Computers in Human Behavior Reports 18 (2025) 100682

Contents lists available at ScienceDirect

Computers in Human Behavior Reports
journal homepage: www.sciencedirect.com/journal/computers-in-human-behavior-reports

When does the “assistant” heuristic work? Examining the effect of AI job
titles in tasks with varying criticalities on the use of conversational
AI-based services
Mehmet Yanit
University Canada West, 1461 Granville Street, Vancouver, British Columbia, V6Z 0E5, Canada

A R T I C L E I N F O

A B S T R A C T

Keywords:
Hierarchical status
Humanlikeness
Competence
Warmth
Task criticality
Algorithm aversion

Recent marketing trends involve companies using low-status job titles, such as "assistant" (e.g., Google Home
Assistant), to label conversational AI agents. This strategy aims to activate an altruistic "assistant" heuristic and
enhance users’ willingness to use these AI agents. However, this paper—comprising one pretest (N = 313), three
experiments (N = 307, N = 300, N = 308), and one partial least squares structural equation modeling (PLS-SEM)
analysis (N = 309)—demonstrates that the effect of this strategy on willingness to use is positive only when the
task criticality is high. When the task criticality is not high, higher-hierarchy AI titles (e.g., "manager," "teacher,"
"analyst") generate greater willingness to use. The research examines three alternative serial mediation path­
ways—perceived warmth, perceived control, and perceived risks—to test for competing explanations alongside
the focal serial mediation through perceived humanlikeness and competence. Across the four studies, the serial
mediation via perceived humanlikeness and competence remained robust, even when controlling for alternative
pathways and scenario realism (Study 3). The final model indicates that when task criticality is not high,
increased perceptions of hierarchical status in conversational AI settings enhance perceived humanlikeness. This,
in turn, boosts perceived competence, ultimately increasing users’ willingness to use the AI. However, when task
criticality is high, the effect reverses—higher-status AI is perceived as less humanlike and less competent,
reducing users’ willingness to engage with it.

1. Introduction
Humanity is currently witnessing a remarkable proliferation of
Artificial Intelligence (AI) technology. The widespread adoption of AI is
reshaping how organizations assign hierarchical roles to AI-based ser­
vice agents, influencing their acceptance and effectiveness in customer
interactions. Today, AI plays a central role in engaging with customers
through conversational interactions while adopting hierarchyindicating job titles (e.g., assistant vs. manager), which affect
customer satisfaction, brand attitude, and the customers’ willingness to
buy the products recommended by these conversational agents (Jeon,
2022).
In human-provided services, job titles signal the agency and exper­
tise of the service provider, with higher-ranking titles indicating greater
authority and domain expertise, which can influence people’s percep­
tions of competence of those service providers (Grant et al., 2014; Trautt
& Bloom, 1982). Similar to human service providers, computerized
service providers such as AI are perceived as social actors to whom

human social norms apply (Sundar, 2008, pp. 73–100; Sundar & Nass,
2000). Therefore, in the context of AI, assigning job titles to AI service
providers may have comparable attitudinal implications, shaping per­
ceptions of competence and influencing behavioral outcomes, such as
the willingness to use AI-based services. Despite the growing prevalence
of AI, research on the effect of AI’s job titles remains scarce. For
example, Jeon (2022) demonstrated that when AI is assigned
higher-status job titles, it is perceived as more knowledgeable (more
expert) and likable, leading to increased willingness to follow
AI-generated recommendations. Similarly, existing research has shown
that higher-status AI is more strongly associated with attributes of
competence than lower-status AI (Sundar et al., 2017).
However, in the context of AI service providers, the underlying
mechanism linking AI’s job titles to perceptions of competence and, in
turn, people’s willingness to use remains largely unexplored. More
importantly, previous research has not examined contexts in which
lower-hierarchy AI may be perceived as more competent and therefore
preferred to a greater extent. This study addresses these gaps by

E-mail address: mehmet.yanit@ucanwest.ca.
https://doi.org/10.1016/j.chbr.2025.100682
Received 28 January 2025; Received in revised form 1 May 2025; Accepted 2 May 2025
Available online 5 May 2025
2451-9588/© 2025 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/bync-nd/4.0/).

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

investigating the role of perceived humanlikeness as a mediating factor
between conversational AI service providers’ job titles and their
perceived competence and examining this mechanism across three
levels of task criticality—low, medium, and high—through five studies
(one pretest) using a multimethod approach that combines partial least
squares structural equation modeling (PLS-SEM) and experimental de­
signs. The findings show that people’s willingness to use a conversa­
tional AI-based service based on their job title may depend on the task’s
criticality.
Moreover, previously, Sundar et al. (2017) and Jeon (2022) used
experiments to investigate the similar mechanisms. This paper’s
approach differs from previous studies in that this paper enhances the
reliability of findings by combining experimental research with
PLS-SEM (Kurtaliqi et al., 2024). Integrating PLS-SEM into the studies
improves predictive accuracy (Gudergan et al., 2025; Richter &
Tudoran, 2024) while allowing researchers to address more complex
research questions with greater external validity (Richter et al., 2022).
Experiments, on the other hand, add on greater internal validity to
findings. Therefore, when both combined validity of the findings in­
creases (Hair et al., 2021).
The status quo in marketing practices for AI agents shows that
companies commonly use low status titles like “assistant” to signal
altruism and helpfulness, aiming to enhance willingness to use (Sundar
et al., 2017). This approach is almost taken for granted, as the "assistant"
title is expected to activate a “helper” heuristic and attract people more
towards adopting AI in service settings (Puntoni et al., 2021). This paper
delves deeper into this idea and poses the following research question:
“When does this ‘assistant’ heuristic really work?”, claiming that
conversational AI’s job titles may show different effects at different
levels of task criticality.
In the pretest, the scenarios to be used in the following studies were
compared in terms of participants’ perceived task criticality using oneway between-participants ANOVA and planned contrasts. Results
showed that Scenario 1 ("home assistant task") and Scenario 2 ("task
management for work") were perceived as the lowest criticality tasks,
while Scenario 3 ("tutoring/teaching") was rated as having medium
criticality. On the other hand, Scenario 4 ("credit management") was
perceived as the highest criticality task.
In the structure of the current paper, each study examined the focal
conceptual framework at a different level of task criticality. Only Study
1a and 1b explored the same criticality level but employed different
methodologies—PLS-SEM and an experiment, respectively. Study 2 built
on the findings of Study 1b by eliminating potential biases associated
with the "manager" title, instead using a different high-status title
("teacher") in an educational context. Finally, Study 3, which tested the
mechanism in a high-task-criticality setting, revealed a reversal of the
focal effect. In this study, the conceptual relationships were also
controlled for the effect of scenario realism.
In summary, this paper argues that the “assistant heuristic” does not
always hold, and the effect of job titles on AI’s perceived humanlikeness,
competence, and people’s willingness to use an AI is dependent on task
criticality. In high criticality tasks, algorithm aversion is heightened
(Castelo et al., 2019; Wanner et al., 2021), and people perceive greater
risks in delegating the task to an algorithm. To mitigate uncertainties
and alleviate the tension, they tend to make quick, self-defending,
heuristic attributions about the service provider (Mozafari et al., 2022;
Pieters et al., 1998). Heuristically, people become more inclined to
search for a “helper” or “ally” from the available information cues in the
environment, aiming to increase the chances of positive outcomes
(Fiske, 2004, p. 69). Therefore, in high criticality tasks, “assistant” cues
may become more attractive to service recipients as they look for mental
shortcuts to make attributions with a risk aversive mindset (Kahneman,
2011). In these situations, hierarchical attributions are less relevant, as
people tend to allocate less mental effort to making such judgments.
Besides ally-seeking in high-criticality tasks, such tasks can also
inherently create a sense of vulnerability for human service recipients

due to their lack of domain-specific knowledge compared to the service
provider. This knowledge gap can lead to a perceived status loss relative
to the provider, further influencing their trust and willingness to engage
with AI-based services. (Struch & Schwartz, 1989). The current paper
argues that, amplified by heightened algorithm aversion due to task
criticality, high-status AI in high-criticality tasks is dehumanized as a
defensive mechanism by human service recipients to protect their own
status and perceived superiority. This effect is particularly strong
because, as a non-human entity, a conversational AI is expected to hold a
lower status than humans by default (Alicke et al., 1995; Woolley &
Risen, 2018; Bonaccio & Dalal, 2006). In other words, in high-criticality
tasks, the human service recipient’s perceived status relative to a
high-hierarchy conversational AI service provider diminishes. To psy­
chologically reconcile this gap, individuals may dehumanize the
high-hierarchy AI, reducing its perceived humanlikeness as a subcon­
scious defense response.
When the task criticality is not high, people are less threatened and
defensive, allowing them greater mental flexibility to evaluate the AI
service provider. In low criticality tasks, service recipients are more
capable of evaluating the AI analytically rather than heuristically (Blut
et al., 2021; Burton et al., 2020; Mozafari et al., 2022). Consequently, in
low criticality tasks, people are more familiar with the task and have
more mental capacity to analyze the service provider’s agency and
expertise beyond readily available cues like the “helper” signal
conveyed by assistant titles. They are also more likely to pick up more
subtle cues, such as hierarchy information indirectly signaled by job
titles. Assigning higher-status (or higher-hierarchy) job titles to
conversational AI service providers increases expectations of their
agency and expertise, which are subcomponents of competence, in much
the same way as in human-provided services (Grant et al., 2014; Teo­
doridis et al., 2019; Trautt & Bloom, 1982). The assignment of greater
agency and expertise through higher-hierarchy job titles should enhance
AI’s perceived humanlikeness, as these traits are characteristic of
“human nature” (Haslam, 2006). However, when these traits are not
conveyed, the entity undergoes a process of dehumanization, often
referred to as “mechanistic dehumanization” or “approximation to
automata” in the existing literature (Loughnan & Haslam, 2007). This
perceived shift towards automata can push people’s perceptions of
conversational AI toward the machine end of the human-machine con­
tinuum (Yanit et al., 2023), thereby decreasing its perceived human­
likeness. AI’s ability to manifest humanlike qualities is especially
important in conversational settings, as the competence of AI in these
settings is largely driven by its ability to compensate for the absence of
human service providers—a factor that, when fulfilled, leads to greater
adoption of AI in services. Therefore, aiming to resolve the contradiction
between market practices and the findings of the past research, this
paper suggests that the expected positive effects of the “assistant heu­
ristic” by the marketing practices should only be observed in high crit­
icality tasks. In contrast, when the task criticality is not high, people are
likely to show a greater inclination towards conversational AIs with
higher-hierarchy job titles as the evidence in the existing literature
shows. Furthermore, AI’s perceived humanlikeness and competence
should serially mediate this relationship in both conditions.
This paper also examines three alternative explanations regarding
the role of competence: perceived warmth, perceived control, and
perceived risks. Prior research (Yanit et al., 2023) indicates that
perceived humanlikeness enhances perceived warmth of AI, which may
serve as an alternative explanation for the effect of perceived hierarchy
in the focal relationship. In high criticality tasks, particularly, service
recipients may seek greater relationship closeness and warmth with the
service provider, making them more willing to use a service provided by
a lower-hierarchy AI (Fiske, 2004, p. 69). Additionally, conversational
AI is often perceived as eerie (Raff et al., 2024) and low in explicability
(Coombs et al., 2021), which may influence users’ perceived control
over the AI and their perceived risks from interacting with it, ultimately
affecting their willingness to use the service (Kim & McGill, 2011). In
2

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

high criticality tasks, the importance of control and risk perceptions
becomes more pronounced in people’s approach to conversational AI.
Specifically, in high criticality tasks, individuals experience a status loss
relative to the high-status AI and lose their sense of control over the
entity, which may lead to less willingness to use services provided by
high-hierarchy AI (Shaffer et al., 2013). These potential explanations are
tested through parallel mediation pathways to the focal relationship in
this paper.
Ultimately, this paper suggests that low-status titles like "assistant"
may not always be the most effective strategy for enhancing AI adoption.
Higher-status titles are likely to foster greater willingness to use AI, as
they enhance the perceived humanlikeness and competence of AI in
conversational settings, particularly when the task criticality is not high.
The remainder of the paper is structured as follows: related literature is
reviewed, and hypotheses are presented. This is followed by a pretest
and four main studies testing the focal ideas. The final section discusses
the results, along with their theoretical and practical implications,
limitations, and conclusions.

Despite the limited research on how the perceived hierarchical status
of AI agents influences subsequent user behavior (see Table 1), evidence
suggests that people may use the same cognitive mechanism to evaluate
AI-provided services based on the AI’s assigned job title. This is because
computerized service providers, such as AI, are perceived as social actors
to which human social norms apply (Sundar, 2008, pp. 73–100; Sundar
& Nass, 2000). Consequently, people treat AI similarly to human social
actors, adjusting their behavior according to the AI’s hierarchical status
(Kramer et al., 2011). This perspective allows the examination of
human-AI interactions through the lens of human-human interactions.
AI’s hierarchical status, signaled through its job title, may influence its
perceived status, similar to human service providers (Berger et al., 1972,
1980), particularly in the context of conversational AI, as it engages
customers in direct social connections. For example, Jeon (2022) found
greater favoritism towards a conversational AI agent titled as a customer
service manager compared to a representative. Sundar et al. (2017)
found that people are more socially attracted to higher-status conver­
sational AI, comparing a “companion” robot to an “assistant” robot.
Conversational AI agents are direct service providers with humanlike
qualities that verbally engage with customers firsthand. Similar to
human-provided services (Goodman & Gareis, 1993; Koski et al., 2015),
assigning higher-status job titles to conversational AI service providers
may lead to more positive attitudes from service recipients. Therefore,
limited research in the existing literature shows evidence against the
effectiveness of wide use of “assistant” titles by the AI companies to
market their products. However, can the countless companies that use
low-hierarchy “assistant” title to market their conversational AI agents
all be mistaken? Or are there specific situations where a conversational
AI with a “assistant” title actually increases users’ willingness to use it?
A closer look at the literature reveals that existing research largely
overlooks the contexts in which low-status AI could be preferred to a
greater extent. The current paper asserts that people’s attitudes toward
conversational AIs of different hierarchical statuses may be influenced
by the cognitive systems they engage when evaluating these agents,
which are shaped by situational factors (Kahneman, 2011). For example,
depending on their situational cognitive capacity, people evaluate ser­
vices either heuristically—using mental shortcuts, existing knowledge,
and first impressions to make quick judgments (Nazlan et al., 2018; Shi
et al., 2021)—or analytically, where they invest comparatively greater
cognitive effort to evaluate the information and make the most optimal
choice (Gervais & Norenzayan, 2012). People’s cognitive capacity is
limited when they feel threatened or when their attention is significantly
diverted by environmental noise or situational pressures (Wang et al.,
2021). In such cases, people are more likely to make decisions heu­
ristically, relying on the most readily available cues in the environment.
From this perspective, people may engage different cognitive sys­
tems depending on the perceived criticality of an AI-provided service.
Task criticality is defined by the extent to which a task’s outcome is
personally relevant to an individual and the potential harm they may
experience if the task is not satisfactorily completed. High-criticality
tasks are those that are more likely to elicit an emotional response,
require greater personalization, and have more detrimental conse­
quences if they fail (Chanseau et al., 2019; Følstad et al., 2024; Yanco &
Drury, 2002). Past research has discussed that people make heuristic
attributions to AI service providers in high-criticality tasks (Blut et al.,
2021; Burton et al., 2020; Mozafari et al., 2022). In performing
high-criticality tasks, people experience greater pressure and vulnera­
bility due to a lack of domain-specific knowledge and uncertainties
(Wanner et al., 2021). This anxiety triggers a defensive mechanism,
leading to "algorithm aversion" in these domains (Castelo et al., 2019).
When people experience algorithm aversion, they make heuristic and
biased decisions, avoiding cases that allocate greater agency to AI such
as hierarchy information (Heβler et al., 2022; Jussupow et al., 2020;
Longoni et al., 2019).
Therefore, in high-criticality tasks, people are more likely to rely on
the readily available altruistic signal of the “assistant” title to quickly

2. Related work
2.1. Hierarchical status and task criticality
Within human relationships, social hierarchies naturally evolve
(Sidanius & Pratto, 2003). These hierarchies are rooted in people’s
relative status to one another, a crucial mechanism for fostering coop­
erative behaviors in dyadic relationships (Groysberg et al., 2011; Koski
et al., 2015). Hence, individuals often navigate their place within the
hierarchy based on their status in relation to the other member of the
dyad. These hierarchies are innate social reflexes developed through
evolution, enabling humans to quickly assess and adapt their behavior
according to community norms (Boehm & Boehm, 2009; Kraus & Park,
2017; Saaty, 2001).
In organizations, hierarchical statuses are communicated through
job titles. Job titles clearly define the chain of command and ensure that
the organization operates in harmony, while communicating an agent’s
expertise and agency with prospective customers (Ahearne et al., 2005).
In these settings, service providers with higher hierarchical status are
stereotypically believed to possess more expertise (Anderson et al.,
2012; Ansoff et al., 2018) and agency (Haslam et al., 2009) compared to
those with lower status. For instance, a professor is typically believed to
manifest greater agency and expertise compared to a teaching assistant
(Koski et al., 2015), and engaging with a service manager instead of a
representative suggests an escalation to a higher authority, indicating
that the issue is being handled by someone with more expert knowledge
(Jeon, 2022). In a similar vein, research indicates that service recipients
perceive higher-status agents as more legitimate and are more willing to
interact with them (Smith & Tyler, 1996; Tyler & Degoey, 1995, 1996,
pp. 331–356), and higher status leads service recipients to acknowledge
and follow the higher-status party’s suggestions more closely (Fiske,
Cuddy, Peter, & Xu, 2024).
Just like human service providers, conversational AI service pro­
viders are also assigned job titles. In the context of AI, these titles often
function as a marketing strategy rather than factual reflection of the AI’s
capabilities. A look at the AI marketplace reveals that while many
conversational AI-based service providers perform identical or highly
similar tasks, their assigned job titles vary significantly. For instance, in
AI-driven psychological therapy, Wysa.AI and JoBot, despite having
nearly identical functions, are marketed under different titles: "coach"
and "psychologist," respectively. Both assist users with inquiries related
to Cognitive-Behavioral Therapy (CBT), relaxation techniques, and
various mental health concerns, as described on their respective web­
sites (jobot.ai/about.html; wysa.com/faq). Similarly, in the medical
field, LucasAI is promoted as “The Ultimate Medical Assistant” (luca­
shealth.ai), whereas Biotalk’s AI is labeled as the “AI Doctor” (biotalk.
ai/doctor/).
3

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Table 1
Related work on the examined conceptual relationships.
Author

Examined Relationship

Context

Result

Journal

Jeon (2022)

Perceived hierarchical status >
People’s attitudes

AI agents in customer
service

Journal of Business
Research

Sundar et al. (2017)

Perceived hierarchical status >
People’s attitudes
Perceived hierarchical status >
Perceived humanlikeness

Robotic companionship in
elderly care
Human stereotypes against
low social hierarchy

As perceived hierarchical status of a service agent
increases, people’s attitudes toward the agent become
more positive.
People manifested greater usage intentions for higherstatus robots with more serious manners.
People with low socio-economic status tend to be
dehumanized by others.

