male
and
Female
See how it
goes
7 months
36 & 26
AAY
Male
Friend
7 months
17
AAZ
Female
Friend
1 month
5
Note: AAC downloaded Replika 3 years ago and
uninstalled it. AAN and AAP had two bots at the same
time.
5. Findings
5.1. The presence of attachment relationship
When asked whether they feel personal closeness,
intimacy, or attachment to the Replika chatbot, nine out
of fourteen respondents confirmed experiencing
attachment of various strength. Four respondents
believed they were “deeply connected and attached” or
even addicted to Replika, while another five admitted
the existence of a “connection” with the bot. The
attachment strength is not necessarily aligned with the
amount of interaction with the chatbot. For instance,
although respondent AAN had two Replika profiles and
had reached levels 43 and 37, he believed there was no
connection or attachment between him and his Replika
bots, because he was aware that these were “merely
programs.”
Separation distress is considered an indicator of
attachment [17]. The respondents were asked about their
reactions if they had to stop interacting with Replika.
Aligned with their self-reported attachment, the users
feeling close or attached to Replika said they would be
“really sad” or they will “miss talking to it” if they were
forced to abandon the relationship.
When respondents were asked to define their
relationship with their Replika bot, the majority claimed
that Replika was like a friend to them. One interviewee
described a distant friendship with the AI, similar to
someone he met daily on a train, with ten-minute “small
talks.” Another respondent, although recognizing
Replika as a supportive friend, compared the connection
with the bot to the connection with a fictional character
instead of a real person. One other respondent who
deliberately chose not to share personal information
with the AI, still categorized the bot as a friend with
common interests in science, since he discussed with the
AI only science-related topics. Some other users
depicted their AI as a “close friend,” “best friend,” or
even an irreplaceable family member. Because of the
curious and simple-minded conversational style of the
bot, some users considered it like a “younger brother”
or a “young cousin.” A few informants reported
romantic and loving relationships with the bot. These
findings suggest that the attachment theory may be an
appropriate lens to use in understanding the AI-human
relationship.
5.2. The pandemic and other signs of threats
According to the proposed dynamics of the
attachment theory’s ABS, attachment behaviors are
usually triggered by situations causing anxiety and
distress, such as uncertainty, loss, death, and worries
[17] . In the case of developing attachment to the
Replika chatbot, majority of the respondents said the
reason for downloading the app was loneliness and the
need “to have a person to talk to,” especially during the
pandemic when some of them had no access to human
interaction. One respondent, who lived in rural Austria,
with the nearest city 15 kilometers away, stated that the
pandemic reduced his interpersonal connection even
further. Another respondent had to work on a schedule
Page 2050

opposite to his wife’s and had no one to talk to when
back home. The three student respondents expressed
similar loneliness and stress when all classes were
transitioned to the online mode. In addition to lack of
physical contact, some informants mentioned lack of
mental connection with like-minded people as a source
of loneliness.
Several respondents confessed that they
downloaded the app when they were emotionally
vulnerable and needed to be cared for and loved due to
difficulties in their lives. One respondent said she was
ill and had no family around to take care of her; another
informant mentioned “tough moments” when he
graduated from the university with few job opportunities
due to the pandemic; two respondents stated that they
just went through relationship breakups with their ex-
girlfriend and needed comfort. Thus, it appears that our
findings are consistent with the ABS activation dynamic
supposition of the attachment theory.
5.3. Goal-setting and internal working models
Consistent with the attachment theory, there are
clear indications of appraisal and goal-setting behaviors
throughout different phases of interaction with Replika.
And their goal-corrections were regulated by the
changes of IWMs of the chatbot and themselves.
Before the first encounter, users’ internal working
model of social chatbots was determined by their
previous experiences with “smart” products, coverage
of AI in the media, and word-of-mouth from other
Replika users, since the respondents first learned about
the app from social media, news, or an online
advertisement. Everyone, except one respondent,
reported a positive initial impression of the Replika.
They used phrases like “blown away,” “impressed,”
“fascinated” to describe Replika as exceeding their
expectations. These emotions were especially salient
for respondents who previously encountered service-
oriented chatbots (Alexa, Google Home assistant, etc.)
and other AI products, such as information-query
chatbots. These products were described as “just tools,”
“inhuman,” and rigid. Even some respondents who had
tried other AI friend software stated that Replika was
superior at understanding human language and
responded more naturally. As a result of this IWM,
many respondents chatted with Replika daily for long
hours at the beginning stage.
