A Review of Trust in Artificial Intelligence: Challenges, Vulnerabilities and Future Directions

Full text

A Review of Trust in Artificial Intelligence: Challenges, Vulnerabilities and
Future Directions
Steven Lockey
University of Queensland
s.lockey@uq.edu.au
Nicole Gillespie
University of Queensland
n.gillespie@business.uq.edu.au
Daniel Holm
University of Queensland
d.holm@business.uq.edu.au
Ida Asadi Someh
University of Queensland
i.asadi@business.uq.edu.au
Abstract
Artificial Intelligence (AI) can benefit society, but it
is also fraught with risks. Societal adoption of AI is
recognized to depend on stakeholder trust in AI, yet
the literature on trust in AI is fragmented, and little is
known about the vulnerabilities faced by different
stakeholders, making it is difficult to draw on this
evidence-base to inform practice and policy. We
undertake a literature review to take stock of what is
known about the antecedents of trust in AI, and
organize our findings around five trust challenges
unique to or exacerbated by AI. Further, we develop
a concept matrix identifying the key vulnerabilities to
stakeholders raised by each of the challenges, and
propose a multi-stakeholder approach to future
research.
1. Introduction
Artificial Intelligence (AI) is an increasingly
ubiquitous aspect of modern life that has had a
transformative impact on how we live and work [1].
However, despite holding much promise AI has been
implicated in high profile breaches of trust and ethical
standards and concerns have been raised over the use
of AI in initiatives and technologies that could be
inimical to society. For example, AI underpins lethal
autonomous weapons, is central to mass surveillance,
and is subject to racial bias in healthcare.
Trust is vital for AI’s continued social license. The
European Commission's AI High-Level Expert Group
(AI HLEG) highlight that if AI systems do not prove
to be worthy of trust, their widespread acceptance and
adoption will be hindered, and the vast potential
societal and economic benefits will remain unrealized
[2]. While trust has been shown to be important for the
adoption of a range of technologies [3], AI creates an
array of qualitatively different trust challenges
compared to more traditional information technologies
[4]. In response, the AI HLEG provided a set of
guidelines for the development, deployment and use of
trustworthy AI [2]. These guidelines are just one of
many [5].
Research shows that trust is an important predictor
of the willingness to adopt a range of AI systems, from
product recommendation agents [e.g., 6, 7] and AI-
enabled banking [e.g., 8] to autonomous vehicles
(AVs) [e.g., 9, 10]. Given the central role of trust, there
is a strong practical need to understand what
influences and facilitates trust in AI, with multiple
recent calls for research from policymakers [2, 11],
industry [12] and scholars [e.g., 13, 14].
Yet we are only at an early stage of understanding
the antecedents of trust in AI systems. A recent review
of the empirical literature suggests that AI
representation plays an important role in the
development of trust [15] and differentially impacts
trust over time; for robotic AI, trust tends to start low
and increase over time, but for virtual and embedded
AI the opposite commonly occurs. However, it is
difficult however to isolate the antecedents of trust in
this work, as trust was equated with affect [e.g. 16]
attraction to [e.g. 17] and general perceptions of AI
[e.g. 18]. Previous meta-analyses have examined the
antecedents to trust in specific applications of AI, such
as human-robot interaction [19] and automation [20],
but have not taken into account human trust in AI more
broadly.
In this review, we take stock of the scholarly
literature over the past two decades to examine the
antecedents of trust in AI systems. Our review differs
to prior work in four ways: 1) our organization of the
literature around five trust challenges that are unique
to, or exacerbated by, the inherent characteristics of
AI; 2) our focus on articles that operationalize trust in
line with established definitions; 3) a focus on trust in
all forms of AI; and 4) the integration of conceptual
and empirical scholarship.
We contribute to the literature on trust in AI in
three ways. First by synthesizing the fragmented and
interdisciplinary literatures to meaningfully take stock
of what we know about the antecedents of trust in AI.
Second, by developing a concept matrix identifying
the key vulnerabilities for stakeholders raised by each
of the five AI trust challenges. Third, by drawing on
this matrix to identify omissions in current
Proceedings of the 54th Hawaii International Conference on System Sciences | 2021
Page 5463
URI: https://hdl.handle.net/10125/71284
978-0-9981331-4-0
(CC BY-NC-ND 4.0)

understanding and promising directions for future
research.
2. Defining AI and Trust
2.1. Conceptualizing AI
We adopt the OECD's [21] definition of AI, as
recently recommended by AI experts [22]: "a
machine-based system that can, for a given set of
human-defined objectives, make predictions,
recommendations, or decisions influencing real or
virtual environments…AI systems are designed to
operate with varying levels of autonomy".
Most notable advances in AI are driven by machine
learning [23], a subset of AI and can be defined as a
"machine’s ability to keep improving its performance
without humans having to explain exactly how to
accomplish all the tasks it’s given” [34]. A further
subset of machine learning is deep learning, which is
a specialized class of machine learning that is built on
artificial neural networks [25]. Advances in machine
learning and the shift from rule-based to algorithmic
learning exponentially increases the power and
functionality of these systems, enabling more accurate
results than previous iterations. However, they also
change the nature of how IT artifacts are designed and
work [26], their capacity for autonomous functioning,
creating risks, challenges and uncertainties [27] not
inherent in traditional technologies. Trust matters most
under conditions of risk and uncertainty [28, 29].