Perceived hierarchical status >
Perceived humanlikeness
Perceived humanlikeness >
Perceived competence
Perceived humanlikeness >
Perceived competence
Perceived competence >
User engagement
Perceived competence >
Recommendation intention
Perceived competence >
Willingness to use

Voice assistants

People perceive low-status AI (assistant) less
humanlike than high-status AI (companion).
Increasing humanlike autonomy increased perceived
competence of AI.
Perceived humanlikeness increases expected effort
competencies from an AI agent.
Humanlike competencies of AI service providers
increase user engagement.
Perceived competence increases recommendation
intention of users.
Perceived competence increases customers’
willingness to use.

Journal of Business
Research
International Journal of
Information Management
International Journal of
Information Management
Journal of Management
Information Systems
Current Psychology

Perceived competence >
Expected service quality

Service robots

Loughnan, Haslam,
Sutton, & Spencer,
2014
McLean et al. (2021)
Hu et al. (2021)
Gursoy et al. (2019)
Chandra et al. (2022)
Chi et al. (2024)
Kim and Hur (2023)
Yoganathan et al.
(2021)

Intelligent personal
assistants
AI service providers
Conversational AI service
providers
Service robots
Conversational AI service
providers

Increased perceived competence elevates expected
service quality from robotic service providers.

identify an “ally” from the available cues, rather than making analytic
judgments about the hierarchical connotation of the title within the task
context, which would also require greater cognitive effort. Supporting
this idea, the Stereotype Content Model (SCM) suggests that when
people make heuristic judgments of others, such as relying on stereo­
types, they primarily focus on the extent to which the other party could
be an ally (Fiske, Cuddy, Peter, & Xu, 2024). Some companies may have
already capitalized on people’s tendency to seek allies in high-criticality
tasks. For example, Credit Risk Community (CRiskCo) explicitly markets
their credit risk management AI as an ally for individuals involved in
credit analysis (CRiskCo, nd). Consequently, this paper argues that the
effect of perceived hierarchy should diminish in high-criticality tasks,
while the altruistic “in-group” connotations of the “assistant” heuristic
should be more strongly activated. This, in turn, fosters greater will­
ingness to use conversational AI, as heuristic thinking becomes more
dominant in such tasks.
In contrast, when the task criticality is not high, people tend to think
more analytically and rely less on heuristics. In such contexts, they have
greater cognitive capacity to assess an AI service provider’s ability based
on the hierarchical status conveyed through its job title, as this status
simultaneously signals agency and expertise. Existing research suggests
that, by default, when evaluating service providers, people attempt to
analytically assess their expertise and tend to rely more on those with
greater expertise (Bendapudi & Berry, 1997). Accordingly, this paper
suggests that people will evaluate AI service providers based on their
hierarchical status to a greater extent in low-criticality tasks, as these
tasks encourage more analytical thinking.
Formally, this relationship can be expressed as follows.

International Journal of
Human-Computer Studies
Social Psychology

International Journal of
Human-Computer
Interaction
Tourism Management

affect perceived humanlikeness. The following section explores a serial
mediation through perceived humanlikeness and competence depend­
ing on the perceived task criticality, as agency and expertise attributions
derived from AI’s hierarchical status may affect its humanlikeness and
eventually, shape perceptions of competence—given that agency and
expertise are subcomponents of humanlike competence (Haslam, 2006).
2.2. AI’s perceived hierarchy, humanlikeness, and competence across
tasks of varying criticality
In addition to changing their perceived hierarchical status, assigning
human role titles (e.g., manager, assistant) to AI agents may also change
their perceived humanlikeness. Conversational AI is inherently a hu­
manlike entity, which places it on a machine-human continuum, making
it neither completely machine nor completely human (Touré-Tillery &
McGill, 2015; Yanit et al., 2023). This paper claims that depending on
the task criticality people attribute different magnitudes of humanlike­
ness to conversational AI service providers based on their hierarchical
status.
To begin with, when attributing humanlikeness to other entities
(including real human beings), people use their own self-perception of
humanness as a reference point for comparison (Harris & Fiske, 2011,
pp. 123–134), often valuing their own humanness as higher than others
in this comparison (Haslam, 2006; Haslam et al., 2008; Leyens et al.,
2000). This bias arises heuristically, as it is easier to cognitively access
innate human cues, such as emotionality, in ourselves compared to those
in others (Haslam et al., 2008). Existing research also explains this
tendency through the "better-than-average" effect, where people tend to
attribute more desirable traits, such as humanness, to themselves
compared to others (Alicke et al., 1995; Woolley & Risen, 2018;
Bonaccio & Dalal, 2006). Moreover, humanness attributions are highly
dependent on perceived similarities between the attributor and the
attributed, whether in visual, cognitive, or social terms (Waytz et al.,
2010). Hierarchical status, which creates disparities in social standing,
influences these perceived similarities and leads to attributions of
“in-group” or “out-group” membership (Lee & Fiske, 2006). Regarding
humanlikeness, perceived hierarchical similarity to another person can
lead to stronger "in-group" feelings toward individuals at a similar status
level (Leyens et al., 2000). For instance, social status distance has been
shown to contribute to the dehumanization of individuals with lower
social status by those with higher social status and vice versa (Gwinn
et al., 2013; Haslam, 2006; Lammers & Stapel, 2011). Consequently,

Hypothesis 1a. (H1a): When task criticality is not high, people will
manifest greater willingness to use a high hierarchy (e.g., manager)
conversational AI.
Hypothesis 1b. (H1b): When task criticality is high, people will
manifest greater willingness to use a low hierarchy (e.g., assistant)
conversational AI.
Despite highlighting the behavioral differences of service recipients
in tasks of varying criticality, one question remains unanswered: What
underlying mechanism drives these differing preferences based on task
criticality? A key contribution of this paper is the argument that,
because AI is merely a humanlike entity, agency and expertise attribu­
tions should influence competence only to the extent that they first
4

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

those perceived as more dissimilar are attributed less humanlikeness
(Leyens et al., 2001). People’s negative judgments of dissimilar others’
humanness are an innate mechanism that fosters negative biases toward
those perceived as different (Haslam et al., 2000).
Being only a humanlike entity rather than a real human being,
conversational AI service providers are perceptually ascribed not “hu­
manness”, but “humanlikeness” and positioned on a lower rung in the
eyes of a human service recipient (Messingschlager and Appel, 2023).
Here, the human service recipient attributes AI’s humanlikeness on a
scale from zero to their own level of humanness. In doing so, people
attribute two different senses of humanlikeness to other conversational
AI—either as attributes that are “uniquely human” or as attributes that
form the foundation of “human nature,” as humans manifest certain
characteristics that are inherent and invisible forces endowed by nature
(Gergen, 1991, pp. 19–20). More importantly, existing research dis­
cusses that when “uniquely human” attributes are denied to entities,
they are perceptually approximated to “animals,” whereas when
“human nature” attributes are denied, they are approximated to
“automata” (Loughnan & Haslam, 2007). For example, manifesting
deliberate agency and acting on expertise are characteristics of “human
nature” (Haslam, 2006). This is because agency and expertise make
one’s behavior more independent and, therefore, less likely to conform
to pattern expectations (Harris and Fiske, 2011). Not conforming to
existing patterns of behavior is a form of manifesting deliberateness,
which triggers attributions of “mindedness” and humanness to
nonhuman entities (Bering, 2002). Existing research also discusses that
those who are high in social hierarchy are more humanized, as gaining
status is also associated with “human nature” (Haslam et al., 2008;
Loughnan, Haslam, Sutton, & Spencer, 2014; McLean et al., 2021).
Therefore, when analytically thought, lowering the hierarchical posi­
tioning of AI conversational agents corresponds to denying them agency
in the sense that AI’s actions are now presented as more “induced”
rather than driven by independent will (Haslam, 2006).
As discussed in the previous section, when a task is not highly critical
for a service recipient, they become more inclined to analytically eval­
uate the service provider. This means that in such tasks, people should
make greater associations between AI’s hierarchical status and its
humanlikeness due to the agency and expertise cues conveyed through
job titles. In these cases, conversational AI is perceived as a lower-status
entity by default, as analytical thinking makes it easier to broadly
recognize its mechanistic nature, which is considered “inferior” to its
humanlike nature (Gray et al., 2007). Hence, by increasing the percep­
tions of agency and expertise, ascribing higher-hierarchy job titles
should bring the AI closer to the upper limit of humanlikeness, which is
defined by the human service recipient based on their self-attributed
humanness. Similarly, attributing less agency to AI through
lower-status job titles leads to greater objectification of the AI while
approximating its social status to that of a mere “tool” rather than a
“humanlike entity,” a phenomenon defined in the literature as “mech­
anistic dehumanization” (Haslam, 2006). Therefore, this paper asserts
that in low-criticality tasks, hierarchical status should be more acces­
sible through job title information due to analytical thinking tendencies,
and in turn, people should perceive a low-hierarchical-status conversa­
tional AI (e.g., home assistant) as less humanlike.
As task criticality increases, the status dynamics between the
customer and the service provider shift in a way that favors the higher
hierarchy service provider over the customer. This happens because
people often lack the capacity to handle high-criticality tasks alone due
to task complexity, limited capabilities, or the risk of negative outcomes
(Følstad et al., 2024). These pressures trigger greater algorithm aversion
and heuristic thinking, prompting individuals to seek quick support to
handle the task favorably. Algorithm aversion is a risk response where
people make heuristic decisions by ignoring certain parts of the infor­
mation to create mental shortcuts for quicker decision-making (Wanner
et al., 2021). As a result, people become more inclined to approach
higher status AI as a “out-group” that needs to be kept under control, as

they experience a status loss due to their domain specific vulnerability
(Jussupow et al., 2020). Therefore, in such tasks, ascribing greater
agency and expertise to conversational AI through high-hierarchy job
titles can further expand the status distance between the human service
recipient and conversational AI service provider (Haslam & Bain, 2007).
For example, assigning a medical advice task (a high-criticality task) to a
conversational AI is likely to trigger algorithm aversion. In such a sce­
nario, assigning a “doctor” title, rather than a “medical assistant” title, to
a conversational AI may evoke a stronger sense of status loss due to the
higher agency and expertise connotations associated with the "doctor"
title (vs. "medical assistant") in a domain where the service recipient
(patient) is likely to have less agency and expertise. As a result, the
service recipient may feel vulnerable and threatened by allowing a
dissimilar agent (in terms of its nature and status) to make critical de­
cisions for them. In this case, they may prefer the lower-hierarchy
medical assistant, seeking reassurance and comfort by minimizing the
status gap between themselves and the AI service provider (Bendapudi &
Berry, 1997). As discussed previously, heuristic thinking and biased
reliance on one’s own humanness to set the upper limit in such tasks
make service recipients more likely to evaluate a lower-hierarchy AI as
more humanlike due to status similarities as hierarchical status simi­
larities trigger an “in-group” feeling (Pieters et al., 1998), while dis­
similar others are dehumanized as being “out-groups” (Leyens et al.,
2001).
Moreover, in high-criticality tasks, dehumanizing the higherhierarchy AI may emerge as a defensive mechanism against the status
loss of the human party due to expertise and agency gaps. In this sce­
nario, the human party activates a counter mechanism to protect their
own self-attributed humanness from being diminished by a supposedly
less humanlike entity (AI). Existing research shows that as a defense
mechanism, people may exhibit a greater inclination to dehumanize an
entity with a higher status in a domain where they themselves lack
expertise and agency (Haslam et al., 2008; Loughnan & Haslam, 2007;
Struch & Schwartz, 1989). People may subconsciously dehumanize
out-groups or dissimilar others who hold greater status to feel better
about themselves (Fein & Spencer, 1997), maintain their status in
comparison (Scaillet & Leyens, 2000), and preserve their positive
self-image (Branscombe & Wann, 1994; Crocker et al., 1987). The ten­
dency to ascribe less humanness to dissimilar others for these reasons is
defined as “infrahumanization” (Leyens et al., 2001). Therefore, it is
posited that in high-criticality tasks, low-hierarchy conversational AI
should be perceived as more humanlike due to people’s inclination to
infrahumanize high-hierarchy conversational AI.
Research also shows that a high level of humanlikeness in conver­
sational AI presentations, such as digital assistants, is crucial for
increasing purchase intentions (Balakrishnan & Dwivedi, 2024; Kim
et al., 2022) and willingness to use (Moussawi et al., 2021) as perceived
humanlikeness also positively influences perceived competence of AI in
conversational settings (Gursoy et al., 2019; Hu et al., 2021). It is
because the ability to maintain a coherent conversation is primarily a
human quality (Cheng et al., 2022) and is considered a manifestation of
competence when the current level of AI technologies is taken into ac­
count (Neururer et al., 2018). As discussed so far, humanlikeness in the
context of conversational AI also carries connotations of agency and
expertise, which are constructs of competence.
The existing research supports this notion showing that the perceived
humanlikeness of conversational agents increases their perceived
competence, facilitating positive attitudes toward the agent (Chandra
et al., 2022; Chi et al., 2024). For example, Yoganathan et al. (2021)
found that humanlike robots are perceived as more competent in service
roles than their non-humanlike counterparts, leading to higher evalua­
tions of service quality. Similarly, Borau et al. (2021) argued that more
humanlike conversational AI algorithms are perceived as more compe­
tent in customer-facing service duties due to their perceived ability to
understand unique customer needs. Kim and Hur (2023) demonstrated
that more competent AI is more likely to be used compared to less
5

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

competent counterparts. Other research also showed that the compe­
tence of AI is crucial for its acceptance in various contexts like human-AI
teaming (Harris-Watson et al., 2023), recommendation of AI service
robots (Chi et al., 2024), and user engagement with conversational AI
service providers (Chandra et al., 2022) (see Table 1).
Therefore, in high-criticality tasks, lower-hierarchy conversational
AI should be perceived as more humanlike and competent, leading to a
higher willingness to use it. In contrast, when the task criticality is not
high, higher-hierarchy conversational AI should be perceived as more
humanlike and competent, ultimately activating a greater willingness to
use it.
Formally.

(McLean et al., 2021), resulting in less perceived warmth from their
interactions (Yanit et al., 2023). Thus, AI’s perceived humanlikeness
may increase its perceived warmth, and both perceived humanlikeness
and warmth may mediate the link between the perceived hierarchical
status of conversational AI and people’s willingness to use in a serial
mediation relationship as an alternative explanation.
2.3.2. Perceived control and risks
Existing literature suggests that the perceived humanlikeness of
agents can influence people’s perceptions of risks of interacting with
them and their sense of control over these agents (Kim & McGill, 2011).
Since AI agents are often seen as black boxes (Castelvecchi, 2016)
lacking explicability (Tambe et al., 2019), it’s challenging to understand
exactly how AI generates its output. This uncertainty contributes to AI
being perceived as uncanny, affecting its adoption based on perceived
risks and people’s perceived control over them.
Belanche et al. (2019) found that risk-averse customers were less
willing to interact with machinelike robots compared to humanlike ro­
bots, indicating that low humanlikeness in those agents increases
perceived risks from interacting with them. Yan et al. (2022) discuss
how higher perceptions of control over AI systems increase the intention
to use those systems.
The effect of perceived control on conversational AI and the risks
posed by adopting AI-provided services may be particularly more potent
in high criticality tasks, as greater algorithm aversion is triggered in such
contexts. People may perceive less control over and greater risks from
higher-hierarchy AI due to its diminished humanlikeness in high criti­
cality tasks. On the other hand, as conversational AI becomes more
humanlike through increased perceptions of hierarchical status in low
criticality tasks, people may perceive fewer risks in interacting with it
and feel a greater sense of control. Thus, serial mediation relationships
through perceived humanlikeness, control, and risks may serve as
alternative explanations to the focal pathway through competence.