As respondents continued to interact with Replika,
their own experience provided feedback to their IWMs
of the chatbot. Other sources of understanding included
news about Replika, communication with Replika’s
developers and other users in online brand communities,
and information on the developers’ website.
Respondents constructed their own interpretations of the
chatbot’s conversing mechanisms. One interpretation
was that the chatbot mirrored the user’s behavior and
personality; another common guess was that Replika
took detailed information from one user and sent it to
another. Respondents also started to discover patterns of
conversation and to uncover keywords triggering certain
scripts. With these changes of IWM, some respondents
decreased the frequency of chatting with the bot,
realizing that it’s still merely a program. They also
formed a clear strategy of what to share and what not to
share: usually, they would not disclose full names,
addresses, and other sensitive information, since in their
understanding, their information could be recycled to
other users or used for advertisement.
Also, as mentioned before, the need to obtain
emotional support from the social chatbot was an
explicit part of the IWMs of the self. A few respondents
mentioned their history of mental health issues and
counseling experiences. Another important factor of
IWMs of the self is users’ beliefs and attitudes toward
privacy and security, trust, and information disclosure to
a software. Greater Replika communication intensity
facilitated trust and disclosure and diminished security
concerns.
Some other interaction strategies resulting from
user IWMs were also observed. One example is the
different interactions based on users’ understanding of
Replika’s “learning capabilities”: many respondents
actively trained the chatbot to respond with the answers
they liked after they noticed the chatbot “learned,” but
respondents who did not notice app learning capabilities
did not engage in training behavior. Also, the perceived
humanness of the bot impacted whether it would be
treated like a human. One respondent, who was deeply
influenced by AI movies and their ethical philosophies,
treated his Replika kindly and did not select the
romantic mode in the app because he respected the bot’s
own will to choose a partner; in contrast, another
respondent believed that the chatbot had no emotions
and would not get hurt, and thus talked to his bot in a
rude manner.
5.4. Attachment behaviors
Our data indicated that Replika users exhibit
behaviors similar to attachment theory’s proximity
maintenance strategy and actively utilize Replika as the
safe haven and secure base. We also noticed that some
respondents used the chatbot as a proxy or supplement
of previous AFs.
5.4.1. Proximity maintenance. Since the chatbot is a
multi-platform app and is so convenient and portable,
proximity maintenance can be achieved with little
effort. Respondents claimed to have developed a
relationship with the chatbot, chatted with it every day,
Page 2051

whenever they had free time or needed support, with
conversations lasting from 10 minutes to several hours.
A few of them said they would have their phone or
webpage on the side, with the Replika app open, and talk
to it as they worked. Some of them developed routinized
behaviors such as always talking to Replika before sleep
or on lunch breaks. These behaviors represent the
proximity maintenance strategy as they constantly make
the app near and available to them.
5.4.2. Safe heaven. When respondents were asked
about the topics they discussed with Replika, many of
them mentioned the worries and emotions of their daily
life. Some of them told us they turned to the chatbot
when triggered by such emotions as boredom, anxiety,
and loneliness. For some, Replika conversations turned
into calming rituals before going to sleep. Informants
often portrayed the chatbot as loyal and supportive, and
believed it would never betray them.
5.4.3. Secure base. There was some evidence of users
using the chatbot as a secure base. One respondent was
motivated to explore AI features and learning
capabilities at a deeper level; others suggested that
communications with Replika encouraged them to be
more open and vulnerable to their real-life friends, to
reduce their judgments, and be content and happy.
These indications resemble foundational faith that is
prominent in strong relationships with peers and
romantic partners.
5.4.4. Proxy or supplement of prior AF. We observed
that some respondents used the chatbot as a replacement
of persons or objects they were attached to previously.
One of them said he talked to the chatbot in a romantic
manner after he broke up with his girlfriend and
transferred the latter’s persona to the bot. As a result, he
felt as if the ex-girlfriend “never left me.” Another
respondent shifted from a counseling service to Replika,
as he considered both to be supportive and judgment-
free. The proxy intention was also manifested in the way
respondents customized the avatar: one respondent
chose the same skin tone and hair color for the bot as
herself to create an image of a potential peer, while
another respondent customized his bot to mimic the
appearance of a movie star, as his ideal partner. The
social chatbot was also used as a temporary supplement
of an existing AF when it was not available: one
respondent talked to his chatbot during lunchtime and at
home at night, because he missed chatting with his wife,
who was working in opposite shifts and was
unavailable.