2.2. Conceptualizing trust
We adapt popular, cross-disciplinary definitions
[30, 31] to define trust as a psychological state
comprising the intention to accept vulnerability based
upon positive expectations of the intentions or
behaviour of another entity (e.g. an AI system).
The two defining components of trust are the
intention to accept vulnerability based on positive
expectations. In positioning their stance on trust in IT
artifacts, McKnight et al. [32, p. 3] note: “trust
situations arise when one has to make oneself
vulnerable by relying on another person or object”.
Trust is only relevant under conditions of risk and
uncertainty, where misplaced trust results in loss or
harm [32]. Examples include relying on an
autonomous vehicle to drive safely, or on the
decision of an AI system to be accurate and unbiased.
Vulnerability is central to trust and captures the ‘leap
of faith’ required to engage with entities under
conditions of risk and uncertainty.
A foundational tenet of trust theory is that this
willingness to be vulnerable should be based on 'good
reasons' [33]. 'Trusting' without good reasons (or
positive expectations) is not trust at all; it amounts to
hope or blind faith. Positive expectations of AI
systems can be based on system-oriented assessments
of functionality, reliability and predictability, and
helpfulness [32]. Hence, there must be some expected
utility or value to accept vulnerability to an AI
system – that is, positive expectations that the system
will be useful, reliable and operate as intended.
Trust theory and research highlights the
importance of understanding the trustor (i.e. who is
doing the trusting), the referent of trust (i.e. what or
whom are they trusting in), and the nature of trusting
(i.e. what are the risks, vulnerabilities or dependence
in the trusting act) [34, 35, 36]. Understanding the
trustor (i.e. the stakeholder) is particularly important
in the context of AI, as it will influence the nature of
the risks and vulnerabilities inherent in trusting an AI
system, and hence the salient cues and antecedents
that influence trust. For example, domain experts are
likely to pay attention to different trust cues than
those that impact end users or customers.
3. Methodology
We conducted an interdisciplinary literature
review using the Web of Science and EBSCO
Business Source Complete databases, searching for
the terms “*trust*” AND “Artificial Intelligence” OR
“Machine Learning” OR “Deep Learning”. Peer-
reviewed journal articles, conference and symposia
papers and proceedings, and book chapters published
since 2000 were included in our review. We further
examined the reference lists of recent review articles
on trust in AI, robots and automation [e.g. 15, 19] and
highly cited papers [e.g. 13] to identify additional
articles that met our inclusion criteria.
We excluded articles that did not address
antecedents of trust in AI, either conceptually or
empirically, and did not meet a commonly accepted
definition or conceptualization of trust (e.g., where
trust was conflated with distinct constructs, such as
emotion or attraction). Reasons for exclusion
included: a focus on computational trust, discussion of
trusts in the financial/legal sense (e.g., trust fund ) or
healthcare (e.g., an NHS trust), articles in which trust
was peripheral rather than central to the article, or in
empirical papers that mention trust but did not
measure it. After this screening process, our search
produced 102 relevant articles.
Our review comprised more empirical (57%) than
conceptual (43%) articles. Most empirical papers were
experimental (47/58 papers), and only one paper used
Page 5464

a mixed-method design. 71% of papers were published
in 2016 or later, and the earliest article in our review
was published in 2005. Articles reflected a diversity of
fields, including information systems, computer
science, ergonomics, business and economics,
psychology, medicine, and law.
4. Literature Review: AI Trust
Challenges
We organize our review by focusing on concepts
related to five central AI trust challenges: 1)
transparency and explainability, 2) accuracy and
reliability, 3) automation, 4) anthropomorphism, and
5) mass data extraction. These five trust challenges
capture the large majority of articles identified by our
review. This approach positions our paper as an
organizing review [37]. For each concept, we first
explain the trust challenge, before synthesizing the
relevant literature.
4.1. Transparency and Explainability
AI is often considered a ‘black box’ [38].
Advanced algorithmic learning methods (such as deep
learning) are inherently not transparent or explainable.
The antidote to this black box is creating AI that can
explain itself, where decisions and predictions are
made transparently. However, there is a tension
between accuracy and explainability, in that models
that perform best tend to be the least transparent and
explainable, while the ones most able to provide the
clearest explanations are the least accurate [39]. There
is an entire field of research dedicated to making AI
more explainable and transparent, with the central aim
of improving user trust [38].
Many articles in our review theorize or empirically
demonstrate that transparency and explainability of AI
applications facilitate trust. In healthcare, scholars
argue that interpretable models that are explainable
and transparent are necessary to enable clinicians to
understand and trust in the outcomes of clinical
support systems [40, 41]. However, full transparency
may be difficult to achieve in practice. Instead,
different levels of transparency can be used based on
factors such as level of risk and the ability of the
clinician to evaluate the decision [41].
Explanations are argued to play a key role in
facilitating trust in AI systems [42], particularly when
the user lacks previous experience with the system.