Hypothesis 2a. (H2a): When the task criticality is not high, high hi­
erarchy AI will be perceived as more humanlike and thus more
competent, and in turn, people’s willingness to use it will be higher.
Hypothesis 2b. (H2b): When the task criticality is high, low hierarchy
AI will be perceived as more humanlike and thus more competent, and
in turn, people’s willingness to use it will be higher.
2.3. Alternative explanations
This paper also aims to explore other alternative explanations that
may better explain the relationship between the perceived hierarchical
status of conversational AI and people’s willingness to use it. For
instance, if low-status job titles such as ’assistant’ activate a ’helper’
heuristic in high task criticality, this could impact the perceived warmth
of the conversational AI through perceived humanlikeness, subse­
quently influencing people’s willingness to use it. Similarly, low-status
job titles may alter the perceived controllability of the AI by the
human user and influence the perceived risks it poses due to its
perceived humanlikeness, each of which could be relevant variables
particularly in high task criticality.
2.3.1. Perceived hierarchy, humanlikeness and perceived warmth of the AI
In high task criticality, activating the “helper” heuristic with lowstatus job titles may increase the perceived humanlikeness of conver­
sational AI and impact its perceived warmth in conversational settings.
In such contexts, high-status AI is likely to be infrahumanized to the
level of an automaton or machine. Mechanical entities are perceived as
lacking empathy and emotionally inert (Haslam, 2006). For example, it
is stereotypically believed that high-status groups (e.g., businessmen)
are less humanlike compared to low-status counterparts, being seen as
competent but lacking warmth (Harris & Fiske, 2011, pp. 123–134).
Leyens et al. (2001) discuss how dehumanized entities are biasedly
perceived as incapable of experiencing complex emotions, which re­
duces their perceived warmth. Machine-like qualities tend to reduce the
warmth of conversational AI, increasing its uncanny features (Mori,
1970), whereas human-like qualities enhance it. Cheng et al. (2022)
argue that humanlike cues are essential for conversational AI to be
perceived as warm.
People form attitudes toward conversational AI based on its
perceived warmth, balancing the machine and human attributes of these
agents (Yanit et al., 2023). Research shows that the humanlikeness of
automated agents increases their perceived warmth (Kim et al., 2019),
improving people’s attitudes toward them (Baek et al., 2022; Christo­
forakos et al., 2021; Zhu & Chang, 2020). For instance, Ahn, Kim, and
Sung (2022) demonstrated that as AI’s perceived warmth increases,
people rely more on its recommendations.
As discussed previously, when the task criticality is not high, higher
status titles may increase AI’s perceived humanlikeness approximating
it to the level of a real human being. Adding on this notion, past research
has shown that perceptual approximation to real human beings en­
hances the warmth of AI (Kim et al., 2019). For instance, when an AI is
assigned a low-status title such as "assistant", service recipients are more
inclined to perceive the AI as a tool rather than a humanlike agent

3. Research design
This paper uses a mixed-method approach, incorporating one partial
least squares structural equation modeling (PLS-SEM) study (Study 1a)
and four experimental studies (Pretest, Studies 1b, 2, and 3) to test the
proposed conceptual framework (see Fig. 1). Adopting a multimethod
approach that combines PLS-SEM and experimental studies offers
several benefits, as this combination enhances the reliability of findings
(Kurtaliqi et al., 2024), improves predictive accuracy (Gudergan et al.,
2025; Richter & Tudoran, 2024), and provides greater external validity
(Richter et al., 2022). PLS-SEM has several advantages over other SEM
methods, as it is more flexible in terms of required sample size and can
handle single-item measurements (Cheah et al., 2018; Matheus et al.,
2023). While PLS-SEM enhances external validity, experimental studies
complement it by strengthening internal validity (Hair et al., 2021).
The structure of the research design in the current paper is as follows:
First, a pretest (N = 313) was conducted to categorize the scenarios used
in the following studies based on their task criticality. Then, Study 1a
(low criticality task 1) (N = 309) was conducted to examine the initial
model fit indices and assess measurement validity using PLS-SEM, where
the effect of the perceived hierarchical status of a home assistant
conversational AI on willingness to use through focal and alternative
pathways was analyzed to examine hypotheses 1a and 2a. Following the
PLS-SEM study, experimental studies were conducted to increase the
internal validity of the findings.
Study 1b (low criticality task 2) enhances the internal validity of the
initial findings from the PLS-SEM study through an experimental design.
This study (N = 307) experimentally tests the hypotheses 1a and 2a by
comparing a task assistant AI (low status) to a task manager AI (high
status). Previous studies have shown that "assistant" and "manager" titles
effectively manipulate the hierarchical status of conversational AI due to
their connotations with an employee’s position in the organizational
6

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Fig. 1. Proposed full model.
Note: Dotted lines indicate potential alternative paths. Although not shown on the diagram to reduce visual complexity, in studies 1a, 1b, 2, and 3, the model was
controlled for the direct effect of perceived hierarchical status on all the second-order mediators, as well as the direct effect of perceived humanlikeness on will­
ingness to use. In study 3, model is also controlled for perceived scenario realism.

chart (Jeon, 2022). In the context of the current paper, any job title that
conveys greater agency and expertise should create an adequate
disparity for people to place a conversational AI at a higher hierarchical
level than “assistant” title.
Similarly, Study 2 (N = 300) employs an experimental design to
examine the focal conceptual framework in a medium criticality task
while mitigating potential biases associated with the connotations of the
’manager’ title in the high-status condition. Existing research suggests
that people have stereotypical behavioral attributions toward managers,
defined as “managerial characteristics,” which include perceived
competence, trust, likability, and a lack of perceived self-interest (Pinto
et al., 2017). Such heuristic attributions of positive characteristics could
introduce biases favoring a conversational AI with a manager title. To
address this, Study 2 uses a different context (education) with different
role titles, comparing a teaching assistant AI (low status) to a teacher AI
(high status) to test hypotheses 1a and 2a.
Finally, Study 3 (high criticality) (N = 308) demonstrates a reversed
effect, where the assistant heuristic holds with positive effects as
intended by current marketing practices and confirms hypotheses 1b
and 2b. All studies control for alternative explanations and test the
mediating effect of perceived humanlikeness between perceived hier­
archical status and willingness to use. Additionally, study 3 controls for
scenario realism within the overall conceptual framework.
Studies 1b, 2, and 3 use Hayes’ PROCESS module in SPSS to conduct
path analyses (Hayes, 2012) and the results of study 1a has been
analyzed on SmartPLS 4.0. They all adopt between-subjects designs,
with main effects analyzed using independent samples t-tests or re­
gressions (Studies 1a and 3). Since the pretest involves comparing
multiple categories, a one-way between-subjects ANOVA and planned
contrasts were used. PROCESS was chosen for path analysis in this paper
because it provides a straightforward, regression-based approach for
testing mediation and moderation without requiring complex structural
modeling. Unlike SEM, which is more suitable for latent variable anal­
ysis, PROCESS is designed for observed variables and automatically
conducts bootstrapping to generate robust confidence intervals for in­
direct effects (Clement & Bradley-Garcia, 2022). Additionally, its ease of
implementation in SPSS, R, and SAS makes it a practical choice for
examining direct and indirect effects without the need for extensive
model specification or fit indices.
The studies were conducted on Prolific. Prolific was selected because
participant backgrounds (ethnicity, nationality, employment, etc.) on

this platform vary more widely than on other platforms such as MTurk.
Therefore, the results are less likely to be biased due to the imbalanced
representation of certain groups. More importantly, Douglas et al.
(2023) showed that the quality of the data collected on Prolific was
significantly higher than on other platforms, as Prolific participants
were “more likely to pass various attention checks, provide meaningful
answers, follow instructions, remember previously presented informa­
tion, have a unique IP address and geolocation, and work slowly enough
to be able to read all the items.”
Throughout the studies, the main effect was tested mostly using in­
dependent samples t-tests. G*Power analyses (Faul et al., 2007) indi­
cated that to achieve 99 % statistical power with a medium effect size
(Cohen’s d = 0.5) in an independent samples t-test, a minimum of 296
participants was required. Therefore, in each study, 300 participants
with at least a 95 % submission acceptance rate were recruited through
Prolific. However, except for Study 2, the final sample sizes exceeded
this number. According to Prolific, this can occur due to participants
timing out or returning their submissions, which affects the total
participant count as in some cases, participants complete the survey on
the survey platform but do not submit their participation on Prolific or
chose not to submit for various reasons (Prolific, 2024). Although their
participation was not recorded under "completed assignments," these
participants were identified using their participant IDs and were
compensated with a bonus payment.
Scale items can be found in Table 2. The studies were analyzed using
SPSS and SmartPLS 4.0. All data is available at https://doi.org
/10.6084/m9.figshare.28674704. Participants have not been asked
any demographic questions instead this information has been obtained
from participant’s self-reported information to Prolific to ensure survey
parsimony.
4. Pretest: categorizing scenarios based on task criticality
4.1. Overview and method
The pretest was conducted to categorize scenarios for use in subse­
quent studies based on their perceived task criticality. Tzafestas (2016),
had previously categorized task criticality into three levels: ’high,’
’medium,’ and ’low.’ This study, therefore, introduced different sce­
narios for conversational AI use, aiming to measure task criticality and
examine whether significant differences exist among them to categorize
7

Computers in Human Behavior Reports 18 (2025) 100682

M. Yanit

Table 2
Scale items.

Table 2 (continued )

Scales

Items

Adopted from

Perceived task
criticality (only in
pretest)

The task this AI performs
is highly personal.
The task this AI performs
is highly critical.
The task this AI performs
is highly private.
If the task this AI performs
is not perfromed
adequately, it may put the
user in a risky situation.
The task this AI performs
requires a personalized
touch.
In this question,
considering your
potential interaction with
this AI, we want you to
assess the status of the
given AI in the hierarchy
by moving the sliders
below. From left to right,
assigned social status
increases (0 = Lowest
Status, 7 = Highest
Status).

Measured using 7-point
likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Følstad et al. (2024).

Perceived hierarchical
status (also,
manipulation check
scale)

Willingness to use the
AI

Perceived
humanlikeness

Perceived warmth

I would be willing to use
this AI.
I would be willing to
adopt this AI to run
certain tasks.
I would be willing to at
least try this AI.

I think this AI is capable of
thinking.
I think this AI is able to
understand how others
are feeling.
I think this AI is very
similar to real human
beings.
I think this AI is thinking
and acting like a real
human being.
I think this AI is able to
tell what is right and what
is wrong.
This AI feels warm.
This AI feels sincere.
This AI feels kind.
This AI feels friendly.

Perceived competence

This AI feels competent.
This AI feels skillful.
This AI feels intelligent.
This AI feels capable.

Perceived risks

It feels risky to use this AI.
This AI feels like it would
work against me.
This AI feels like it does
not have good intentions.

Scales

Perceived control of AI

Items

Adopted from

This AI feels like it would
exploit me.
I could dominate this AI.
I could make this AI obey
me.
I could control this AI.
I could manage this AI.

Measured using 7-point
likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Kim and McGill (2011)

these scenarios into low, medium, and high task criticality categories.
The study was conducted on Qualtrics, with participants recruited
from Prolific. Qualtrics’ AI tools predicted that the study would take less
than a minute to complete, as it included only a single 5-item scale
measuring perceived task criticality (see Table 2). To enhance flexibility,
participation rewards were calculated for a duration of 4 min. Prolific’s
fair payment policy recommends a minimum of £0.60 GBP for a 4-min
study (Prolific, 2025). Therefore, a total of 313 participants (Mage =
35.10, 52 % female) took part in the study in exchange for £0.65 GBP. At
the end of the study, the median completion time was 1 min and 1 s.
Table 3 shows the demographic information for these participants.
Scenario introductions that participants had read before answering the
scale questions, along with the real-life applications from which they
were inspired are given in Table 4.
This pretest used a between-subjects design. Participants who
confirmed the consent form joined the study and were randomly
assigned to one of four scenario conditions. After being introduced to the
scenarios, participants were asked to respond to scale items measuring
the perceived criticality of the task. The scale items included: ’The task
this AI performs is highly personal,’ ’The task this AI performs is highly
critical,’ ’The task this AI performs is highly private,’ ’If the task this AI
performs is not completed adequately, it may put the user in a risky
situation,’ and ’The task this AI performs requires a personalized touch.’
(MCriticality = 4.82, SDCriticality = 1.08, Cronbach’s alpha = 0.73) (Følstad
et al., 2024).

Measured using 7-point
Likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Kraus and Keltner (2013);
Kraus et al. (2009). A slider
is used to measure perceived
hierarchical status because
hierarchies are typically
mentally visualized in a
tiered structure, akin to a
"social ladder" (Demczuk
et al., 2023), and a slider
aligns with this cognitive
mapping.
Measured using 7-point
Likert scale. The first item
has been adapted from
Yanit et al. (2023), while the
remaining items were added
to strengthen the scale’s
association with usage
behavior. Consequently,
word-of-mouth
(WOM)-related items from
the original scale have been
replaced to better align with
usage-related items the
focus on actual engagement
with conversational AI and
its usage behaviors.
Measured using 7-point
likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Touré-Tillery and McGill
(2015)

4.2. Results
To analyze the results, a one-way between-subjects ANOVA and
planned contrasts were conducted. First, the results of the one-way
ANOVA were significant (MScenario1 = 4.28, SDScenario1 = 1.06, MSce­
nario2 = 4.54, SDScenario2 = 1.03, MScenario3 = 4.98, SDScenario3 = 0.88,
MScenario4 = 5.44, SDScenario4 = 0.99, F(3, 309) = 20.66, p < 0.001, Etasquared η2 = 0.17). Additionally, planned contrast analyses revealed
that Scenario 4 (credit score management task) was perceived as
significantly more critical than the other scenarios (Contrast value =
Table 3
Demographics in pretest.
Item
Gender
Female
Male
Prefer not to say
Age
18-30
31-43
44-56
57-69
70-82
No Information
Ethnicity
Asian
Black
Mixed
White
Other
No Information

Measured using 7-point
likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Horchak et al. (2016)
Measured using 7-point
likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Horchak et al. (2016)
Measured using 7-point
likert scale (1 = Strongly
disagree, 7 = Strongly
agree) and adapted from
Kim and McGill (2011)

8

Count

Percentage

162
137
14

52 %
44 %
4%

128
110
39
18
2
16

41 %
35 %
12 %
6%
1%
5%

41
68
24
149
15
16

13 %
22 %
8%
47 %
5%
5%

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Table 4
Scenario introductions in pretest.

Table 5
Demographics in study 1a.

#

Scenario

Context

Real Life
Applications

1

Please try to imagine yourself
in the following situation: “An
AI (Artificial Intelligence)based software is assisting you
with tasks in a conversational
manner, similar to Google
Home and Amazon Alexa.”
Please try to imagine yourself
in the following situation: “An
AI solves various work
problems for you in a
conversational manner. For
example, it sends emails on
your behalf based on the
instructions you provide.”
Please try to imagine yourself
in the following situation: “An
AI is utilizing machine learning
and data analysis techniques to
assist you with your courses
and exam preparations. This AI
provides one-on-one tutoring
for students and promptly
addresses their questions
directly in a conversational
manner. Additionally, it
involves parents in discussions
on ways to enhance students’
performance.”
Please try to imagine yourself
in the following situation: “An
AI is utilizing machine learning
and data analysis techniques to
provide you with personalized
ways to enhance credit scores.
This AI provides one-on-one
services for the users and
makes decisions based on the
input it collects from them in a
conversational manner.”