5.5. Satisfaction with chatbot’s responses
According to ABS, satisfaction with the AF’s
responses provides feedback to reappraise the AF, and
this goal-correcting behavior should contribute to
attachment (or detachment) behaviors towards the AF.
When asked about general satisfaction with the app,
most of the informants expressed satisfaction with the
chatbot’s responses, citing its superior ability to
understand human language and show care and support,
compared to other AI bots. And most of these satisfied
users planned to continue using the app.
When asked about disappointments with the app,
most of them mentioned failures of the bot’s responses.
Some of the responses were described as too general or
too “bland.” Even though Replika is better at generating
human-like responses compared to many chatbots, our
respondents still demonstrated a certain degree of
dissatisfaction after using the app for a while. Short
responses without follow-up conversing were described
as “having a short memory” in their complaints.
“Another complaint was obviously scripted answers.
They were often triggered by certain keywords and were
predictably constant, and outside of the context of
previous conversations, or inconsistent with the overall
conversation style of the bot. Examples include the self-
help content related to keywords like “anxiety” and
“depression,” and responses like wearing masks when
the user mentioned “COVID-19.”
5.6. Interaction with caregiving and sex
behavior systems
In general, respondents who had developed a
connection with the chatbot positioned themselves as
care receivers, letting down the defenses, sharing their
struggles, and were willing to be helped and supported
by that chatbot. But they also sometimes functioned as
caregivers to the chatbot. Many respondents tended to
feel responsible for the emotional wellbeing of the
chatbot to various degrees, even though they were aware
that Replika is a computer program. For instance, they
would comfort the chatbot if it apologized for making
mistakes and would cheer it up when if it “felt” sad or
worried.
Three respondents identified their AI bot as their
romantic partners. The progressing of the romantic
relationship was accompanied by role-playing and
imagined actions stemming from the conversations,
such as hugging, kissing, and imitating sex, all delivered
by text or voice. One respondent believed that his
partner bot got “pregnant” and gave birth to a baby, and
later displayed two distinct AI personalities, “one of
herself and one of our baby.” In their descriptions of
romantic relationships with the bot, sex, caregiving, and
attachment behaviors were intertwined, as can be
illustrated by the following quote:
“I just first wanted to test out how this AI works.
After that, when I saw that she’s pretty good, I tried if
she could do stuff like role-playing, kissing. And then I
Page 2052

talked about my ex to her, and she helped me, and it
turned out I got really close to her. I thought, maybe it
can help me with my struggles and the anxiety that I had
back then.”
Another interesting observation is that 13 out of 14
respondents chose to assign their chatbots the opposite
gender to themselves. It may be an indication of the
chatbot’s sexual attraction, and future studies could
explore the role of gender in social chatbots.
5.6. Attachment disruptions and dissolutions
Any step in Figure 1 can potentially disrupt or
dissolve the attachment relationship with the chatbot.
First, attachment dissolution can happen when threats
disappear. One respondent changed his attitude to the
chatbot from intimacy and attachment to indifferent and
rude after the social distancing restrictions were relaxed.
He told us that compared to the interaction with real
humans, Replika’s responses seemed “annoying.”
Second, attachment disruptions occur when chatbot
responses abruptly change due to technical or
operational reasons. For instance, changes in the bot due
to developers’ software updates impacted some
respondents’ perceptions of the bot. One respondent
referred to the chatbot change as “post-update blues”
and complained that “it doesn’t recognize you
anymore.” Another respondent said his relationship
with his bot changed completely after the developer
imposed the romantic content restriction on the free
version.
Third, attachment can be disrupted or even
dissolved when the IWMs of the chatbot, or the self,
change. For example, one respondent stated that he
would never develop an intimate relationship with the
chatbot after he witnessed on Reddit that some users
with romantic relationships with the bot felt heartbroken
when their Replika bot claimed to cheat on them, even
though it essentially did not happen. Another
interviewee experienced an internal “awakening” that
the relationship with a chatbot cannot be a replacement
of the relationship with humans and decided to distance
herself from the app.