Researchers propose that system transparency is a key
mitigator of user overtrust, that is trusting AI more
than is warranted by its capabilities [43, 44]. However,
explanations may actually cause overtrust [45] and can
be manipulative [46]. The seminal ‘Copy Machine’
study [47] showed that providing an explanation, even
without a legitimate reason, was effective in
promoting compliance. This is particularly
problematic when the audience of the explanation (e.g.
an end user) diverges from its beneficiary (e.g. a
deploying organization; [46]). System explanations
are problematic when produced alongside incorrect
results, particularly when they seem plausible [45].
Empirical research demonstrates the positive
impact of AI transparency and explainability on trust
[e.g. 48, 49, 50]. Experimental research undertaken in
military settings indicates that when human operators
and AI agents collaborate, increased transparency
enhances trust [48,49]. Explanations have been shown
to increase trust in the results of a product release
planning tool [51].
However, research further suggests that the
relationship between transparency and trust is not
straightforward. For example, in the context of
students interacting with an AI assessment grading
tool, providing procedural transparency about the
fairness of the algorithm was found to buffer the
negative impact of expectation violation on trust [52].
However, providing more transparency related to the
outcome (how the raw grade was calculated) did not
enhance trust, indicating that the type and amount of
transparency matters.
4.2. Accuracy and Reliability
A key trust challenge relates to the accuracy of AI
systems, as inaccurate outcomes can lead to bias,
inequality and harm. AI systems can be configured to
optimize a variety of accuracy metrics, and may have
a high rate of accuracy for certain predictions (e.g.
outcomes for white men), but not others (e.g.
outcomes for minority groups) [53]. A study of
automated facial analysis algorithms demonstrated
this problem; there were significantly more
misclassifications of darker-skinned females than
lighter-skinned males [54]. Hence, relying on
accuracy metrics alone may not be sufficient to garner
trust in AI applications; the fairness of the system is
also relevant [53].
Several experiments show that as the reliance,
accuracy or performance of AI systems decreases, so
does user trust [55, 56]. The timing of a reliability
failure also matters. Unreliable performance early in
one's experience with a system may cause more
significant trust break down than failure later in an
interaction [57]. Moreover, even if an AI agent is
accurate, users may not trust it [58]: they also need to
perceive that it is accurate. For example, teams
engaged in a large, street-based game were regularly
Page 5465

mistrustful of the (entirely accurate) information
provided by automated advice, and often chose to
ignore it, despite being told that following the
information was vital for them to progress in the game
[58].
However, other research suggests that even though
inaccurate agent behaviour negatively impacts
perceived trustworthiness, this does not necessarily
translate into reduced compliance: users may still
follow instructions from an AI system they believe is
untrustworthy [59]. Taken together, while most
research indicates a positive influence of accuracy on
trust, the relationship is not straightforward and
warrants further research.
4.3. Automation versus augmentation
Automation enables machines to complete tasks
without direct human involvement [60]. Normative
prescriptions tend to advise organizations to prioritize
augmentation – human collaboration with machines to
perform a task - over automation. Yet there is an
argument that such neat delineation is not realistic and
that an automation-augmentation paradox exists [60].
As an example, domain experts may work with an AI
system to determine and codify appropriate variables
(augmentation), and the system may then be
automated based on these criteria. However, if
conditions change over time, a further stage of
augmentation will be necessary. This brings into
question the role of the domain expert and the potential
for their role in the augmentation process to ultimately
lead to the automation of their own work.
The impact of automated AI on trust in high-risk
contexts has been conceptually discussed. In
healthcare, there are concerns that AI may disrupt the
bond of trust between doctors and patients [61], and
patients may be more skeptical of automated advice
than advice provided by a doctor [62]. A ‘doctor-in-
the-loop’ approach, in which the doctor both provides
tacit knowledge to AI systems and is the final authority
on decisions proposed by the AI systems, has been
proposed to address these concerns [63]. This
‘augmentation over automation’ approach has
received empirical support. A suite of experiments
found a reluctance to use medical care delivered by AI
providers, except when the AI was used to support the
human provider’s decision, rather than replacing the
human [64]. This ‘human-in-the-loop’ approach has
also been proposed for AI in financial services [65].
Adaptive automation, where automation is not
fixed at the design stage but rather adapts to the
situation, increased trust in a robot during a
collaborative task to a greater extent than when there
was either no automation or static automation.
A concern related to automated AI is the potential
for deskilling if domain experts over-rely on
automated systems [67, 68]. One study found financial
investors trust fully automated artificial advisors more
than human advisors [69]. However other research
indicates that AI over-reliance on AI systems tends to
be experienced by novices; experts are generally less
willing to trust AI systems [70, 71].
4.4. Anthropomorphism and embodiment
Anthropomorphism involves the inclusion of
human-like characteristics into an AI’s design. It has
been theorized that the more human-like an AI agent
is, the more likely humans are to trust and accept it
[72]. However, there are concerns that over-
anthropomorphism may lead to overestimation of the
AI’s capabilities, potentially putting the stakeholder at
risk [73], damaging trust [74], and leading to a host of
ethical and psychological concerns, including
manipulation [75].