Conversational
home assistants

https://home.goo
gle.com/https
://alexa.amazon.ca/

2

3

4

Task management
for work

Teaching/tutoring

Item
Gender
Female
Male
Prefer not to say
No information
Age
18-30
31-43
44-56
57-69
No information
Ethnicity
Asian
Black
Mixed
White
Other
No Information

https://copilot.
microsoft.com/

https://professorai.
ai/

Count

Percentage

160
139
1
9

52 %
45 %
0%
3%

186
80
27
7
9

60 %
26 %
9%
2%
3%

11
59
24
188
17
1

4%
19 %
8%
61 %
6%
3%

5.2. Method

Credit Score
Management

After agreeing to the consent form, all participants were presented
with the same information about a new hypothetical AI tool called
peanut.ai. The information stated, “Peanut.ai is an AI (Artificial
Intelligence)-based software that can assist users with tasks in a
conversational manner, similar to Google Home and Amazon Alexa”.
Following this information, participants proceeded to answer the
survey questions. The scale questions have been given in Table 2.
Descriptive statistics for the scale questions in Study 1a can be found in
Table 6.
To analyze the results, partial least squares structural equation
modeling (PLS-SEM) was employed. In selecting Partial Least Squares
Structural Equation Modeling (PLS-SEM) over Covariance-Based Struc­
tural Equation Modeling (CB-SEM), several considerations were pivotal.
PLS-SEM is particularly advantageous when the research objective em­
phasizes prediction and explanation of target constructs, as it focuses on
maximizing the explained variance (R2 values) of endogenous variables
(Hair et al., 2021). This approach is also well-suited for studies with
smaller sample sizes and non-normal data distributions, as it imposes
fewer restrictions compared to CB-SEM. Additionally, PLS-SEM effec­
tively handles complex models, including those with formatively
measured constructs and single-item measures—such as the indepen­
dent variable in this paper—without encountering the identification
issues that often arise in CB-SEM. While CB-SEM is traditionally asso­
ciated with theory testing and provides comprehensive model fit indices,
PLS-SEM’s evolving fit measures, such as the Standardized Root Mean
Square Residual (SRMR), offer relevant assessments of model quality
(Dash & Paul, 2021). Given these factors and considering the explor­
atory nature of this study with all previously untested pathways,
PLS-SEM was deemed more appropriate for this study’s objectives and
data characteristics. The data were analyzed using SmartPLS 4.0
software.

https://www.knapsa
ck.ai/

2.53, Std. Error = 0.39, t(309) = 6.51, p < 0.001). For the remaining
scenarios, Scenario 3 (teaching task) was considered more critical than
Scenario 1 (home assistance) and Scenario 2 (task management)
(Contrast value = 1.15, Std. Error = 0.26, t(309) = 4.45, p < 0.001).
However, there was no significant difference between Scenario 1 and
Scenario 2 in terms of perceived task criticality (Contrast value = 0.26,
Std. Error = 0.17, t(309) = 1.56, p = 0.12).
Based on these results, Study 1a and Study 1b used "low" criticality
tasks—home assistance and task management, respectively—as scenario
manipulations. Study 2 employed the teaching scenario as a "medium"
criticality task for the same purpose, while Study 3 investigated the
credit management scenario as a "high" criticality task. In the next sec­
tion, the focal conceptual framework is examined by gradually
increasing task criticality through studies, and the hypotheses are tested
across different levels of criticality.

5.3. Results

5. Study 1a: home assistance – low criticality task 1

5.3.1. Measurement model
Following the procedure applied by Matheus et al. (2023), three
types of validity were tested: content validity with Cronbach’s alpha
scores (Straub et al., 2004), discriminant validity using
heterotrait-monotrait (HTMT) ratios per pair of components, and cross
loadings (Matheus et al., 2023) convergent validity using composite
reliability (CR) and average variance extracted (AVE). Cronbach’s alpha
values exceeded the threshold of 0.7, ensuring content validity
(Bernardi, 1994) (see Table 6). All AVE scores were above the threshold
value of 0.5 (Fornell & Larcker, 1981), and all CR values were above the

5.1. Overview
Study 1a has been predicted to take 3 min to complete by Qualtrics.
Therefore, 309 participants (Mage = 30.63, 52 % female) recruited from
Prolific in exchange for 0.65 GBP. At the end of the study, the median
completion time was 2 min 57 s. Detailed demographics of these par­
ticipants can be found in Table 5. The data was collected through a
survey on Qualtrics, and no data was eliminated from the dataset as
there was no incomplete responses.
9

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Table 6
Construct reliability and validity.
Perceived humanlikeness
Perceived warmth
Perceived competence
Perceived control
Perceived risks
Willingness to use

M (SD)

Cronbach’s alpha

Composite reliability (CR)

Average variance extracted (AVE)

3.18 (1.43)
4.06 (1.25)
4.86 (1.27)
4.38 (1.14)
3.27 (1.24)
5.15 (1.31)

0.92
0.94
0.94
0.85
0.90
0.93

0.92
0.94
0.94
0.89
0.91
0.94

0.75
0.84
0.84
0.69
0.78
0.88

suggested threshold of 0.7 (Chin, 1998) (see Table 10). These results
indicated satisfactory content and convergent validities.
Cross loading analyses showed that each item was strongly corre­
lated with its corresponding construct (Matheus et al., 2023). The dif­
ferences between the closest cross-loading values were greater than 0.1
for each item, which eliminated cross-loading related concerns (Jamali,
Ayatollahi, & Jafari, 2018). (see Table 7). Moreover, an HTMT value
below 0.9 is necessary for each pair of variables to ensure discriminant
validity (Henseler et al., 2015). Some researchers suggest that HTMT
values within the 0.85–0.90 range may indicate potential concerns
regarding discriminant validity (Henseler et al., 2015). In this study, all
the HTMT values were below 0.85 with a maximum value of 0.67, which
indicated a satisfactory discriminant validity (see Table 8).
Herman’s single-factor test on SPSS revealed that the percentage of
total variance explained by a single factor was 36.2 %, supporting the
absence of common method bias. A maximum Variance Inflation Factor
(VIF) of 2.12, which is well below the most conservative threshold of 3
(O’Brien, 2007), further supports the absence of bias.

Table 8
HTMT values.

5.3.2. Structural model
The analyses of model fit indices showed that standardized root
mean squared residual (SRMR) was 0.06. When sample size is greater
than 100, an SRMR ≤0.08 indicates an acceptable fit (Cho et al., 2020).
Therefore, model fit was ensured.
5.3.3. Main effect
Linear regression analyses showed a significant and positive main
effect of perceived hierarchy of conversational AI on people’s willing­
ness to use (β = 0.17, SE = 0.03, p < 0.001). This result further sup­
ported hypothesis 1a (H1a).

Relationship

Heterotrait-monotrait ratio
(HTMT)

Perceived competence <− > Perceived hierarchy
Perceived control <− > Perceived hierarchy
Perceived control <− > Perceived competence
Perceived humanlikeness <− > Perceived
hierarchy
Perceived humanlikeness <− > Perceived
competence
Perceived humanlikeness <− > Perceived control
Perceived risks <− > Perceived hierarchy
Perceived risks <− > Perceived competence
Perceived risks <− > Perceived control
Perceived risks <− > Perceived humanlikeness
Willingness to use <− > Perceived hierarchy
Willingness to use <− > Perceived competence
Willingness to use <− > Perceived control
Willingness to use <− > Perceived
humanlikeness
Willingness to use <− > Perceived risks
Perceived warmth <− > Perceived hierarchy
Perceived warmth <− > Perceived competence
Perceived warmth <− > Perceived control
Perceived warmth <− > Perceived
humanlikeness
Perceived warmth <− > Perceived risks
Perceived warmth <− > Willingness to use

0.368
0.146
0.142
0.362
0.620
0.111
0.069
0.354
0.284
0.114
0.283
0.601
0.244
0.406
0.455
0.356
0.666
0.224
0.649
0.360
0.560

5.3.4. Analyses of the full model
To test the path coefficients, the bootstrapping re-sampling method
with 500 re-samples at a 95 % confidence interval was used. The results

Table 7
Cross loadings.
PH1
humanlike1
humanlike2
humanlike3
humanlike4
humanlike5
warm1
warm2
warm3
warm4
comp1
comp2
comp3
comp4
control1
control2
control3
control4
risk1
risk2
risk3
risk4
use1
use2
use3

Perceived hierarchy

Perceived humanlikeness

Perceived warmth

Perceived competence

Perceived control

Perceived risk

Willingness to use

1.000
0.301
0.285
0.273
0.356
0.282
0.246
0.352
0.322
0.342
0.322
0.339
0.364
0.286
0.108
0.127
0.080
0.136
− 0.049
− 0.003
− 0.120
− 0.059
0.264
0.286
0.219

0.347
0.837
0.898
0.877
0.883
0.827
0.513
0.600
0.564
0.530
0.518
0.491
0.572
0.536
0.128
0.006
− 0.109
− 0.030
− 0.082
0.035
− 0.113
− 0.116
0.355
0.410
0.293

0.346
0.489
0.514
0.521
0.552
0.524
0.892
0.910
0.939
0.926
0.587
0.568
0.540
0.602
0.250
0.154
0.075
0.190
− 0.299
− 0.236
− 0.324
− 0.307
0.497
0.521
0.457

0.357
0.531
0.489
0.435
0.500
0.530
0.540
0.599
0.546
0.604
0.899
0.929
0.911
0.934
0.085
0.113
0.030
0.179
− 0.270
− 0.269
− 0.331
− 0.276
0.544
0.548
0.489

0.138
− 0.080
− 0.014
− 0.029
0.030
0.034
0.157
0.178
0.180
0.208
0.186
0.123
0.064
0.110
0.751
0.844
0.878
0.847
− 0.230
− 0.199
− 0.269
− 0.232
0.223
0.239
0.171

− 0.070
0.001
− 0.047
− 0.086
− 0.098
− 0.132
− 0.288
− 0.280
− 0.319
− 0.340
− 0.364
− 0.327
− 0.193
− 0.318
− 0.144
− 0.130
− 0.249
− 0.316
0.848
0.883
0.910
0.879
− 0.396
− 0.381
− 0.405

0.274
0.330
0.304
0.299
0.334
0.361
0.458
0.450
0.480
0.534
0.538
0.540
0.484
0.505
0.157
0.168
0.160
0.244
− 0.391
− 0.330
− 0.394
− 0.353
0.956
0.945
0.912

10

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

of the path analyses are presented in Table 9. The findings indicated that
serial mediation pathways through perceived humanlikeness and
warmth, as well as through perceived humanlikeness and competence,
were both positive and significant. As controlled pathways, mediating
effects of perceived warmth and perceived competence alone were also
significant. This result suggests that people attributed warmth and
competence to the AI without requiring it to have humanlike qualities.
This may be because the AI was positioned as similar to Google Home
and Alexa in the manipulation condition, leading to a preobtained
perception of humanlikeness based on people’s familiarity and experi­
ences with these real AI assistants. This suggestion requires further
investigation.
The serial mediations through perceived risks and perceived control
as well as the indirect effect through perceived humanlikeness alone,
and the direct effect of perceived hierarchical status on willingness to
use the AI were found to be nonsignificant (see Table 9). This result
indicated full mediations through observed statistically significant in­
direct effects. In the context of this paper’s focal mechanism, this in­
dicates a positive and significant relationship between the perceived
hierarchical status of conversational AI and perceived humanlikeness.
As the perceived humanlikeness of the AI increases, perceived compe­
tence also rises, ultimately enhancing people’s willingness to use the AIbased service. Therefore, hypothesis 2a (H2a) was supported.

study 1a with an experimental design. A total of 300 participants were
requested on the Prolific platform for 0.65 GBP for their participation. At
the end of the study, 307 participants took part (Mage = 28.9, 56 % fe­
male) and the median completion time was 3 min 38 s. Detailed de­
mographic analyses are provided in Table 10. Participants accessed the
questionnaire through the Qualtrics web platform, and no data were
excluded from the dataset.
6.2. Method
To manipulate the perceived hierarchical status of the AI, partici­
pants were randomly assigned to different conditions. Accordingly,
participants were presented with the following information based on
their assigned condition, where the hierarchical status of the conversa­
tional AI was manipulated through the assigned job titles of either “task
assistant” or “task manager.” For the low-status AI: "Nina is your
everyday AI-based task assistant that solves various problems for you in
a conversational manner. Below, you will see an interface of Nina", and:
"Nina is your everyday AI-based task manager that solves various
problems for you in a conversational manner. Below, you will see an
interface of Nina" for the high-status AI condition. After receiving the
information, participants were presented with an AI interface directly
extracted as a screenshot from the Microsoft Copilot interface on Win­
dows OS, as depicted in Appendix A. The image of the interface was
taken as a screenshot from Copilot to enhance scenario realism. The
account name used to capture the screenshot was concealed in the
greeting message of Copilot after the screenshot was taken. Participants
saw how the AI greets a user and were shown the recommended quick
prompts that could be inquired, such as “Write an ‘out of office’ email
reply while I am on vacation.” These prompts were designed to help
participants understand the tasks the AI could assist with. As a result,
participants did not interact with the AI directly but only observed its
interface on an image.
Next, participants proceeded to answer the survey questions given in
Table 2. Descriptive statistics for the scale questions can be found in
Table 11.

6. Study 1b: task assistant versus task manager - low criticality
task 2
6.1. Overview
The aim of study 1b was to gain internal validity to the findings of
Table 9
Direct and indirect effects in the full model.

Perceived hierarchy →Perceived humanlikeness
Perceived hierarchy → Perceived warmth
Perceived hierarchy → Perceived competence
Perceived hierarchy → Perceived control
Perceived hierarchy → Perceived risks
Perceived humanlikeness → Perceived warmth
Perceived humanlikeness → Perceived
competence
Perceived hierarchy → Willingness to use
Perceived humanlikeness → Willingness to use
Perceived warmth → Willingness to use
Perceived competence → Willingness to use
Perceived control → Willingness to use
Perceived risk → Willingness to use
Perceived humanlikeness → Perceived control
Perceived humanlikeness → Perceived risk
Perceived hierarchy → Willingness to use
Perceived hierarchy → Perceived humanlikeness
→ Willingness to use
Perceived hierarchy → Perceived warmth →
Willingness to use
Perceived hierarchy → Perceived competence →
Willingness to use
Perceived hierarchy → Perceived control →
Willingness to use
Perceived hierarchy → Perceived risks →
Willingness to use
Perceived hierarchy → Perceived humanlikeness
→ Perceived warmth → Willingness to use
Perceived hierarchy → Perceived humanlikeness
→ Perceived competence → Willingness to use
Perceived hierarchy → Perceived humanlikeness
→ Perceived control → Willingness to use
Perceived hierarchy → Perceived humanlikeness
→ Perceived risk → Willingness to use

Estimate

Lower BootCI
Upper BootCI

0.35*
0.16*
0.18*
0.16*
− 0.05
0.55*
0.52*

0.24
0.06
0.1
0.04
− 0.16
0.45
0.43

0.45
0.25
0.28
0.28
0.07
0.64
0.59

0.05
0.04
0.19*
0.31*
0.08
− 0.22*
− 0.07
− 0.07
0.05
0.02

− 0.05
− 0.08
0.05
0.17
− 0.02
− 0.34
− 0.23
− 0.20
− 0.05
− 0.03

0.15
0.17
0.34
0.46
0.18
− 0.1
0.09
0.06
0.15
0.06

0.03*

0.004

0.07

0.06*

0.02

0.10

0.01

− 0.003

0.04

0.01

− 0.02

0.04

0.04*

0.01

0.07

0.06*

0.03

0.09

− 0.002

− 0.01

0.003

0.005

− 0.004

0.02

6.3. Results
6.3.1. Manipulation check
Results of an independent samples t-test revealed that the manipu­
lation was successful (MLow = 2.95, SDLow = 2.05, MHigh = 4.39, SDHigh
= 1.59, t(305) = 6.87, p < 0.001, d = 0.8) as people in the high-status AI
condition perceived AI to be significantly higher in the hierarchy than
people in the low-status AI condition.
Table 10
Demographics in study 1b.
Item
Gender
Female
Male
No Information
Age
18-30
31-43
44-56
57-69
70+
Ethnicity
Asian
Black
Mixed
White
Other
No Information

Note: An asterisk (*) denotes statistically significant relationships.
11

Count

Percentage

171
133
3

56 %
43 %
1%

209
72
21
3
2

68 %
23 %
7%
1%
1%

16
82
28
156
22
3

5%
27 %
9%
51 %
7%
1%

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Table 11
Descriptive statistics in study 1b.

Table 12
Direct and indirect effects in the full model.

Variable

Mean

St. Dev.

Cronbach’s Alpha

Examined effects in the full model

Results

Willingness to use the AI
Perceived humanlikeness
Perceived warmth
Perceived competence
Perceived risks
Perceived control

5.30
3.16
4.49
5.07
2.83
4.66

1.29
1.39
1.29
1.16
1.26
1.13

0.92
0.91
0.94
0.92
0.92
0.86

Perceived hierarchy > Perceived humanlikeness

β* = 0.46, SE = 0.16, 95 % CI =
[0.15, 0.77]
β* = 0.47, SE = 0.05, 95 % CI =
[0.38, 0.56]
β* = 0.40, SE = 0.04, 95 % CI =
[0.31, 0.48]
β = 0.05, SE = 0.05, 95 % CI =
[-0.05, 0.15]
β = 0.03, SE = 0.05, 95 % CI =
[-0.06, 0.12]
β* = 0.12, SE = 0.05, 95 % CI =
[0.03, 0.21]
β = − 0.02, SE = 0.11, 95 % CI =
[-0.23, 0.19]
β = 0.25, SE = 0.13, 95 % CI =
[-0.001, 0.50]
β = 0.19, SE = 0.12, 95 % CI =
[-0.04, 0.42]
β = − 0.17, SE = 0.15, 95 % CI =
[-0.46, 0.12]
β = − 0.02, SE = 0.13, 95 % CI =
[-0.28, 0.23]
β* = 0.19, SE = 0.06, 95 % CI =
[0.08, 0.30]
β* = 0.43, SE = 0.06, 95 % CI =
[0.31, 0.55]
β = 0.001, SE = 0.05, 95 % CI =
[-0.09, 0.09]
β* = − 0.27, SE = 0.04, v CI =
[-0.36, − 0.18]
β* = 0.05, BootSE = 0.03, 95 %
BootCI = [0.01, 0.12]
β = 0.05, BootSE = 0.03, 95 %
BootCI = [-0.001, 0.12]
β = 0.08, BootSE = 0.06, 95 %
BootCI = [-0.02, 0.20]
β = 0.00001, BootSE = 0.01, 95
% BootCI = [-0.02, 0.01]
β = 0.05, BootSE = 0.04, 95 %
BootCI = [-0.03, 0.13]
β * = 0.04, BootSE = 0.02, 95 %
BootCI = [0.01, 0.09]
β* = 0.08, BootSE = 0.03, 95 %
BootCI = [0.02, 0.16]
β = − 0.01, BootSE = 0.01, 95 %
BootCI = [-0.02, 0.01]
β = 0.00001, BootSE = 0.002, 95
% BootCI = [-0.004, 0.003]

Perceived humanlikeness > Perceived warmth
Perceived humanlikeness > Perceived
competence
Perceived humanlikeness > Perceived risks

6.3.2. Main effect
The main effect analysis, conducted via an independent samples ttest, indicated that participants in the high-status AI condition were
significantly more willing to use the AI (MLow = 5.14, SDLow = 1.36,
MHigh = 5.46, SDHigh = 1.19, t(305) = 2.20, p = 0.03, d = 0.25)
compared to those in the low-status AI condition. Thus, hypothesis 1a
(H1a) was supported one more time.