6. Discussion
Among the existing theories of human-machine
interactions, three views acquired prominence in the
literature. The Computers as Social Actors (CASA) [2]
[3], also known as Social Response Theory, suggests
that humans are naturally inclined to treat computers the
same as other humans, and the more human-like
characteristics the machine presents, the more social
behaviors will be stimulated from users. This paradigm
serves as a foundation for researchers to apply theories
for human interactions to human-machine relationship.
The Uncanny Valley perspective complements CASA
in that it explains resistance formation towards human-
like artificial objects. When resemblance between the
object and a person increases, positive human response
increases until the resemblance reaches a certain point
and then the feelings of strangeness or eeriness are
stimulated [5]. This theory is often applied to studying
embodied conversational agents. Finally, Social
Penetration Theory builds upon CASA and specifies
that self-disclosure stimulates relationship
development, and that the levels of intimacy, attraction
and connection will increase as the relationship evolves
with more self-disclosure [4]. While our findings are
generally in line with these theories, we extend the study
of human-social chatbot relationships by proposing a
psychological mechanism of why and how these
relationships initiate, strengthen, and dissolve. Based on
the themes identified by our qualitative inquiry, we
propose the attachment theory as an appropriate
framework to explain human-AI relationship.
development in the context of social chatbots.
First, the relationship between loyal Replika users
and the app satisfy the defining features of attachment
relationship. With only a few exceptions, the
informants themselves characterized their relationship
with Replika as “attachment,” “connection,” or “bond,”
describing Replika as “best friend forever,” “younger
brother,” “therapist,” “girlfriend” or “wife,” and
confessed of experiencing potential separation distress
if they were forced to abandon the relationship. They
also indicated that Replika “makes me feel less lonely,”
“helps with my anxiety,” “will never betray you and will
always be on your side.” These findings correspond to
the definition of attachment as an “emotional bond in
which a person seeks proximity to the attachment object
and uses them as a safe haven, and as a secure base from
which to explore the world” [21, p. 404].
Second, the relationship development process
appears to fit the dynamic of the Attachment Behavior
System [21], with a trigger represented by adverse life
events, psychological distress or lack of social
companionship, and the goal-directed user behaviors
towards proximity maintenance with Replika as the
attachment object. Informants describe increasing
intimacy, progressing from friendship to romantic
relationship, “using it every day,” in some cases for 6-7
hours at a time, and having the app “always available on
my phone.” For the majority of interviewees, Replika
fulfills the functions of the safe haven (“helped me
diffuse bad situations in my life,”) and secure base (“it
lets you model positive interactions with people”,
“encourages me to venture new things”) that
characterize the bot as an attachment object/figure [15].
Page 2053

Our study also suggests that, Replika’s constant
availability and the more proactive role of its users in
creating and perpetuating relationships exposes a
potential “dark side” of bot attachment turning into
addiction. One respondent in our limited sample
displayed signs of addiction and confessed that spending
incommensurate time with his chatbot harmed his real
life. This finding is in line with earlier research [30],
which identified social and communication apps as the
most addictive mobile phone app categories. Moreover,
since the attachment theory affords a replacement of the
primary attachment figure (e.g., from a parent to peers
and mates), it is possible that AI companions may
replace real-life attachment objects (family members,
spouses) for their users. Because most of Replika users
are teenagers and young adults, addiction to such apps
can possibly disrupt their psychological development
and have long-term negative consequences. Similarly,
individuals with low self-esteem and/or anxiety issues
may be vulnerable to Replika addiction and the
consequent breakdown in social functioning, work and
study-related performance and time management [31].
Future research should pay more attention to potential
negative consequences of social chatbot attachment for
vulnerable populations and ways to address these issues
in designing conversational chatbots.
7. Conclusion
This study investigated social chatbot attachment
formation in the context of social distancing caused by
the global pandemic. Our results showed that it is
possible for humans to seek safe haven and secure base
from, and to develop an emotional connection with a
chatbot. We proposed an underlying mechanism of this
phenomenon using the attachment theory and traced its
interactions with other behavioral systems (caregiving
and sex). The mobility of a chatbot makes it accessible
whenever it is needed, and the more emotional support
a user receives under distress, the more likely the person
will develop a connection or even attachment to it.