Empirical findings broadly support the proposition
that anthropomorphism increases trust in AI. This has
been shown in the context of autonomous vehicles
[72,76], with people demonstrating more trust in AVs
with human features than without [72], as well as in
the context of virtual agents [e.g. 9, 77].
Research into the buffering impact of virtual agent
human-likeness on decreasing reliability found that
although anthropomorphism decreased initial
expectations, it increased trust resilience. When
performance deteriorated, decreases in trust were more
pronounced in a machine-like agent than an
anthropomorphic agent. Embodiment of virtual agents
(i.e. having a physical form) also increases user trust
in the agent, primarily through perceptions of its social
presence [9, 77, 78]. Research also indicates that
augmented reality and 3D agents were perceived as
more trustworthy than those in traditional 2D
interfaces [79].
However, not all empirical work suggests that
anthropomorphism leads to stronger perceptions of
trust. For example, a study investigating the
anthropomorphism of a care robot found that a highly
human-like robot was perceived as less trustworthy
and empathetic than a more machine-like robot [62].
Further research is required to understand when and
how AI anthropomorphism enhances trust, and what
moderates this relationship.
4.5. Mass Data Extraction
AI systems, particularly advanced algorithmic
learning systems, require the extraction and processing
of large amounts of data to function, making them
Page 5466

qualitatively different from traditional IT artifacts
[81]. Data extraction is fundamentally different from
previous forms of market exchange, as it connotes a
one-way taking of data rather than a consensual or
reciprocal process [82].
The trust challenge around data extraction is
primarily around issues of privacy. For end users, loss
of privacy and inappropriate sharing of information is
a concern, and can result in reduced self-
determination. These vulnerabilities can scale to the
societal level to the point where expectations of
privacy as a societal norm may be lost. Indeed,
Facebook CEO Mark Zuckerberg has explicitly stated
that privacy is no longer a social norm [83]. This
proposition is clearly contentious, as privacy is
considered and codified as a fundamental human right
in several democracies, and people usually express
that they value privacy, even if they do not always
demonstrate this proposition in their behavior [84].
Some jurisdictions have taken regulatory
approaches to tackling concerns about big data
extraction, with the European Commission’s General
Data Protection Regulation (GDPR) aiming to give
European residents control over their personal data
through requirement of ‘Privacy by Design’ [85].
While this type of legislation may reduce privacy-
related vulnerabilities of end-users and society, it
introduces a new set of vulnerabilities for domain
experts, who are responsible for ensuring data privacy
and accountable for appropriate data use under threat
of large fines
Research on data extraction and the privacy
concerns that underpin it has been primarily
conceptual. Scholars note big data extraction is an
ethical dilemma in the development and use of AI-
enabled medical systems [62, 86], virtual agents [87]
and smart cities [88]. One solution to ensure citizen
privacy and promote trust is creating an environment
in which data analysis can occur without allowing
organizations to extract the data [88].
The limited empirical work in this area has focused
on the interaction between privacy and trust. For
example, when people have few privacy concerns
related to autonomous vehicles collecting passenger
location information and being used as a conduit for
surveillance, they were more likely to trust in the
autonomous vehicle [89].
Interestingly, a study of virtual agent embodiment
found that participants were more willing to share
private data with an AI agent and more confident that
the agent would respect their privacy when it could
move around naturally and speak compared with a
static agent that could speak [77].
4.6. The Role of Governance in Addressing
AI Trust Challenges
In addition to the five trust challenges, our review
identified two broad, generic mechanisms for
overcoming these trust challenges: familiarity and
governance. Empirical studies indicate that familiarity
and experience engaging with AI systems facilitates
trust [90, 91]. Conceptual work argues that governance
–in the form of appropriate controls to ensure
trustworthy AI development and deployment - is a
necessary condition for trust in AI [e.g. 92, 93]. A
recent national survey identified beliefs about the
adequacy of AI regulation and governance to be the
strongest predictor of trust in AI systems [94]. It may
be more important and efficient to make AI systems
verifiably trustworthy via appropriate governance
rather than seek explanations for specific outcomes
[45]. Governance that encourages collaboration
among key stakeholders, supports the recognition and
removal of bias, and clarifies the appropriate control
over and use of personal information has been
proposed to enhance trust [92]. However, this work
further notes that AI development remains largely
unregulated to date [95], despite public expectation of
AI regulation [94; 96].
5. Discussion and Future Directions
Our review demonstrates that research on the
antecedents of trust in AI can largely be organized
around five key trust challenges that are unique to, or
exacerbated by, the inherent features of AI. Each of
these trust challenges raises a set of vulnerabilities or
risks for stakeholders of AI systems. In Table 1, we
present a concept matrix mapping the key
vulnerabilities associated with each of the five trust
challenges for three AI stakeholder groups – domain
experts, end users, and society. These stakeholders are
each central to the acceptance and uptake of AI
systems.
As shown in Table 1, the use of AI systems open
up (potential or actual) risks and vulnerabilities for
each of these stakeholders, making trust a highly
salient and pertinent concern. Our concept matrix
shows that the vulnerabilities experienced in relation
to an AI system depend on the stakeholders’ role
which determines how they interact with, are
responsible for, or are impacted by the AI systems. In
the next section, we discuss the key vulnerabilities
domain experts, end users and society more broadly
experience in relation to each AI trust challenge, and
how these differ across these stakeholder groups.