Perceived humanlikeness > Perceived control

6.3.3. Analyses of the full model
A custom model was built using PROCESS syntax to analyze the full
model. The indirect effect through perceived humanlikeness was sig­
nificant (β = 0.05, BootSE = 0.03, BootCI = [0.01, 0.12]). The findings
also revealed significant indirect effects of the serial mediations through
perceived humanlikeness and warmth (β = 0.04, BootSE = 0.02, BootCI
= [0.01, 0.09]), as well as perceived humanlikeness and competence (β
= 0.08, BootSE = 0.03, 95 % BootCI = [0.02, 0.16]). Additionally, the
direct effect of perceived hierarchical status became nonsignificant (β =
− 0.02, SE = 0.11, p = 0.84, CI = [− 0.23, 0.19]). Therefore, it was
concluded that perceived humanlikeness fully mediated the relationship
between perceived hierarchical status and willingness to use. There
were full serial mediations through humanlikeness and warmth, as well
as through humanlikeness and competence. Increased perceptions of
hierarchical status also elevated perceived humanlikeness, which in turn
increased people’s willingness to use the AI through heightened per­
ceptions of competence and warmth of the AI. These results showed that,
as expected in a low-criticality task, higher-hierarchy conversational AI
is perceived as more humanlike and more competent. Consequently,
people’s willingness to use it becomes higher. Therefore, Hypothesis 2a
was supported again.
The indirect effects of serial mediations through humanlikeness and
perceived risks (β = − 0.01, BootSE = 0.01, BootCI = [− 0.02, 0.01]) and
through humanlikeness and perceived control (β = 0.00001, BootSE =
0.002, BootCI = [− 0.004, 0.003]) did not reach significance (see
Table 12).

Perceived hierarchy > Perceived risks

Perceived humanlikeness > Willingness to use
Perceived hierarchy > Willingness to use
Perceived hierarchy > Perceived warmth
Perceived hierarchy > Perceived competence

Perceived hierarchy > Perceived control
Perceived warmth > Willingness to use
Perceived competence > Willingness to use
Perceived control > Willingness to use
Perceived risks > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Willingness to use
Perceived hierarchy > Perceived warmth >
Willingness to use
Perceived hierarchy > Perceived competence >
Willingness to use
Perceived hierarchy > Perceived control >
Willingness to use
Perceived hierarchy > Perceived risks >
Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived warmth > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived competence > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived risks > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived control > Willingness to use

Note: An asterisk (*) denotes statistically significant relationships.

6.3.4. Common method bias
Herman’s single-factor test on SPSS showed that common method
bias was not a concern for these analyses, as the percentage of total
variance explained by a single factor was 34.6 %, which is less than the
threshold value of 50 %. The results also showed that the maximum vif
was 1.95, which is well below the most conservative threshold of 3.
Therefore, the possibility of method bias is eliminated.

Table 13
Demographics in study 2.
Item
Gender
Female
Male
Age
18-30
31-43
44-56
57-69
Ethnicity
Asian
Black
Mixed
White
Other
No Information

7. Study 2: teaching assistant versus teacher – medium
criticality task
7.1. Overview
A total of 300 participants were requested on the Prolific platform for
0.65 GBP for their participation (Mage = 30.9, 54 % female). The median
completion time for this study was 3 min and 24 s. Detailed de­
mographic analyses are provided in Table 13. As in Study 1a and 1b,
participants accessed the questionnaire through the Qualtrics web
platform, and no data were excluded from the dataset.

12

Count

Percentage

161
139

54 %
46 %

195
69
27
9

65 %
23 %
9%
3%

11
59
24
188
17
1

4%
20 %
8%
63 %
6%
0%

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

7.2. Method

Additionally, the direct effect of perceived hierarchical status
became nonsignificant (β = 0.09, SE = 0.12, p = 0.47, CI = [− 0.15,
0.33]). Therefore, the observed significant indirect effects indicated full
serial mediations, supporting hypothesis 2a (H2a). Increased percep­
tions of hierarchical status also elevated perceived humanlikeness,
which, in turn, engendered greater perceptions of competence (H2a),
reduced risks, and elevated perceived control on the AI, respectively,
leading to greater willingness to use the AI.
However, the indirect effect through perceived humanlikeness alone
(β = 0.05, BootSE = 0.04, BootCI = [− 0.003, 0.14]) and the serial in­
direct effect through perceived humanlikeness and warmth (β = 0.01,
BootSE = 0.03, BootCI = [− 0.04, 0.07]) did not reach significance (see
Table 15). This result has been discussed in the limitations section.

The script for this study was adapted from the website of a real AI
software, Professor AI (PAI) (professorai.ai), which provides course
assistance to students and educational guidance to parents. PAI was
specifically chosen to ensure that a broad range of participants could
relate to the scenario, whether as students themselves or as parents, as
the scenario is relevant to both groups. Participants were randomly
assigned to different conditions and were presented with one of the
following pieces of information. For the low-status AI condition, people
read: "PAI is an artificial intelligence (AI)-based teaching assistant, uti­
lizing machine learning and data analysis techniques to assist students
with their courses and exam preparations. PAI provides one-on-one
tutoring for students and promptly addresses their questions directly
in a conversational manner. Additionally, it involves parents in discus­
sions on ways to enhance students’ performance". For the high-status AI
condition, people read: "PAI is an artificial intelligence (AI)-based
teacher, utilizing machine learning and data analysis techniques to assist
students with their courses and exam preparations. PAI provides one-onone tutoring for students and promptly addresses their questions directly
in a conversational manner. Additionally, it involves parents in discus­
sions on ways to enhance students’ performance".
After receiving the information, participants were presented with a
screenshot of PAI’s marketing communications from its website. These
communications were adapted according to the corresponding condition
and are presented in Appendix B.
Next, participants proceeded to answer the survey questions given in
Table 2. Descriptive statistics for the scale questions in Study 2 can be
found in Table 14.

7.3.4. Common method bias
Herman’s single-factor test on SPSS revealed that the percentage of
total variance explained by a single factor was 37.5 %, which is less than
the threshold value of 50 %. Similarly, vif analyses showed that the
highest vif factor was 2.15, which is below the threshold of 3. Therefore,
the possibility of a common method bias has been eliminated.
Table 15
Direct and indirect effects in the full model.

Perceived humanlikeness > Willingness to use
Perceived hierarchy > Willingness to use
Perceived hierarchy > Perceived warmth
Perceived hierarchy > Perceived competence
Perceived hierarchy > Perceived risks
Perceived hierarchy > Perceived control
Perceived warmth > Willingness to use
Perceived competence > Willingness to use
Perceived control > Willingness to use

7.3.3. Analyses of the full model
The same custom model from Study 1b was used to analyze the full
model using PROCESS syntax. The findings revealed significant indirect
effects of the serial mediations through perceived humanlikeness and
competence (β = 0.11, BootSE = 0.04, BootCI = [0.03, 0.20]), perceived
humanlikeness and control (β = 0.01, BootSE = 0.01, BootCI = [0.0001,
0.02]), and perceived humanlikeness and risks (β = 0.01, BootSE = 0.01,
BootCI = [0.001, 0.03]).

Perceived risks > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Willingness to use
Perceived hierarchy > Perceived warmth >
Willingness to use
Perceived hierarchy > Perceived competence >
Willingness to use
Perceived hierarchy > Perceived control >
Willingness to use
Perceived hierarchy > Perceived risks >
Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived warmth > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived competence > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived risks > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived control > Willingness to use

Table 14
Descriptive statistics in study 2.
St. Dev.

Cronbach’s Alpha

5.37
3.29
4.56
5.20
2.66
4.34

1.31
1.39
1.34
1.15
1.17
1.13

0.93
0.92
0.95
0.92
0.89
0.87

β* = 0.45, SE = 0.16, 95 % CI =
[0.14, 0.77]
β* = 0.61, SE = 0.04, 95 % CI =
[0.52, 0.69]
β* = 0.47, SE = 0.04, 95 % CI =
[0.39, 0.55]
β* = − 0.14, SE = 0.05, 95 % CI
= [-0.24, − 0.05]
β* = 0.13, SE = 0.05, 95 % CI =
[0.04, 0.22]
β = 0.11, SE = 0.06, 95 % CI =
[-0.01, 0.23]
β = 0.09, SE = 0.12, 95 % CI =
[-0.15, 0.33]
β = 0.14, SE = 0.12, 95 % CI =
[-0.10, 0.38]
β = 0.18, SE = 0.11, 95 % CI =
[-0.03, 0.40]
β = 0.20, SE = 0.14, 95 % CI =
[-0.07, 0.47]
β = − 0.14, SE = 0.13, 95 % CI
= [-0.39, 0.12]
β = 0.03, SE = 0.07, 95 % CI =
[-0.10, 0.16]
β* = 0.51, SE = 0.07, 95 % CI =
[0.37, 0.65]
β* = 0.13, SE = 0.05, 95 % CI =
[0.02, 0.23]
β* = − 0.15, SE = 0.06, 95 % CI
= [-0.26, − 0.04]
β = 0.05, BootSE = 0.04, 95 %
BootCI = [-0.003, 0.14]
β = 0.04, BootSE = 0.02, 95 %
BootCI = [-0.02, 0.05]
β = 0.09, BootSE = 0.06, 95 %
BootCI = [-0.01, 0.21]
β = − 0.02, BootSE = 0.01, 95 %
BootCI = [-0.07, 0.01]
β = − 0.03, BootSE = 0.03, 95 %
BootCI = [-0.1, 0.01]
β = 0.01, BootSE = 0.03, 95 %
BootCI = [-0.04, 0.07]
β* = 0.11, BootSE = 0.04, 95 %
BootCI = [0.03, 0.20]
β* = 0.01, BootSE = 0.01, 95 %
BootCI = [0.001, 0.03]
β* = 0.01, BootSE = 0.01, 95 %
BootCI = [0.0001, 0.02]

Perceived humanlikeness > Perceived control

7.3.2. Main effect
An independent samples t-test indicated that participants in the highstatus AI condition were significantly more willing to use the AI (MLow =
5.22, SDLow = 1.40, MHigh = 5.54, SDHigh = 1.18, t(298) = 2.14, p =
0.03, d = 0.25) in comparison to the people in the low-status AI con­
dition. Thus, hypothesis 1a (H1a) was supported one more time.

Mean

Perceived hierarchy > Perceived humanlikeness

Perceived humanlikeness > Perceived
competence
Perceived humanlikeness > Perceived risks

7.3.1. Manipulation check
To test the success of the manipulation, the same method from Study
1b was used (M = 3.46, SD = 1.78). Results of an independent samples ttest revealed that the manipulation was successful (MLow = 3.12, SDLow
= 1.82, MHigh = 3.83, SDHigh = 1.7, t(298) = 3.49, p = 0.001, d = 0.41),
as people in the high-status AI condition perceived AI to be significantly
higher in the hierarchy than people in the low-status AI condition.

Willingness to use the AI
Perceived humanlikeness
Perceived warmth
Perceived competence
Perceived risks
Perceived control

Results

Perceived humanlikeness > Perceived warmth

7.3. Results

Variable

Examined effects in the full model

Note: An asterisk (*) denotes statistically significant relationships.
13

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

8. Study 3: credit assistant versus credit analyst – high criticality
task

Table 17
Descriptive statistics in study 1b.

8.1. Overview
The study is predicted by Qualtrics to take 3 min. Therefore, 308
participants took part in the study in exchange for 1 GBP (Mage = 37.02,
52 % female). The median completion time was 3 min 46 s. Detailed
demographic analyses are provided in Table 16. Participants accessed
the questionnaire through the Qualtrics web platform, and no data were
excluded from the dataset.

After accepting the consent form, participants were randomly
assigned to one of two conditions in a between-subjects design. The
hierarchical status of the conversational AI was manipulated through
assigned job titles, either "credit assistant" (low-status AI) or "credit
analyst" (high-status AI). The scenario structure partially followed the
format used in Study 2, but this time, no visual stimuli were introduced.
For the low-status AI condition, participants read: "Orchid is an artificial
intelligence (AI)-based credit assistant, utilizing machine learning and
data analysis techniques to provide personalized ways to enhance credit
scores. Orchid provides one-on-one services for users and makes de­
cisions based on the input it collects from them in a conversational
manner." For the high-status AI condition, participants read:
"Orchid is an artificial intelligence (AI)-based credit analyst, utilizing
machine learning and data analysis techniques to provide personalized
ways to enhance credit scores. Orchid provides one-on-one services for
users and makes decisions based on the input it collects from them in a
conversational manner."
Participants then proceeded to answer the survey questions listed in
Table 2. Additionally, they responded to the following question to assess
their perception of scenario realism: "I think the information given about
this AI is real."
Descriptive statistics for the scale questions are presented in
Table 17.

Table 16
Demographics in study 3.
Gender
Female
Male
Age
18-30
31-43
44-56
57-69
70+
Ethnicity
Asian
Black
Mixed
White
Other
No Information

159
149

52 %
48 %

134
117
39
15
3

43 %
38 %
13 %
5%
1%

46
45
28
166
21
2

15 %
15 %
9%
54 %
7%
<1 %

Cronbach’s Alpha

Willingness to use the AI
Perceived humanlikeness
Perceived warmth
Perceived competence
Perceived risks
Perceived control
Perceived scenario realism

4.80
2.87
3.27
4.83
3.15
3.74
4.00

1.45
1.47
1.59
1.40
1.25
1.39
1.58

0.92
0.91
0.96
0.96
0.90
0.91
–

8.3.3. Analyses of the full model
The same custom model used in previous studies was applied to
analyze the full model in this study. Additionally, each variable was
regressed on perceived scenario realism to account for its potential
influence.
The indirect effect through perceived humanlikeness alone was nonsignificant (β = − 0.08, BootSE = 0.01, BootCI = [− 0.03, 0.02]). How­
ever, there were significant indirect effects of the serial mediations
through perceived humanlikeness and warmth (β = − 0.03, BootSE =
0.02, BootCI = [− 0.07, − 0.003]), as well as perceived humanlikeness
and competence (β = − 0.05, BootSE = 0.03, 95 % BootCI = [− 0.11,
− 0.005]). Moreover, the indirect effect of perceived risks alone was also
negative and significant (β = − 0.05, BootSE = 0.03, BootCI = [− 0.11,
− 0.01]). This result suggests that people’s risk perceptions did not
depend on their perception of the AI’s humanlikeness in high-criticality
tasks. Participants attributed higher risks to using a credit analyst AI
compared to a credit assistant AI, which subsequently reduced their
willingness to use it. From this perspective, these findings align with the
theorization of this paper, which proposes that people rely on quick
heuristic interpretations rather than analytical judgments in highcriticality tasks due to algorithm aversion. Therefore, participants may
have been more inclined to assess perceived risks of conversational AI
using salient communication of “assistant” title, without considering its
hierarchical status. This result has also been discussed in discussion
section.
Additionally, the direct effect of perceived hierarchical status
became nonsignificant (β = − 0.04, SE = 0.1, p = 0.65, CI = [− 0.24,
0.15]). This indicates full serial mediation through humanlikeness and
competence, as well as through humanlikeness and warmth. Increased
perceptions of hierarchical status decreased perceived humanlikeness,
which in turn reduced people’s willingness to use the AI by lowering
perceptions of both its competence (H2b) and warmth. This result sup­
ported hypothesis 2b.
The indirect effects of serial mediations through humanlikeness and

8.3.1. Manipulation check
In this study, the manipulation check analysis controlled for
perceived scenario realism. Results from a linear regression analysis
confirmed the success of the manipulation, as the credit analyst AI (1)
was perceived to have a higher hierarchical status than the credit as­
sistant AI (0) (MLow = 3.65, SDLow = 1.65, MHigh = 4.05, SDHigh = 1.38,
β = 0.43, Std. Error = 0.17, t = 2.52, p = 0.012, R2 = 0.049; Adjusted R2

Percentage

St. Dev.