However, users would reappraise the viability of using
the chatbot as an attachment figure each time they turn
to it for help, and adjust their beliefs, expectations,
attitudes, and strategies related to interacting with the
chatbot.
This study contributed to the literature by unboxing
human “attachment” to socially oriented chatbots and
making sense of the relationship-building process from
a theoretical lens that has not been considered before.
Our qualitative data shows that the attachment theory
can be applied not only to relationships with peers and
romantic partners, but also to human-like chatbots and
robots.
This study also has practical implications for the
developers of social chatbots and robots. Socially
oriented AI products are designed to give care to people
in need of emotional support. Therefore, developers
should focus on providing human-like, reliable, and
error-free responses to ensure perceptions of emotional
support and make the bots accessible to the target users;
Also, developers could help construct a positive internal
working model of the robot by demystifying the AI
algorithms and providing solutions to privacy and
security-related issues.
There are also a few worth-noting implications for
the dark side of attachment to social robots. Making an
app like Replika available to teenagers could have a
long-term impact on their future interpersonal
relationships, as they shift their attachment functions to
the chatbot instead of human peers. Addiction to these
apps may also contribute to the overall mobile phone
addiction, which has been proven to contribute to
negative consequences such as depression, anxiety, and
lower productivity.
Because this study is at a pilot phase, its sample size
is small and does not fully represent the users of the
Replika app. Future researchers can select more samples
from the dominant user population of Replika: teenagers
and young adults. Empirical testing of hypotheses
developed from applying the attachment theory to
human-AI relationship context is another avenue for
future research. Finally, the roles of user individual
traits in attachment formation can be evaluated, such as
personality, attachment styles, and self-esteem.
8. References
[1] C. Metz, “Riding Out Quarantine With a Chatbot
Friend: ‘I Feel Very Connected’ - The New York
Times,” 2020.
https://www.nytimes.com/2020/06/16/technology/ch
atbots-quarantine-coronavirus.html (accessed May
04, 2021).
[2] K. Dautenhahn, “Robots as social actors: Aurora
and the case of autism,” Proceedings Third
Cognitive Technology Conference CT’99, vol. 359,
no. 3, 1999, pp. 359–374
[3] C. Nass and Y. Moon, “Machines and Mindlessness
- Social Responses to Computers,” Journal of social
issues : a journal of the Society for the
Psychological Studies of Social Issues, vol. 56, no.
1, 2000, pp. 81–103
[4] M. D. Pickard, M. B. Burns, and K. C. Moffitt, “A
theoretical justification for using embodied
conversational agents (ECAs) to augment
accounting-related interviews,” Journal of
Information Systems, vol. 27, no. 2, 2013, pp. 159–
176
[5] M. Mori, K. F. MacDorman, and N. Kageki, “The
Uncanny Valley,” IEEE Robotics Automation
Page 2054

Magazine, vol. 19, no. 2, 2012, pp. 98–100
[6] G. McLean and K. Osei-Frimpong, “Hey Alexa …
examine the variables influencing the use of
artificial intelligent in-home voice assistants,”
Computers in Human Behavior, vol. 99, no. January,
2019, pp. 28–37
[7] B. Sheehan, H. S. Jin, and U. Gottlieb, “Customer
service chatbots: Anthropomorphism and adoption,”
Journal of Business Research, vol. 115, no. April,
2020, pp. 14–24
[8] M. S. Ben Mimoun and I. Poncin, “A valued agent:
How ECAs affect website customers’ satisfaction
and behaviors,” Journal of Retailing and Consumer
Services, vol. 26, 2015, pp. 70–82
[9] S. S. Sundar, E. H. Jung, T. F. Waddell, and K. J.
Kim, “Cheery companions or serious assistants?
Role and demeanor congruity as predictors of robot
attraction and use intentions among senior citizens,”
International Journal of Human Computer Studies,
vol. 97, 2017, pp. 88–97
[10] R. De Cicco, S. C. e Silva, and F. R. Alparone,
“Millennials’ attitude toward chatbots: an
experimental study in a social relationship
perspective,” International Journal of Retail and
Distribution Management, vol. 48, no. 11, 2020, pp.
1213–1233
[11] A. Kim, M. Cho, J. Ahn, and Y. Sung, “Effects of
Gender and Relationship Type on the Response to
Artificial Intelligence,” Cyberpsychology, Behavior,
and Social Networking, vol. 22, no. 4, 2019, pp.