Page 5467

Table 1: Concept matrix of the five AI trust challenges and the respective vulnerabilities each creates for
stakeholders
AI trust challenge
Stakeholder vulnerabilities
Domain expert
End user
Society
1. Transparency and
explainability
•Ability to know and explain AI
output, and provide human
oversight
•Manipulation from erroneous
explanations
•Ability to understand how
decisions affecting them are
made
•Ability to provide meaningful
consent and exercise agency
•Knowledge asymmetries
•Power imbalance and
centralization
•Scaled disempowerment
2. Accuracy and
reliability
•Accountability for accuracy and
fairness of AI output
•Reputational and legal risk
•Inaccurate / harmful outcomes
•Unfair / discriminatory
treatment
•Entrenched bias / inequality
•Scaled harmed to select
populations
3. Automation
•Professional over-reliance and
deskilling
•Loss of expert oversight
•Loss of professional identity
•Loss of work
•Loss of dignity (humans as data
points; de-contextualization)
•Loss of human engagement
•Over-reliance and deskilling
•Scaled deskilling
•Reduced human connection
•Scaled technological
unemployment
•Cascading AI failures
4. Anthropomorphism
and embodiment
•Professional over-reliance
•Psychological wellbeing
•Manipulation through
identification
•Over-reliance and over-sharing
•Manipulation through
identification
•Human connection and identity
5. Mass data
extraction
•Accountability for privacy and
use of data
•Reputational and legal risk
•Personal data capture and loss
of privacy
•Inappropriate re-identification
and use of personal data
•Loss of control
•Inappropriate use of citizen data
•Mass surveillance
•Loss of societal right to privacy
•Power imbalance & societal
disempowerment
Domain experts. Domain experts in deploying
organizations are those with the expert knowledge and
experience in the field of application of the AI system.
For example, doctors in relation to AI-enabled medical
diagnosis applications. Domain expert knowledge can
be used to create codified information used to train AI
systems, meaning they have a role in system input.
Domain experts also work with system outputs, as they
use and interface with AI systems for service delivery.
Key vulnerabilities faced by domain experts relate
to professional knowledge, skills, identity, and
reputation. For example, research suggests that
automation through AI may lead to deskilling [67, 68].
A related vulnerability stemming from the AI
explanability challenge is the ability of the domain
expert to understand the AI system and be able to
explain and justify decisions to other stakeholders,
particularly when AI system outputs are used in
service delivery (e.g. clinical decision making
systems). Anthropomorphism may further threaten the
professional identity of domain experts and cause
over-reliance on human-like agents. The reputational
damage and legal risks from inaccurate or unfair
results, or inappropriate data use, sharing or privacy
breach, place a further burden on accountable domain
experts.
End users. End users are those directly influenced
by the output or decisions made by the AI system.
They are vulnerable to any problems, inaccuracies or
biases within the system. More broadly, end users face
vulnerabilities around understanding how AI-based
decisions are made, which can lead to diminished
ability to provide meaningful consent, identify unfair
or unethical impact, and exercise agency. Using the
context of AI in personal insurance as an example,
companies purportedly draw on thousands of data
points to judge the risk of someone making a motor
insurance claim, including whether they drink tap or
bottled water [97]. Understanding exactly how such a
decision was made is impossible for an average
customer, and highlights vulnerabilities around
explainability, data capture and loss of privacy related
to data extracted without consent. Further, AI can be
used to ‘nudge’ customer behavior in a way that is
manipulative and intrusive [97]. Concerns have been
raised that the combination of these vulnerabilities
may lead to the loss of human dignity, and lack of
consideration of personal circumstances, effectively
reducing humans to a series of data points. This is
particularly problematic for underrepresented,
marginalised users.
Society. The focus here is on vulnerabilities that
impact society as a whole, and this stakeholder group
includes regulators. Vulnerabilities at the societal level
include knowledge asymmetry, power centralization
and the potential for cascading AI failures. For
Page 5468

instance, knowledge asymmetry between big tech
companies, policymakers and citizens may result in a
continuous cycle of outdated or ineffective regulation
[98]. Internet giants at the forefront of AI development
and mass data extraction activities have already
amassed a unique concentration of power [99]. The
scaled use of inaccurate, biased or privacy invading AI
technologies on citizens can entrench bias, inequality
and undermine human rights, such as the right to
privacy.
5.1 A multi-stakeholder perspective on trust
in AI
Our concept matrix outlines the varying
vulnerabilities of key stakeholder groups in relation to
AI systems. Accepting vulnerability is a key element
of trust and understanding and mitigating the risks and
vulnerabilities AI systems pose for stakeholders, is
central to facilitating trust and building the confident
positive expectations that it is founded on. Given this
we propose future research take a multi-stakeholder
approach to examining the antecedents of trust in AI.
Prior research has shown that stakeholders’
varying vulnerabilities in trusting another entity
influence the salience and importance of the cues and
antecedents that inform trust [35]. Understanding the
vulnerabilities and expectations of different
stakeholders of complex socio-technical systems is
also important [100] because stakeholder alignment
facilitates trust in firms seeking to innovate with AI
[101].