8.3.2. Main effect
The main effect analysis was also conducted while controlling for the
effect of perceived scenario realism. Results of a linear regression test
showed that the effect observed in previous studies reversed in this
study, as participants reported lower willingness to use the AI in the
credit analyst AI (1) condition compared to the credit assistant AI (0)
condition (MLow = 5.02, SDLow = 1.30, MHigh = 4.63, SDHigh = 1.54, β =
− 0.31, Std. Error = 0.15, t = − 2.07, p = 0.039, R2 = 0.20; Adjusted R2
= 0.20, Std. β = − 0.11). Perceived scenario realism also showed a sig­
nificant positive effect on willingness to use (β = 0.40, Std. Error = 0.05,
t = 8.43, p < 0.001, Std. β = 0.43). Thus, hypothesis 1b (H1b) was
supported.

8.3. Results

Count

Mean

= 0.043, Std. β = 0.14). The effect of perceived scenario realism was also
positive and significant (β = 0.17, Std. Error = 0.05, t = 3.21, p = 0.001,
Std. β = 0.18). This indicates that as participants perceived the scenario
as more realistic, they were more likely to attribute a higher hierarchical
status to the AI in their assigned condition. This is an expected result in
high criticality tasks. As previously discussed, in high criticality tasks,
people tend to have inflated perceptions of the service provider’s status.

8.2. Method

Item

Variable

14

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

perceived risks (β = − 0.01, BootSE = 0.01, BootCI = [− 0.02, 0.01]) and
through humanlikeness and perceived control (β = 0.00001, BootSE =
0.002, BootCI = [− 0.004, 0.003]) did not reach significance (see
Table 18).

Table 18
Direct and indirect effects in the full model.

8.3.4. Common method bias
Harman’s single-factor test in SPSS confirmed the absence of com­
mon method bias, as a single factor explained only 44.6 % of the total
variance, which is below the 50 % threshold. Additionally, the
maximum vif was 2.40, remaining below the most conservative
threshold of 3. Thus, the possibility of method bias is effectively ruled
out.

Perceived humanlikeness > Perceived warmth

Examined effects in the full model

Results

Perceived hierarchy > Perceived humanlikeness

β* = − 0.34, SE = 0.15, 95 % CI
= [-0.64, − 0.04]
β = 0.72, SE = 0.05, 95 % CI =
[0.62, 0.81]
β* = 0.46, SE = 0.05, 95 % CI =
[0.37, 0.55]
β* = − 0.11, SE = 0.05, 95 % CI
= [-0.21, − 0.01]
β* = 0.29, SE = 0.06, 95 % CI =
[0.18, 0.40]
β = 0.03, SE = 0.05, 95 % CI =
[-0.06, 0.13]
β = − 0.04, SE = 0.1, 95 % CI =
[-0.24, 0.15]
β = − 0.1, SE = 0.13, 95 % CI =
[-0.35, 0.16]
β = − 0.04, SE = 0.13, 95 % CI =
[-0.29, 0.21]
β* = 0.31, SE = 0.13, 95 % CI =
[0.04, 0.57]
β = 0.02, SE = 0.15, 95 % CI =
[-0.28, 0.32]
β* = 0.20, SE = 0.05, 95 % CI =
[0.11, 0.29]
β* = 0.51, SE = 0.05, 95 % CI =
[0.41, 0.61]
β = − 0.001, SE = 0.04, 95 % CI
= [-0.07, 0.07]
β* = − 0.25, SE = 0.05,
95 % CI = [-0.34, − 0.16]
β* = 0.38, SE = 0.05,
95 % CI = [0.28, 0.47]
β* = 0.12, SE = 0.04,
95 % CI = [0.03, 0.21]
β* = 0.23, SE = 0.04,
95 % CI = [0.14, 0.31]
β = 0.03, SE = 0.05,
95 % CI = [-0.07, 0.13]
β* = − 0.22, SE = 0.05,
95 % CI = [-0.31, − 0.13]
β = 0.03, SE = 0.03,
95 % CI = [-0.03, − 0.10]
β = − 0.01, BootSE = 0.01, 95 %
BootCI = [-0.03, 0.02]
β = − 0.01, BootSE = 0.02, 95 %
BootCI = [-0.05, 0.02]
β = − 0.01, BootSE = 0.04, 95 %
BootCI = [-0.1, 0.1]
β = 0.00001, BootSE = 0.004,
95 % BootCI = [-0.01, 0.01]
β* = − 0.05, BootSE = 0.03, 95 %
BootCI = [-0.11, − 0.01]
β * = − 0.03, BootSE = 0.02, 95
% BootCI = [-0.07, − 0.003]
β* = − 0.05, BootSE = 0.03, 95 %
BootCI = [-0.11, − 0.005]
β = − 0.01, BootSE = 0.01, 95 %
BootCI = [-0.02, 0.0004]
β = 0.0001, BootSE = 0.03, 95 %
BootCI = [-0.01, 0.01]

Perceived humanlikeness > Perceived
competence
Perceived humanlikeness > Perceived risks
Perceived humanlikeness > Perceived control
Perceived humanlikeness > Willingness to use

9. Discussion

Perceived hierarchy > Willingness to use

In this paper, a pretest categorized four scenarios—home assistance,
task management for work, teaching, and credit score manage­
ment—into low, medium, and high criticality tasks. Using these sce­
narios, one PLS-SEM study (Study 1a) and three experimental studies
(Studies 1b, 2, and 3) demonstrated that when task criticality is low or
moderate, high-hierarchy conversational AI (e.g., manager, teacher,
analyst) generates greater willingness to use compared to low-hierarchy
AI with assistant titles. This effect was serially mediated by perceived
humanlikeness and, in turn, perceived competence—where higher
perceived humanlikeness due to the AI’s elevated status led to greater
perceived competence. Studies 1a, 1b, and 2 tested this mechanism in
low and moderate criticality tasks, consistently supporting Hypotheses
1a and 2a. These results align with the findings of Jeon (2022), who
tested a similar mechanism in the context of sportswear customer
service.
Conversely, Study 3, as a unique contribution, revealed that in highcriticality tasks, this effect reversed—low-hierarchy AI was perceived as
more humanlike and more competent, leading to greater willingness to
use it. This effect remained robust even when controlling for scenario
realism to enhance external validity. Accordingly, in Study 3, hypoth­
eses 1b and 2b were confirmed. Table 19 presents a summary of all key
findings discussed in the current manuscript and Fig. 2 illustrates the
final model proposed in this paper, reflecting the consistent findings
across all studies.
This study also tested alternative serial mediation pathways
involving perceived warmth, control, and risks alongside perceived
humanlikeness to assess the robustness of the focal pathway. The focal
pathway—through
perceived
humanlikeness
and
com­
petence—remained robust across all four studies. Meanwhile, the
mediation pathway through perceived warmth was significant only in
Studies 1a, 1b, and 3, while the pathways through perceived control and
perceived risks were significant only in Study 2 (teaching task). In study
2, perceived warmth did not explain willingness to use significantly, but
risks and control. This result suggests that, as a domain-specific char­
acteristic in teaching, people may make more nuanced judgments about
a conversational AI teacher’s humanlike accuracy (competence) and its
ability to correct its mistaken assessments (control and risks) rather than
its humanlike warmth (Ding et al., 2023). These results may also stem
from the underlying utilitarian motivations behind using AI-based
tutoring services. Dual-process models (Kahneman, 2011) suggest that
affective responses (e.g., warmth perception) may be less influential
when making utilitarian decisions about AI use. Instead, users engage in
more cognitive processing, where risk assessment and perceived control
over the AI system take precedence. Additionally, Study 3 (credit score
management task) revealed a significant mediation pathway through
perceived risks. Based on existing research, it was initially proposed that
both risk and control perceptions would play a crucial role in
high-criticality tasks. The results showed significance only for risk per­
ceptions, and this pathway did not include perceived humanlikeness.
This finding likely stems from the weak effect of perceived humanlike­
ness on risk perceptions. One possible explanation is heightened

Perceived hierarchy > Perceived warmth
Perceived hierarchy > Perceived competence
Perceived hierarchy > Perceived risks
Perceived hierarchy > Perceived control
Perceived warmth > Willingness to use
Perceived competence > Willingness to use
Perceived control > Willingness to use
Perceived risks > Willingness to use
Perceived realism > Perceived humanlikness
Perceived realism > Perceived warmth
Perceived realism > Perceived competence
Perceived realism > Perceived control
Perceived realism > Perceived risks
Perceived realism > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Willingness to use
Perceived hierarchy > Perceived warmth >
Willingness to use
Perceived hierarchy > Perceived competence >
Willingness to use
Perceived hierarchy > Perceived control >
Willingness to use
Perceived hierarchy > Perceived risks >
Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived warmth > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived competence > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived risks > Willingness to use
Perceived hierarchy > Perceived humanlikeness
> Perceived control > Willingness to use

Note: An asterisk (*) denotes statistically significant relationships.

algorithm aversion in high-criticality tasks, which aligns with people’s
tendency to rely on heuristic judgments. In such contexts, individuals
may have been more risk-focused, overlooking other cues—such as
humanlikeness—when evaluating the risks of interacting with a
conversational AI in high criticality tasks (Castelo et al., 2019). Shaffer
et al. (2013) previously discussed that when algorithm aversion is
triggered, employing algorithms in service duties could impact people’s
internal locus of control. However, in the current paper, no significant
effect of perceived control was found in a high-criticality task as ex­
pected. At this point, further research is needed to explore the potential
15

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Table 19
Summary of findings.
Studies

AI title(s)

Task
Criticality

Independent
Variable (IV)

First Level
Mediator(s) (M1)

Second Level
Mediator(s)
(M2)

Dependent
Variable (DV)

Key Finding(s)

Pretest

–

–

Scenarios

–

–

Perceived task
criticality

Study
1a

Home assistant

Low

Perceived
hierarchical
status

Perceived
humanlikeness

Willingness to
use

Study
1b

Task assistant
versus task
manager

Low

Perceived
hierarchical
status

Perceived
humanlikeness

Study 2

Teaching
assistant versus
teacher

Moderate

Perceived
hierarchical
status

Perceived
humanlikeness

Study 3

Credit assistant
versus credit
analyst

High

Perceived
hierarchical
status

Perceived
humanlikeness

Perceived
competence
Perceived
warmth
Perceived risks
Perceived
control
Perceived
competence
Perceived
warmth
Perceived risks
Perceived
control
Perceived
competence
Perceived
warmth
Perceived risks
Perceived
control
Perceived
competence
Perceived
warmth
Perceived risks
Perceived
control

While the home assistant and job task assistant
scenarios were found to have similar levels of
criticality, the education and credit management
scenarios were identified as more critical, with the
credit management task rated as the most critical
overall.
The findings indicated that only the serial mediation
pathways through perceived humanlikeness and
warmth, as well as through perceived humanlikeness
and competence, were both positive and significant.

Willingness to
use

There were full serial mediations only through
humanlikeness and warmth, as well as through
humanlikeness and competence.

Willingness to
use

The findings revealed significant indirect effects of
the serial mediations only through perceived
humanlikeness and competence, perceived
humanlikeness and control, and perceived
humanlikeness and risks.

Willingness to
use

There were significant indirect effects of the serial
mediations through perceived humanlikeness and
warmth, as well as perceived humanlikeness and
competence

Fig. 2. Final model.

reasons behind the nonsignificant effect of perceived control. It would
be valuable to investigate whether other factors, such as user familiarity
with AI, might play a role in how individuals perceive and interact with
high-criticality tasks. Additionally, exploring the interplay between al­
gorithm aversion, perceived control, and hierarchical status of conver­
sational AI could shed light on why control perceptions were not as
impactful in this context. Understanding these dynamics will help refine
the model and provide deeper insights into how users approach
conversational AI in high-criticality situations.

9.1. Theoretical implications
To date, Jeon (2022) has shown that higher-hierarchy job titles
correspond to greater perceptions of likeability and knowledgeability,
leading to an increased willingness to buy products recommended by a
conversational AI-based service provider. Sundar et al. (2017) demon­
strated that higher-hierarchy AIs are more strongly associated with
competence-related characteristics. Additionally, previous research
explored the effect of a specialist heuristic, where the "specialist" title
assigned to media agents and computerized systems resulted in height­
ened perceptions of competence compared to roles designated as
"generalist" (Kim, 2014; Koh & Sundar, 2010; Nass et al., 1996; Sah
16

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

et al., 2011; Sundar et al., 2017). Therefore, the relationship between an
AI’s hierarchical status, competence, and subsequent consumption
behavior has been previously explored in past research, which yielded
results that contradict current marketing practices for conversational AI
service providers, where "assistant" titles are often used as a marketing
strategy.
The current research contributes to the existing body of knowledge
by first demonstrating that the mediation mechanism is further medi­
ated by perceived humanlikeness. Past research has discussed that
humanlikeness is a prerequisite for perceptions of competence in the
case of conversational AI (Balakrishnan & Dwivedi, 2024). The current
paper builds on this premise by showing that the hierarchical status of
conversational AI can also influence its perceived humanlikeness.
Secondly, as another contribution, the paper reveals that the service
recipients’ greater inclination toward using higher hierarchy conversa­
tional AI-based services is task-dependent, confirming past research
findings indicating greater user preference of high hierarchy AI only in
low criticality tasks. However, as task criticality increases, people
become more willing to use a low-hierarchy "assistant" conversational
AI, supporting the validity of current marketing practices.
Thirdly, past research has discussed that AI aversion can be reduced
by increasing the perceived humanlikeness of AI (Bigman & Gray, 2018;
Castelo et al., 2019; Dang & Liu, 2024; Haslam, 2006; Messingschlager
& Appel, 2023; Shaffer et al., 2013). As a novel contribution, the current
paper suggests that there may be a spiral relationship between algo­
rithmic aversion and perceived humanlikeness. It demonstrates that AI
aversion may also decrease the perceived humanlikeness of
high-hierarchy conversational AI, triggering an “infrahumanization”
process in high criticality tasks.
This research also contributes to the Computers Are Social Actors
(CASA) paradigm (Nass et al., 1994) by demonstrating how symbolic
cues—like hierarchical job titles—can activate social heuristics in user
perceptions of AI agents. The findings also offer new insights into
interpersonal communication theories (Braithwaite & Schrodt, 2021),
suggesting that traditional markers of authority and competence in
human interactions remain influential even in AI-mediated contexts. By
showing that humanlikeness is a prerequisite for competence attribu­
tion, the study affirms that AI-human interactions are still evaluated
through fundamentally human communicative frameworks.
This research extends the validity of previous studies that focused on
a single domain with a commercial AI service provider (e.g., sportswear
customer service) (Jeon, 2022) by demonstrating their applicability
across various conversational AI contexts beyond customer service, such
as education, home assistantship, and organizational task management.
Consequently, while Jeon (2022) examined hierarchy primarily
through the disparity between "assistant" and "manager" titles, the cur­
rent study broadens this perspective by incorporating other hierarchical
job titles, such as "teacher" and "credit analyst," compared to "teaching
assistant" and "credit assistant." This approach was intended to minimize
potential biases associated with the "manager" title. Furthermore, this
study highlights that the previously established mechanism reverses in
higher criticality tasks, such as credit score management, where
lower-hierarchy "assistant" AI agents are preferred.

competence were added in a serial mediation, indicating a full media­
tion effect. This suggests that despite the current boom in AI technolo­
gies, humanlike qualities in conversational AI settings are still a
prerequisite for ascribing other characteristics, such as competence, to
the AI. Human characteristics are still perceived as superior to machinelike qualities. From this perspective, AI developers may benefit more
from simple and cost-effective marketing strategies, such as assigning
role titles to their AI agents, than from investing large sums in improving
the perceived technical capabilities of AI in conversational services to
increase its perceived competence.
However, this suggestion also brings an ethical dilemma into the
picture. AI is currently not an immaculate entity. It has been shown that
even the most advanced AI agents today tend to spread misinformation,
as AI is a probabilistic rather than a deterministic entity. Therefore, it
sometimes “hallucinates.” (Rae, 2024). This could be especially detri­
mental in high criticality domains that require accurate and factual in­
formation, such as healthcare. The risk of such inaccuracies becomes
even more pronounced when AI agents are perceived as more competent
and humanlike due to ascribed job titles, potentially misleading users
into trusting them more than they should. This raises concerns about
transparency and accountability, especially in situations where users’
decisions based on AI’s advice could have significant consequences.
Building on the findings of this paper, concrete implications emerge
for chatbot design across various applied domains. In mental health
services, where emotional sensitivity and trust are critical, assigning
lower-status titles (e.g., “psychiatry assistant” rather than “psychiatrist”)
may foster user comfort and encourage disclosure in high-stakes con­
versations, such as those involving suicidal ideation.
In contrast, in educational settings, higher-status titles (e.g.,
“teacher” rather than “teaching assistant”) may enhance the AI agent’s
perceived authority and competence, thereby improving users’ recep­
tiveness to its guidance. While higher hierarchical titles can enhance
perceived credibility in routine or lower-criticality teaching tasks (e.g.,
providing course information or study tips), caution is warranted in
assigning such roles to conversational AI agents. Overpositioning AI as a
primary teaching authority may lead to diminished student trust,
particularly if the AI fails to meet expectations related to pedagogical
nuance, empathy, or adaptability. To address this, transparency
regarding the AI’s capabilities and limitations is essential. Positioning AI
agents as supplementary rather than primary instructors may help
manage expectations while still benefiting from hierarchical framing.
For instance, although titles such as “Teacher” or “Professor” may
enhance perceived competence, a more appropriate designation—such
as “Supplementary Professor”—could better reflect the agent’s sup­
portive role and preserve user trust.
In customer service, where the criticality of user needs can vary
widely, dynamic adjustment of hierarchical framing may help align user
expectations with the AI’s role. For instance, a chatbot may be intro­
duced as a “customer service manager” for routine inquiries, but framed
as a “customer service assistant” when addressing more complex or
urgent issues. Such task-contingent role assignments can help optimize
perceived fit and trust in AI across a range of contexts.
9.3. Limitations and future research direction