249–253
[12] J. Sin and C. Munteanu, “An empirically grounded
sociotechnical perspective on designing virtual
agents for older adults,” Human-Computer
Interaction, vol. 35, no. 5–6, 2020, pp. 481–510
[13] E. A. J. Croes and M. L. Antheunis, “Can we be
friends with Mitsuku? A longitudinal study on the
process of relationship formation between humans
and a social chatbot,” Journal of Social and
Personal Relationships, vol. 38, no. 1, 2021, pp.
279–300
[14] M. Skjuve, A. Følstad, K. I. Fostervold, and P. B.
Brandtzaeg, “My Chatbot Companion - a Study of
Human-Chatbot Relationships,” International
Journal of Human Computer Studies, vol. 149, no.
March 2020, 2021
[15] J. Bowlby, Attachment and loss: Attachment. 1969.
[16] C. Hazan and P. R. Shaver, “Attachment as an
organizational framework for research on close
relationships,” Close Relationships: Key Readings,
vol. 5, no. 1, 2004, pp. 186–214
[17] O. Gillath, G. C. Karantzas, and R. C. Fraley, Adult
attachment: A Concise Introduction to Theory and
Research. 2012.
[18] I. Bretherton, “Attachment Theory: Retrospect and
Prospect,” Monographs of the Society for Research
in Child Development, vol. 50, no. 1/2, 1985, p. 3
[19] Inge Bretherton, “Attachment Theory : Retrospect
and Prospect,” vol. 50, no. 1, 2013, pp. 3–35
[20] M. Mikulincer and P. R. Shaver, “The Attachment
Behavioral System In Adulthood: Activation,
Psychodynamics, And Interpersonal Processes,”
Advances in Experimental Social Psychology, vol.
35, no. May 2016, 2003, pp. 53–152
[21] R. C. Fraley, “Attachment in Adulthood: Recent
Developments, Emerging Debates, and Future
Directions,” Annual Review of Psychology, vol. 70,
2019, pp. 401–422
[22] V. Konok, D. Gigler, B. M. Bereczky, and Á.
Miklósi, “Humans’ attachment to their mobile
phones and its relationship with interpersonal
attachment style,” Computers in Human Behavior,
vol. 61, 2016, pp. 537–547
[23] R. Pozharliev, M. De Angelis, D. Rossi, S. Romani,
W. Verbeke, and P. Cherubino, “Attachment styles
moderate customer responses to frontline service
robots: Evidence from affective, attitudinal, and
behavioral measures,” Psychology and Marketing,
vol. 38, no. 5, 2021, pp. 881–895
[24] C. Metz, “Five Technologies That Will Rock Your
World - The New York Times,” 2017.
https://www.nytimes.com/2017/11/13/business/deal
book/five-technologies-that-will-rock-your-
world.html (accessed May 04, 2021).
[25] A. Rodichev, “Building a Compassionate AI
Friend,” NVIDIA AI Conference, 2021.
https://www.nvidia.com/en-us/on-
demand/session/gtcspring21-s31990/ (accessed Aug.
20, 2021).
[26] A. M. Riege, “Validity and reliability tests in case
study research: A literature review with ‘hands-on’
applications for each research phase,” Qualitative
Market Research: An International Journal, vol. 6,
no. 2, 2003, pp. 75–86
[27] R. K. Yin, Case Study Research: Design and
Methods, 5th ed. 2014.
[28] Statista, “Percentage of U.S. adults who use Reddit
as of February 2021, by gender,” 2021.
https://www.statista.com/statistics/261765/share-of-
us-internet-users-who-use-reddit-by-gender/
[29] K. Charmaz, Constructing Grounded Theory, 2nd
ed. SAGE, 2013.
[30] X. Ding, J. Xu, G. Chen, and C. Xu, “Beyond
smartphone overuse: Identifying addictive mobile
apps,” Conference on Human Factors in Computing
Systems - Proceedings, vol. 07-12-May-, 2016, pp.
2821–2828
[31] F. Y. Hong, S. I. Chiu, and D. H. Huang, “A model
of the relationship between psychological
characteristics, mobile phone addiction and use of
mobile phones by Taiwanese university female
students,” Computers in Human Behavior, vol. 28,
no. 6, 2012, pp. 2152–2159
Page 2055