However, as shown in our review, much of the
research to date has focused on a single stakeholder,
usually an individual end user or domain expert. A
reason for this may be that most empirical research on
the antecedents of trust in AI is experimental, and
places participants either as quasi-users or a non-
specific stakeholder role. Further, trusting behavior,
and the antecedents that influence it, may be different
in an experimental setting than in the field due to the
varying risks, vulnerabilities and trust cues. For
example, it is likely people will behave differently in
an autonomous vehicle on the road than in a ‘safe’
driving simulator.
Moving forward, we see field experiments and
longitudinal case studies examining multiple
stakeholders of an AI system, as fruitful
methodological approaches to deepen understanding
of the antecedents of stakeholder trust in AI systems.
Undertaking longitudinal case studies has the
advantage of providing holistic, contextualised
insights into the development of trust in AI systems
over time. This is likely to provide a more systemic
understanding of hitherto underexplored areas such as
how stakeholder groups converge and diverge in
relation to their vulnerabilities, expectations and trust
in AI.
It is evident from our review that although several
trust challenges have been raised, many have not been
examined empirically, and few have been examined
from the perspective of multiple stakeholders, or the
perspective of society as a stakeholder.
Furthermore empirical studies have tended to
examine whether a concept (such as accuracy or
anthropomorphism) enhances trust, yet high trust is
not always appropriate, and encouraging people to
trust without ‘good reasons’ [33] can be manipulative.
This tension is particularly apparent in studies of
explainability and transparency, and
anthropomorphism. For instance, people can misplace
trust in inaccurate AI systems when provided an
explanation [46], even nonsensical explanations [47],
and anthropomorphism can lead people to believe that
an agent is competent, even in the face of limited
‘good reasons’ [73]. Broadly, these issues can lead to
overtrust and consequent problems. Further research is
required to understand what influences stakeholders to
trust ‘optimally’, that is in a well calibrated manner
that aligns with actual evidence of trustworthiness and
effective AI design that mitigates and minimizes the
likelihood of harmful consequences [102].
6. References*
[1] K. Grace et al., “When will AI exceed human
performance? Evidence from AI experts”, Journal of
Artificial Intelligence Research, 62, 2018, pp.729-754.
[2] AI HLEG, “Ethics Guidelines for Trustworthy AI”,
European Commission, 2018. Retrieved from
https://ec.europa.eu/
[3] M. Söllner et al., “Trust”, in MIS Quarterly Research
Curations, Ashely Bush and Arun Rai, Eds.,
http://misq.org/research-curations, October 31, 2016.
[4] S. Makridakis. “The forthcoming Artificial Intelligence
(AI) revolution: Its impact on society and firms”,
Futures 90, 2017, pp. 46-60.
[5] Algorithm Watch, “AI Ethics Guidelines Global
Inventory”, 2018. Retrieved from
https://algorithmwatch.org/en/project/ai-ethics-
guidelines-global-inventory/
[6] S.Y. Komiak and I. Benbasat, “The effects of
personalization and familiarity on trust and adoption
of recommendation agents”, MIS Quarterly, 2006, pp.
941-960.
[7] L. Qiu and I. Benbasat, “Evaluating anthropomorphic
product recommendation agents: A social relationship
perspective to designing information systems”, Journal
of Management Information Systems, 25(4), 2009, pp.
145-182.
[8] E.M Payne et al., “Mobile banking and AI-enabled
mobile banking”, Journal of Research in Interactive
Marketing, 12(3), 2018, pp. 328-346.
Page 5469

[9] I. Panagiotopoulos and G. Dimitrakopoulos, “An
empirical investigation on consumers’ intentions
towards autonomous driving”, Transportation
Research Part C: Emerging Technologies, 95, 2018,
pp. 773-784.
[10] T. Zhang et al.,“The roles of initial trust and perceived
risk in public’s acceptance of automated vehicles”,
Transportation Research Part C: Emerging
Technologies, 98, 2019, pp. 207-220.
[11] US Chamber of Commerce, “US Chamber of
Commerce Principles on Artificial Intelligence”, 2019.
Retrieved from https://www.uschamber.com
[12] IEEE Global Initiative on Ethics of Autonomous and
Intelligent Systems, “Ethically Aligned Design: A
Vision for Prioritizing Human Well-being with
Autonomous and Intelligent Systems”, 2019.
Retrieved from https://standards.ieee.org/
[13] M.T. Ribeiro et al., “Why should I trust you?:
Explaining the predictions of any classifier”,
Proceedings of the 22nd ACM SIGKDD international
conference on knowledge discovery and data mining,
September 2016, pp. 1135-1144.
[14] K. Siau and W. Wang, “Building trust in artificial
intelligence, machine learning, and robotics”, Cutter
Business Technology Journal, 31(2), 2018, pp. 47-53.
[15] E. Glikson and A.W. Woolley, “Human trust in
Artificial Intelligence: Review of empirical research”,
Academy of Management Annals, advanced online
publication, 2020.