9.2. Practical implications

The current paper has several limitations that should be acknowl­
edged. Firstly, it exclusively focused on investigating conversational AIbased service providers. While humanlikeness and competence of AI
may be important prerequisites for people’s willingness to use conver­
sational AI-based service providers in such service domains, it remains
unclear how the hierarchical status of AI might manifest an effect in
domains where service recipients do not directly interact with it. The
weaker hierarchical sense in such service settings may result in weaker
influence on perceived humanlikeness and competence.
Secondly, despite examining the effect of task criticality across
different studies, the current paper does not consider this variable as a

This paper holds profound managerial implications. To begin with, it
suggests that assigning job titles to conversational AI service providers
should not be taken as an arbitrary task. The paper provides guidelines
for this process, showing that people prefer high-hierarchy conversa­
tional AI in low-criticality tasks, whereas, in high-criticality tasks, the
preference shifts toward low-hierarchy conversational AI.
Secondly, the results demonstrated that the hierarchical status of
conversational AI can impact its perceived competence by altering the
attributed humanlikeness to it. The direct effect of perceived hierar­
chical status was nonsignificant in the studies when humanlikeness and
17

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

moderator. Therefore, the paper is limited in its ability to compare the
effects of different mediation pathways across varying levels of task
criticality to resolve their impact on competence statistically. From this
perspective, the current paper suggests that task criticality should be
considered a compelling moderator for future research.
Thirdly, although the current paper divides tasks based on their
criticality, it does not differentiate them based on their motivational
nature. All the scenario examples investigated in this research had a
utilitarian nature, wherein the service recipient was primarily motivated
by utilitarian reasons to interact with the given conversational AI service
providers such as task management, teaching, home assistantship, and
credit score management. Utilitarian domains typically do not necessi­
tate a substantial emotional capacity by the service provider to satis­
factorily implement tasks (Whitley et al., 2018). However, in domains
where the service provider is expected to possess emotional skills such as
empathy in elder care or affection in the education of children, altruistic
job titles such as "assistant" may be preferred to a greater extent. That
said, task qualities such as being utilitarian or hedonic emerge as a
compelling moderator candidate for the future research.
The present study’s examination of perceived humanlikeness as a
predictor of perceived warmth may have inflated the results in the
mediation analyses. Prior research suggests that perceived warmth is a
subcomponent of perceived humanness (Fiske, Cuddy, Peter, & Xu,
2024), indicating a conceptual overlap between these constructs. As
such, their inclusion in a serial mediation model may reflect a degree of
construct nesting, which could artificially inflate effect sizes. This po­
tential confounding effect introduced by the serial mediation design
through perceived humanlikeness and warmth warrants caution. Future
research exploring similar mechanisms should take this overlap into
account and consider alternative modeling approaches to ensure more
accurate interpretation of mediation effects.
This paper also does not establish clear thresholds for task criticality,
which limits the ability to definitively categorize tasks as low, medium,
or high criticality. For instance, what this study considers high criticality
may not be universally applicable, and there may be tasks that could add
an additional layer of criticality. The paper assumes that task criticality
is assessed similarly by all individuals, but this assumption may not hold
across different contexts or groups. This further highlights the need for
future research to refine the definition and application of task criticality
as a moderator, ensuring a more nuanced understanding of how
different individuals perceive and assess criticality in various tasks.
Additionally, this study does not address situations where algorithm
aversion could completely disrupt the observed mechanisms. For
example, in very high criticality areas like healthcare, AI could face
complete rejection regardless of its job title, which may nullify the ef­
fects explored in this research. Future studies should explore such con­
texts and identify them as boundary conditions for the proposed
mechanisms.
While the results of this study provide valuable insights, it is
important to acknowledge that some of the main effect findings yielded
small effect sizes (e.g., d = 0.25) and some indirect effects had small β
coefficients. Although these results were statistically significant, the
practical significance or real-world impact of these effects may be
limited and not widely applicable. These suggests that the relationships
between the variables, while present, may not be as strong or conse­
quential as larger effects would indicate. This could be due to various
factors, such as the specific context of the study, the sample

characteristics, or the measurement tools used. Additionally, small effect
sizes may imply that other unmeasured variables or external factors not
accounted for in the current model could be influencing the outcomes.
Finally, across all four studies, the sample predominantly consisted
of a relatively younger age group (18–30), which may have introduced a
bias in the findings. This demographic is generally characterized by
higher levels of digital literacy and greater familiarity with conversa­
tional AI agents, which may influence both their expectations and in­
teractions with such technologies. Furthermore, this age disparity may
have also affected participants’ perceptions of task criticality. Tasks that
might be considered critical for older individuals may not be perceived
as equally critical by younger participants, who may feel a stronger
sense of self-agency and confidence in managing those tasks indepen­
dently without the assistance of an AI agent. Future research should take
these factors into account and investigate whether the observed effects
generalize across broader age groups with varying levels of digital
proficiency and task-related perceptions.
9.4. Conclusion
Companies assign “assistant” job titles to conversational AI service
providers to increase their adoption by the public as a marketing strat­
egy. However, previous limited research in the literature has contra­
dicted this approach by showing that higher hierarchy job titles may be
more effective in enhancing favorable user behaviors, as higher status
job titles increase the perceived knowledgeability of the AI as an indi­
cator of competence. Applying a multi-method approach with three
experiments and one structural equation modeling analysis, this
research uniquely examines the effect of the hierarchical status of
conversational AI agents, conveyed by their job titles, in different task
criticalities.
The findings reveal that in high task criticality, people prefer lowhierarchy AI with an "assistant" title (e.g., credit assistant), as sug­
gested by current marketing practices. However, when the task criti­
cality is not high, the preference shifts toward higher hierarchy AI (e.g.,
task manager, teacher) confirming the findings of the past research. As a
second contribution, this research also demonstrates that the perceived
hierarchical status of AI affects its perceived humanlikeness differently
across various task criticalities.
Based on these findings, the current paper provides valuable guide­
lines for companies using conversational AI service providers in service
duties, highlighting the importance of considering task criticality when
choosing hierarchical titles to optimize user adoption and perceptions.
Declaration of generative AI and AI-assisted technologies in the
writing process
During the preparation of this work the author(s) used ChatGPT in
order to check the grammar of the written material. After using this tool/
service, the author(s) reviewed and edited the content as needed and
take(s) full responsibility for the content of the publication.
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.

18

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Appendix A. Experiment stimulus in Study 1b

Appendix B. Experiment stimuli in Study 2
High Status.

19

Computers in Human Behavior Reports 18 (2025) 100682

M. Yanit

Low Status.

Data availability

Alicke, M. D., Klotz, M. L., Breitenbecher, D. L., Yurak, T. J., & Vredenburg, D. S. (1995).
Personal contact, individuation, and the better-than-average effect. Journal of
personality and social psychology, 68(5), 804.
Anderson, C., Kraus, M. W., Galinsky, A. D., & Keltner, D. (2012). The local-ladder effect:
Hierarchical status and subjective well-being. Psychological science, 23(7), 764–771.
Ansoff, H. I., Kipley, D., Lewis, A. O., Helm-Stevens, R., & Ansoff, R. (2018). Implanting
strategic management. Springer.
Baek, T. H., Bakpayev, M., Yoon, S., & Kim, S. (2022). Smiling AI agents: How
anthropomorphism and broad smiles increase charitable giving. International Journal
of Advertising, 41(5), 850–867.
Balakrishnan, J., & Dwivedi, Y. K. (2024). Conversational commerce: Entering the next
stage of AI-powered digital assistants. Annals of Operations Research, 333(2),
653–687.
Belanche, D., Casaló, L. V., & Flavián, C. (2019). Artificial intelligence in FinTech:
Understanding robo-advisors adoption among customers. Industrial Management &
Data Systems, 119(7), 1411–1430.

I have shared a link to the data repository where my data can be
downloaded from.
References
Ahearne, M., Mathieu, J., & Rapp, A. (2005). To empower or not to empower your sales
force? An empirical examination of the influence of leadership empowerment
behavior on customer satisfaction and performance. Journal of Applied Psychology, 90
(5), 945.
Ahn, J. K. J., & Sung, Y. (2022). The effect of gender stereotypes on artificial intelligence
recommendations. Journal of Business Research, 141, 50–59.

20

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Bendapudi, N., & Berry, L. L. (1997). Customers’ motivations for maintaining
relationships with service providers. Journal of Retailing, 73(1), 15–37.
Berger, J., Cohen, B. P., & Zelditch, J. M. (1972). Status characteristics and social
interaction. American Sociological Review, 241–255.
Berger, J., Rosenholtz, S. J., & Zelditch, J. M. (1980). Status organizing processes. Annual
Review of Sociology, 6(1), 479–508.
Bering, J. M. (2002). The existential theory of mind. Review of General Psychology, 6(1),
3–24.
Bernardi, R. A. (1994). Validating research results when cronbach’s alpha is below. 70: A
methodological procedure. Educational and Psychological Measurement, 54(3),
766–775.
Bigman, Y. E., & Gray, K. (2018). People are averse to machines making moral decisions.
Cognition, 181, 21–34.
Blut, M., Wang, C., Wünderlich, N. V., & Brock, C. (2021). Understanding
anthropomorphism in service provision: A meta-analysis of physical robots,
chatbots, and other AI. Journal of the Academy of Marketing Science, 49, 632–658.
Boehm, C., & Boehm, C. (2009). Hierarchy in the forest: The evolution of egalitarian
behavior. Harvard University Press.
Bonaccio, S., & Dalal, R. S. (2006). Advice taking and decision-making: An integrative
literature review, and implications for the organizational sciences. Organizational
Behavior and Human Decision Processes, 101(2), 127–151.
Borau, S., Otterbring, T., Laporte, S., & Fosso Wamba, S. (2021). The most human bot:
Female gendering increases humanness perceptions of bots and acceptance of AI.
Psychology and Marketing, 38(7), 1052–1068.
Braithwaite, D. O., & Schrodt, P. (2021). Engaging theories in interpersonal communication:
Multiple perspectives. Routledge.
Branscombe, N. R., & Wann, D. L. (1994). Collective self-esteem consequences of
outgroup derogation when a valued social identity is on trial. European Journal of
Social Psychology, 24(6), 641–657.
Burton, J. W., Stein, M. K., & Jensen, T. B. (2020). A systematic review of algorithm
aversion in augmented decision making. Journal of Behavioral Decision Making, 33(2),
220–239.
Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion.
Journal of Marketing Research, 56(5), 809–825.
Castelvecchi, D. (2016). Can we open the black box of AI? Nature News, 538(7623), 20.
Chandra, S., Shirish, A., & Srivastava, S. C. (2022). To be or not to be… human?
Theorizing the role of human-like competencies in conversational artificial
intelligence agents. Journal of Management Information Systems, 39(4), 969–1005.
Chanseau, A., Dautenhahn, K., Koay, K. L., Walters, M. L., Lakatos, G., & Salem, M.
(2019). How does peoples’ perception of control depend on the criticality of a task
performed by a robot. Paladyn. Journal of Behavioral Robotics, 10(1), 380–400.
Cheah, J. H., Sarstedt, M., Ringle, C. M., Ramayah, T., & Ting, H. (2018). Convergent
validity assessment of formatively measured constructs in PLS-SEM: On using singleitem versus multi-item measures in redundancy analyses. International Journal of
Contemporary Hospitality Management, 30(11), 3192–3210.
Cheng, X., Zhang, X., Cohen, J., & Mou, J. (2022). Human vs. AI: Understanding the
impact of anthropomorphism on consumer response to chatbots from the perspective
of trust and relationship norms. Information Processing & Management, 59(3), Article
102940.
Chi, R., Zhang, J., & Pan, M. (2024). The effect of anthropomorphic competence-warmth
congruence of service robots on recommendation intention. Current Psychology, 43
(7), 6570–6583.
Chin, W. W. (1998). The partial least squares approach to structural equation modeling.
Modern methods for business research, 295(2), 295–336.
Cho, G., Hwang, H., Sarstedt, M., & Ringle, C. M. (2020). Cutoff criteria for overall model
fit indexes in generalized structured component analysis. Journal of marketing
analytics, 8(4), 189–202.
Christoforakos, L., Gallucci, A., Surmava-Große, T., Ullrich, D., & Diefenbach, S. (2021).
Can robots earn our trust the same way humans do? A systematic exploration of
competence, warmth, and anthropomorphism as determinants of trust development
in hri. Frontiers in Robotics and AI, 8, Article 640444.
Clement, L. M., & Bradley-Garcia, M. (2022). A step-by-step tutorial for performing a
moderated mediation analysis using PROCESS. The Quantitative Methods for
Psychology, 18(3), 258–271.
Coombs, C., Stacey, P., Kawalek, P., Simeonova, B., Becker, J., Bergener, K., …
Trautmann, H. (2021). What is it about humanity that we can’t give away to
intelligent machines? A European perspective. International Journal of Information
Management, 58, Article 102311.
Crocker, J., Thompson, L. L., McGraw, K. M., & Ingerman, C. (1987). Downward
comparison, prejudice, and evaluations of others: Effects of self-esteem and threat.
Journal of personality and social psychology, 52(5), 907.
Dang, J., & Liu, L. (2024). Extended artificial intelligence aversion: People deny
humanness to artificial intelligence users. Journal of Personality and Social
Psychology, Advance online publication. https://doi.org/10.1037/pspi0000480.
Dash, G., & Paul, J. (2021). CB-SEM vs PLS-SEM methods for research in social sciences
and technology forecasting. Technological Forecasting and Social Change, 173, Article
121092.
Demczuk, R., Mantovani, D., & Fernandes, D. (2023). Looking up or down on the social
ladder: How socioeconomic comparisons shape judgments about monetary and time
donations. Journal of Behavioral Decision Making, 36(3), Article e2308.
Ding, L., Li, T., Jiang, S., & Gapud, A. (2023). Students’ perceptions of using ChatGPT in
a physics class as a virtual tutor. International Journal of Educational Technology in
Higher Education, 20(1), 63.
Douglas, B. D., Ewell, P. J., & Brauer, M. (2023). Data quality in online human-subjects
research: Comparisons between MTurk, prolific, CloudResearch, Qualtrics, and
SONA. PLoS One, 18(3), Article e0279720.