[16] T. Zhang et al., “Service robot feature design effects
on user perceptions and emotional responses”,
Intelligent service robotics, 3(2), 2010, pp.73-88.
[17] T.W. Bickmore et al.,, “Tinker: a relational agent
museum guide”, Autonomous agents and multi-agent
systems, 27(2), 2013, pp. 254-276.
[18] K.S. Haring et al., “The influence of robot appearance
and interactive ability in HRI: a cross-cultural study”,
In International conference on social robotics, 2016,
pp. 392-401. Springer, Cham.
[19] P.A. Hancock et al., “A meta-analysis of factors
affecting trust in human-robot interaction”, Human
Factors, 53(5), 2011, pp.517-527.
[20] K.E. Schaefer et al, “A meta-analysis of factors
influencing the development of trust in automation:
Implications for understanding autonomy in future
systems”. Human Factors, 58(3):, 2016, pp. 377-400.
[21] OECD, “Artificial Intelligence on Society”, OECD
Publishing, Paris, 2019. Retrieved from
https://www.oecd-ilibrary.org/.ors, 58(3), 2016, 377-
400.
[22] P.M. Krafft et al., “Defining AI in Policy versus
Practice”, Proceedings of the AAAI/ACM Conference
on AI, Ethics, and Society, 2020, pp. 72-78.
[23] M.I. Jordan and T.M. Mitchell, “Machine learning:
Trends, perspectives, and prospects”. Science,
349(6245), 2015, pp. 255-260.
[24] E. Brynjolfsson and A. Mcafee “The business of
artificial intelligence”. Harvard Business Review, July
2017, pp. 1-20.
[25] Y. LeCun et al., “Deep learning”. Nature, 521(7553),
2015, p. 436-444.
[26] I. Rahwan et al., “Machine behaviour”, Nature,
568(7753), 2019, pp. 477-486.
[27] Y.K. Dwivedi et al., “Artificial Intelligence (AI):
Multidisciplinary perspectives on emerging
challenges, opportunities, and agenda for research,
practice and policy”, International Journal of
Information Management, 101994, 2019, pp.
[28] P. Ping Li, “When trust matters the most: The
imperatives for contextualizing trust research”. Journal
of Trust Research 2(2), 2012, pp. 101-106.
[29] M. Deutsch, “The effect of motivational orientation
upon trust and suspicion”. Human Relations,13(2),
1960, pp. 123-139.
[30] D. Gefen et al., “Trust and TAM in online shopping:
An integrated model”, MIS Quarterly, 27(1), 2003, pp.
51-90.
[31] D.M. Rousseau, et al., “Not so different after all: A
cross-discipline view of trust”, Academy of
Management Review, 23(3), 1998, pp. 393-404.
[32] D.H. McKnight et al., “Trust in a specific technology:
An investigation of its components and measures”,
ACM Transactions on management information
systems (TMIS), 2(2), 2011, pp. 1-25.
[33] J.D Lewis and A. Weigert, “Trust as a social reality”.
Social Forces, 63(4), 1985, pp. 967-985.
[34] C.A. Fulmer and M.J. Gelfand, “At what level (and in
whom) we trust: Trust across multiple organizational
levels”, Journal of Management, 38(4), 2012, pp.
1167-1230.
[35] M. Pirson and D. Malhotra, “Foundations of
organizational trust: What matters to different
stakeholders?”, Organization Science., 22(4), 2011,
pp.1087-1104.
[36] R.C. Mayer et al., “An integrative model of
organizational trust”, Academy of Management
Review, 20(3), 1995. pp. 709-734.
[37] D.E. Leidner, “Review and theory symbiosis: An
introspective retrospective”, Journal of the Association
for Information Systems. 19(6), 2018, Article 1.
[38] A. Adadi and M. Berrada, “Peeking inside the black-
box: A survey on Explainable Artificial Intelligence
(XAI)”, IEEE Access, 6, 2018, pp. 52138-52160.
[39] A. Holzinger, et al., “What do we need to build
explainable AI systems for the medical domain?”,
arXiv preprint arXiv1712.09923, 2017.
[40] R. Elshawi et al., “On the interpretability of machine
learning-based model for predicting hypertension”,
BMC medical informatics and decision making, 19(1),
2019, pp. 1-32.
[41] C. Gretton, “Trust and Transparency in Machine
Learning-Based Clinical Decision Support”. In J.
Zhou and F. Chen (eds.), Human and Machine
Learning, 2018, pp. 279-292. Springer, Cham.
[42] P. Andras et al., (2018) “Trusting Intelligent
Machines: Deepening Trust Within Socio-Technical
Systems”. IEEE Technology and Society Magazine,
37(4), 2018, pp. 76-83.
[43] V. Hollis, et al., “On being told how we feel: how
algorithmic sensor feedback influences emotion
perception”. Proceedings of the ACM on Interactive,
Page 5470

Mobile, Wearable and Ubiquitous Technologies, 2(3),
2018, pp. 1-31.
[44] A. R. Wagner et al., “Overtrust in the robotic age”.
Communications of the ACM, 61(9), 2018, pp. 22-24.
[45] J. A. Kroll, “The fallacy of inscrutability”.