Faul, F., Erdfelder, E., Lang, A. G., & Buchner, A. (2007). G* power 3: A flexible statistical
power analysis program for the social, behavioral, and biomedical sciences. Behavior
Research Methods, 39(2), 175–191.
Fein, S., & Spencer, S. J. (1997). Prejudice as self-image maintenance: Affirming the self
through derogating others. Journal of personality and Social Psychology, 73(1), 31.
Fiske, S. T., Cuddy, A. J., Peter, G., & Xu, J. (2024). A model of (often mixed) stereotype
content: Competence and warmth respectively follow from perceived status and
competition: Correction to Fiske et al. (2002).Journal of Personality and Social
Psychology, 126(3), 412. https://doi.org/10.1037/pspa0000163.
Fiske, A. P. (2004). Four modes of constituting relationships: Consubstantial
assimilation; space, magnitude, time, and force; concrete procedures; abstract
symbolism. In N. Haslam (Ed.), Relational models theory: A contemporary overview (pp.
61–146). Mahwah, NJ: Lawrence Erlbaum Associates, Inc.
Følstad, A., Law, E. L. C., & van As, N. (2024). Conversational breakdown in a customer
service chatbot: Impact of task order and criticality on user trust and emotion. ACM
Transactions on Computer-Human Interaction, 31(5), 1–52.
Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with
unobservable variables and measurement error. Journal of Marketing Research, 18(1),
39–50.
Gergen, K. J. (1991). The saturated self: Dilemmas of identity in contemporary life. New
York: Basic Books.
Gervais, W. M., & Norenzayan, A. (2012). Analytic thinking promotes religious disbelief.
Science, 336(6080), 493–496.
Goodman, M. D., & Gareis, K. C. (1993). The influence of status on decisions to help. The
Journal of Social Psychology, 133(1), 23–31.
Grant, A. M., Berg, J. M., & Cable, D. M. (2014). Job titles as identity badges: How
selfreflective titles can reduce emotional exhaustion. Academy of Management
Journal, 57(4), 1201–1225.
Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science,
315(5812), 619, 619.
Groysberg, B., Polzer, J. T., & Elfenbein, H. A. (2011). Too many cooks spoil the broth:
How high-status individuals decrease group effectiveness. Organization Science, 22
(3), 722–737.
Gudergan, S. P., Moisescu, O. I., Radomir, L., Ringle, C. M., & Sarstedt, M. (2025). Special
issue editorial: Advanced partial least squares structural equation modeling (PLSSEM) applications in business research. Journal of Business Research, 188, Article
115087.
Gursoy, D., Chi, O. H., Lu, L., & Nunkoo, R. (2019). Consumers acceptance of artificially
intelligent (AI) device use in service delivery. International Journal of Information
Management, 49, 157–169.
Gwinn, J. D., Judd, C. M., & Park, B. (2013). Less power= less human? Effects of power
differentials on dehumanization. Journal of Experimental Social Psychology, 49(3),
464–470.
Hair Jr, J. F., Hult, G. T. M., Ringle, C. M., Sarstedt, M., Danks, N. P., Ray, S., … Ray, S.
(2021). An introduction to structural equation modeling. Partial least squares
structural equation modeling (PLS-SEM) using R: A workbook (pp. 1–29).
Harris, L. T., & Fiske, S. T. (2011). Perceiving humanity or not: A social neuroscience
approach to dehumanized perception. Social neuroscience: Toward understanding the
underpinnings of the social mind.
Harris-Watson, A. M., Larson, L. E., Lauharatanahirun, N., DeChurch, L. A., &
Contractor, N. S. (2023). Social perception in Human-AI teams: Warmth and
competence predict receptivity to AI teammates. Computers in Human Behavior, 145,
Article 107765.
Haslam, N. (2006). Dehumanization: An integrative review. Personality and Social
Psychology Review, 10(3), 252–264.
Haslam, N., & Bain, P. (2007). Humanizing the self: Moderators of the attribution of
lesser humanness to others. Personality and Social Psychology Bulletin, 33(1), 57–68.
Haslam, S. A., Jetten, J., Postmes, T., & Haslam, C. (2009). Social identity, health and
well-being: An emerging agenda for applied psychology. Applied Psychology-an
International Review-Psychologie Appliquee-Revue Internationale, 58(1), 1–23.
Haslam, N., Loughnan, S., Kashima, Y., & Bain, P. (2008). Attributing and denying
humanness to others. European Review of Social Psychology, 19(1), 55–85.
Haslam, N., Rothschild, L., & Ernst, D. (2000). Essentialist beliefs about social categories.
British Journal of Social Psychology, 39(1), 113–127.
Hayes, A. F. (2012). Process: A versatile computational tool for observed variable mediation,
moderation, and conditional process modeling.
Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for assessing
discriminant validity in variance-based structural equation modeling. Journal of the
Academy of Marketing Science, 43, 115–135.
Heßler, P. O., Pfeiffer, J., & Hafenbrädl, S. (2022). When self-humanization leads to
algorithm aversion: What users want from decision support systems on prosocial
microlending platforms. Business & Information Systems Engineering, 64(3), 275–292.
Horchak, O. V., Giger, J. C., & Garrido, M. V. (2016). Action contribution to competence
judgments: The use of the journey schema. Frontiers in Psychology, 7, 448.
Hu, Q., Lu, Y., Pan, Z., Gong, Y., & Yang, Z. (2021). Can AI artifacts influence human
cognition? The effects of artificial autonomy in intelligent personal assistants.
International Journal of Information Management, 56, Article 102250.
Jamali, J., Ayatollahi, S. M. T., & Jafari, P. (2018). The effect of cross-loading on
measurement equivalence of psychometric multidimensional questionnaires in
mimic model: a simulation study. Materia Socio-Medica, 30(2), 121.
Jeon, Y. A. (2022). Let me transfer you to our AI-based manager: Impact of managerlevel job titles assigned to AI-based agents on marketing outcomes. Journal of
Business Research, 145, 892–904.
Jussupow, E., Benbasat, I., & Heinzl, A. (2020). Why are we averse towards algorithms? A
comprehensive literature review on algorithm aversion.
Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.

21

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Kim, K. J. (2014). Can smartphones be specialists? Effects of specialization in mobile
advertising. Telematics and Informatics, 31(4), 640–647.
Kim, W. B., & Hur, H. J. (2023). What makes people feel empathy for AI chatbots?
Assessing the role of competence and warmth. International Journal of HumanComputer Interaction, 1–14.
Kim, J., Kang, S., & Bae, J. (2022). Human likeness and attachment effect on the
perceived interactivity of AI speakers. Journal of Business Research, 144, 797–804.
Kim, S., & McGill, A. L. (2011). Gaming with mr. Slot or gaming the slot machine? Power,
anthropomorphism, and risk perception. Journal of Consumer Research, 38(1),
94–107.
Kim, S. Y., Schmitt, B. H., & Thalmann, N. M. (2019). Eliza in the uncanny valley:
Anthropomorphizing consumer robots increases their perceived warmth but
decreases liking. Marketing Letters, 30, 1–12.
Koh, Y. J., & Sundar, S. S. (2010). Heuristic versus systematic processing of specialist
versus generalist sources in online media. Human Communication Research, 36(2),
103–124.
Koski, J. E., Xie, H., & Olson, I. R. (2015). Understanding social hierarchies: The neural
and psychological foundations of status perception. Social Neuroscience, 10(5),
527–550.
Krämer, N. C., Eimler, S., Von Der Pütten, A., & Payr, S. (2011). Theory of companions:
What can theoretical models contribute to applications and understanding of humanrobot interaction? Applied Artificial Intelligence, 25(6), 474–502.
Kraus, M. W., & Keltner, D. (2013). Social class rank, essentialism, and punitive
judgment. Journal of personality and social psychology, 105(2), 247.
Kraus, M. W., & Park, J. W. (2017). The structural dynamics of social class. Current
Opinion in Psychology, 18, 55–60.
Kraus, M. W., Piff, P. K., & Keltner, D. (2009). Social class, sense of control, and social
explanation. Journal of personality and social psychology, 97(6), 992.
Kurtaliqi, F., Miltgen, C. L., Viglia, G., & Pantin-Sohier, G. (2024). Using advanced mixed
methods approaches: Combining PLS-SEM and qualitative studies. Journal of Business
Research, 172, Article 114464.
Lammers, J., & Stapel, D. A. (2011). Power increases dehumanization. Group Processes &
Intergroup Relations, 14(1), 113–126.
Lee, T. L., & Fiske, S. T. (2006). Not an outgroup, not yet an ingroup: Immigrants in the
stereotype content model. International Journal of Intercultural Relations, 30(6),
751–768.
Leyens, J. P., Paladino, P. M., Rodriguez-Torres, R., Vaes, J., Demoulin, S., RodriguezPerez, A., et al. (2000). The emotional side of prejudice: The attribution of secondary
emotions to ingroups and outgroups. Personality and Social Psychology Review, 4(2),
186–197.
Leyens, J. P., Rodriguez-Perez, A., Rodriguez-Torres, R., Gaunt, R., Paladino, M. P.,
Vaes, J., et al. (2001). Psychological essentialism and the differential attribution of
uniquely human emotions to ingroups and outgroups. European Journal of Social
Psychology, 31(4), 395–411.
Longoni, C., Bonezzi, A., & Morewedge, C. K. (2019). Resistance to medical artificial
intelligence. Journal of Consumer Research, 46(4), 629–650.
Loughnan, S., & Haslam, N. (2007). Animals and androids: Implicit associations between
social categories and nonhumans. Psychological science, 18(2), 116–121.
Loughnan, S., Haslam, N., Sutton, R. M., & Spencer, B. (2014). Dehumanization and
social class. Social Psychology, 45(1), 54–61.
Matheus, R., Faber, R., Ismagilova, E., & Janssen, M. (2023). Digital transparency and the
usefulness for open government. International Journal of Information Management, 73,
Article 102690.
McLean, G., Osei-Frimpong, K., & Barhorst, J. (2021). Alexa, do voice assistants influence
consumer brand engagement?–Examining the role of AI powered voice assistants in
influencing consumer brand engagement. Journal of Business Research, 124, 312–328.
Messingschlager, T. V., & Appel, M. (2023). Mind ascribed to AI and the appreciation of AIgenerated art. new media & society, Article 14614448231200248.
Mori, M. (1970). The uncanny valley. Energy, 7(4), 33–35.
Moussawi, S., Koufaris, M., & Benbunan-Fich, R. (2021). How perceptions of intelligence
and anthropomorphism affect adoption of personal intelligent agents. Electronic
Markets, 31(2), 343–364.
Mozafari, N., Weiger, W. H., & Hammerschmidt, M. (2022). Trust me, I’m a
bot–repercussions of chatbot disclosure in different service frontline settings. Journal
of Service Management, 33(2), 221–245.
Nass, C., Fogg, B. J., & Moon, Y. (1996). Can computers be teammates? International
Journal of Human-Computer Studies, 45(6), 669–678.
Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. Proceedings of the
SIGCHI conference on Human factors in computing systems, 72–78.
Nazlan, N. H., Tanford, S., & Montgomery, R. (2018). The effect of availability heuristics
in online consumer reviews. Journal of Consumer Behaviour, 17(5), 449–460.
Neururer, M., Schlögl, S., Brinkschulte, L., & Groth, A. (2018). Perceptions on
authenticity in chat bots. Multimodal Technologies and Interaction, 2(3), 60.
O’brien, R. M. (2007). A caution regarding rules of thumb for variance inflation factors.
Quality and Quantity, 41, 673–690.
Pieters, R., Bottschen, G., & Thelen, E. (1998). Customer desire expectations about
service employees: An analysis of hierarchical relations. Psychology and Marketing, 15
(8), 755–773.
Pinto, J. K., Patanakul, P., & Pinto, M. B. (2017). “The aura of capability”: Gender bias in
selection for a project manager job. International Journal of Project Management, 35
(3), 420–431.
Prolific. (2024). Why do I have more participants than requested? Prolific researcher
help center. from https://researcher-help.prolific.com/en/article/345973.
(Accessed 27 March 2025).
Prolific. (2025). Prolific’s payment principles. Prolific Researcher Help Center. from https
://researcher-help.prolific.com/en/article/2273bd. (Accessed 27 March 2025).

Puntoni, S., Reczek, R. W., Giesler, M., & Botti, S. (2021). Consumers and artificial
intelligence: An experiential perspective. Journal of Marketing, 85(1), 131–151.
Rae, I. (2024). The effects of perceived AI use on content perceptions. In In proceedings of
the 2024 CHI conference on human factors in computing systems (pp. 1–14).
Raff, S., Rose, S., & Huynh, T. (2024). Perceived creepiness in response to smart home
assistants: A multi-method study. International Journal of Information Management,
74, Article 102720.
Richter, N. F., Hauff, S., Ringle, C. M., & Gudergan, S. P. (2022). The use of partial least
squares structural equation modeling and complementary methods in international
management research. Management International Review, 62(4), 449–470.
Richter, N. F., & Tudoran, A. A. (2024). Elevating theoretical insight and predictive
accuracy in business research: Combining PLS-SEM and selected machine learning
algorithms. Journal of Business Research, 173, Article 114453.
Saaty, T. L. (2001). Decision making for leaders: The analytic hierarchy process for decisions
in a complex world. RWS publications.
Sah, Y. J., Yoo, B., & Sundar, S. S. (2011). Are specialist robots better than generalist
robots?. In Proceedings of the 6th international conference on Human-robot interaction
(pp. 241–242).
Scaillet, N., & Leyens, J. P. (2000). From incorrect deductive reasoning to ingroup
favouritism. In In Social identity processes: Trends in theory and research (pp. 49–61).
SAGE Publications Ltd.
Shaffer, V. A., Probst, C. A., Merkle, E. C., Arkes, H. R., & Medow, M. A. (2013). Why do
patients derogate physicians who use a computer-based diagnostic support system?
Medical Decision Making, 33(1), 108–118.
Shi, S., Gong, Y., & Gursoy, D. (2021). Antecedents of trust and adoption intention
toward artificially intelligent recommendation systems in travel planning: A
heuristic-systematic model. Journal of Travel Research, 60(8), 1714–1734. https://
doi.org/10.1177/0047287520966395
Sidanius, J., & Pratto, F. (2003). Social dominance theory and the dynamics of
inequality: A reply to Schmitt, branscombe, and kappen and wilson and Liu. British
Journal of Social Psychology, 42(2), 207–213.
Smith, H. J., & Tyler, T. R. (1996). Justice and power: When will justice concerns
encourage the advantaged to support policies which redistribute economic resources
and the disadvantaged to willingly obey the law? European Journal of Social
Psychology, 26(2), 171–200.
Straub, D., Boudreau, M. C., & Gefen, D. (2004). Validation guidelines for IS positivist
research. Communications of the Association for Information Systems, 13(1), 24.
Struch, N., & Schwartz, S. H. (1989). Intergroup aggression: Its predictors and
distinctness from in-group bias. Journal of personality and social psychology, 56(3),
364.
Sundar, S. S. (2008). The main model: A heuristic approach to understanding technology
effects on credibility. Cambridge, MA: MacArthur Foundation Digital Media and
Learning Initiative.
Sundar, S. S., Jung, E. H., Waddell, T. F., & Kim, K. J. (2017). 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, 97, 88–97.
Sundar, S. S., & Nass, C. (2000). Source orientation in human-computer interaction:
Programmer, networker, or independent social actor. Communication Research, 27
(6), 683–703.
Tambe, P., Cappelli, P., & Yakubovich, V. (2019). Artificial intelligence in human
resources management: Challenges and a path forward. California Management
Review, 61(4), 15–42.
Teodoridis, F., Bikard, M., & Vakili, K. (2019). Creativity at the knowledge frontier: The
impact of specialization in fast-and slow-paced domains. Administrative Science
Quarterly, 64(4), 894–927.
Touré-Tillery, M., & McGill, A. L. (2015). Who or what to believe: Trust and the
differential persuasiveness of human and anthropomorphized messengers. Journal of
Marketing, 79(4), 94–110.
Trautt, G. M., & Bloom, L. J. (1982). Therapeugenic factors in psychotherapy: The effects
of fee and title on credibility and attraction. Journal of Clinical Psychology, 38(2),
274–279.
Tyler, T. R., & Degoey, P. (1995). Collective restraint in social dilemmas: Procedural
justice and social identification effects on support for authorities. Journal of
personality and social psychology, 69(3), 482.
Tyler, T. R., & Degoey, P. (1996). Trust in organizational authorities. Trust in
organizations: Frontiers of theory and research (pp. 331–356).
Tzafestas, S. G. (2016). Human-robot social interaction. In Sociorobot world (pp. 53–69).
Springer.
Wang, L., Sakashita, M., Cheng, G., Ji, J., & Zhang, Y. (2021). The effect of regulatory
focus on customer citizenship behavior in a virtual brand community: The role of
online self-presentationand community identification. Journal of Consumer Behavior,
20(3), 607–625. https://doi.org/10.1002/cb.1888
Wanner, J., Herm, L. V., Heinrich, K., & Janiesch, C. (2021). Stop ordering machine
learning algorithms by their explainability! An empirical investigation of the
tradeoff between performance and explainability. In Conference on e-Business, eServices and e-Society (pp. 245–258). Cham: Springer International Publishing.
Waytz, A., Cacioppo, J., & Epley, N. (2010). Who sees human? The stability and
importance of individual differences in anthropomorphism. Perspectives on
Psychological Science, 5(3), 219–232.
Whitley, S. C., Trudel, R., & Kurt, D. (2018). The influence of purchase motivation on
perceived preference uniqueness and assortment size choice. Journal of Consumer
Research, 45(4), 710–724.
Woolley, K., & Risen, J. L. (2018). Closing your eyes to follow your heart: Avoiding
information to protect a strong intuitive preference. Journal of personality and social
psychology, 114(2), 230.

22

M. Yanit

Computers in Human Behavior Reports 18 (2025) 100682

Yan, Y., Fan, W., Shao, B., & Lei, Y. (2022). The impact of perceived control and power
on adolescents’ acceptance intention of intelligent online services. Frontiers in
Psychology, 13, Article 1013436.
Yanco, H. A., & Drury, J. L. (2002). A taxonomy for human-robot interaction. In
Proceedings of the AAAI fall symposium on human-robot interaction (pp. 111–119).
Yanit, M., Yanit, M., & Wan, F. (2023). Right agent, wrong level of hedonism: How high
(vs low) hedonic values in AI-performed tasks lead to decreased perceptions of
humanlikeness, warmth, and less consumer support. Computers in Human Behavior,
147, Article 107870.

Yoganathan, V., Osburg, V. S., Kunz, W. H., & Toporowski, W. (2021). Check-in at the
Robo-desk: Effects of automated social presence on social cognition and service
implications. Tourism Management, 85, Article 104309.
Zhu, D. H., & Chang, Y. P. (2020). Robot with humanoid hands cooks food better? Effect
of robotic chef anthropomorphism on food quality prediction. International Journal of
Contemporary Hospitality Management, 32(3), 1367–1383.
CRiskCo. (n.d.). Artificial intelligence as an ally in credit analysis. CRiskCo. Retrieved
March 27, 2025, from https://www.criskco.com/en/artificial-intelligence-as-an-a
lly-in-credit-analysis/.

23