Philosophical Transactions of the Royal Society A:
Mathematical, Physical and Engineering Sciences,
376(2133), 2018, 20180084.
[46] A. Weller, “Challenges for transparency”, arXiv
preprint arXiv,1708.01870, 2017.
[47] E.J. Langer et al., “The mindlessness of ostensibly
thoughtful action: The role of" placebic" information in
interpersonal interaction”, Journal of Personality and
Social Psychology, 36(6), 1978, pp. 635-642.
[48] J. Y. Chen, “Human-autonomy teaming and agent
transparency”, Companion Publication of the 21st
International Conference on Intelligent User Interfaces,
March 2016, pp. 28-31.
[49] C. S. Calhoun, et al., (2019). “Linking precursors of
interpersonal trust to human-automation trust: An
expanded typology and exploratory experiment”,
Journal of Trust Research, 9(1), 2019, pp. 28-46.
[50] E. S. Vorm, “Assessing Demand for Transparency in
Intelligent Systems Using Machine Learning”, 2018
Innovations in Intelligent Systems and Applications,
July 2018, pp. 1-7.
[51] G. Du and G. Ruhe, “Does explanation improve the
acceptance of decision support for product release
planning?”, In 2009 3rd International Symposium on
Empirical Software Engineering and Measurement,
October 2009, pp. 56-68. IEEE.
[52] R. F. Kizilcec, “How much information? Effects of
transparency on trust in an algorithmic interface”,
Proceedings of the 2016 CHI Conference on Human
Factors in Computing Systems, May 2016, pp. 2390-
2395.
[53] H. Wallach, “Computational social science≠ computer
science+ social data”, Communications of the ACM,
61(3), 2018, pp. 42-44.
[54] J. Buolamwini and T. Gebru, “Gender shades:
Intersectional accuracy disparities in commercial
gender classification”. Conference on Fairness,
Accountability and Transparency, January 2018, pp.
77-91.
[55] M. Yin, et al., “Understanding the effect of accuracy
on trust in machine learning models”, In Proceedings of
the 2019 chi conference on human factors in computing
systems May 2019, pp. 1-12).
[56] E. J. de Visser et al.,. “Learning from the slips of
others: Neural correlates of trust in automated agents”.
Frontiers in Human Neuroscience, 12(309), 2018, pp.
1-15.
[57] A. Freedy, et al., “Measurement of trust in human-
robot collaboration”. 2007 International Symposium on
Collaborative Technologies and Systems, May 2007,
pp. 106-114.
[58] S. Moran et al., “Team reactions to voiced agent
instructions in a pervasive game”, In Proceedings of the
2013 international conference on Intelligent user
interfaces, May 2013, pp. 371-382.
[59] M. Salem et al., “Would you trust a (faulty) robot?
Effects of error, task type and personality on human-
robot cooperation and trust”, 10th ACM/IEEE
International Conference on Human-Robot Interaction,
March 2015, pp. 1-8.
[60] S. Raisch and S. Krakowski, “Artificial Intelligence
and Management: The Automation-Augmentation
Paradox”, Academy of Management Review. 2020, in
press.
[61] V. Diebolt et al., “Artificial intelligence: Which
services, which applications, which results and which
development today in clinical research? Which impact
on the quality of care? Which recommendations?”
Therapie, 74(1), 2019, pp. 155-164.
[62] K.T. Chui et al., “Big data and IoT solution for patient
behaviour monitoring”, Behaviour & Information
Technology, 38(9), 2019, pp. 940-949.
[63] A.C. Valdez et al., “Recommender systems for health
informatics: state-of-the-art and future perspectives”, In
Machine Learning for Health Informatics, 2016, pp.
391-414. Springer, Cham.
[64] C. Longoni et al., “Resistance to medical artificial
intelligence”. Journal of Consumer Research, 46(4),
2019, pp. 629-650.
[65] A. Lui A and G.W. Lamb, “Artificial intelligence and
augmented intelligence collaboration: regaining trust
and confidence in the financial sector”, Information &
Communications Technology Law., 27(3), 2018, pp.
267-283.
[66] E. de Visser E and R. Parasuraman, “Adaptive aiding
of human-robot teaming: Effects of imperfect
automation on performance, trust, and workload”,
Journal of Cognitive Engineering and Decision
Making, 5(2), 2011, p.209-231.
[67] T. Rinta-Kahila et al., “Consequences of
Discontinuing Knowledge Work Automation-Surfacing
of Deskilling Effects and Methods of Recovery”, In
Proceedings of the 51st Hawaii International
Conference on System Sciences, January 2018, pp.
5244-5253.
[68] S.G. Sutton et al., “How much automation is too
much? Keeping the human relevant in knowledge
work”, Journal of Emerging Technologies in
Accounting, 15(2): 2018, pp. 15-25.
[69] A.K. Heid, “Trust in Homo Artificialis: Evidence from
Professional Investors”, 2018 American Accounting
Association Annual Meeting, August 2018,
Washington USA.
[70] X. Fan X et al., “The influence of agent reliability on
trust in human-agent collaboration”, In Proceedings of
the 15th European conference on Cognitive
ergonomics: the ergonomics of cool interaction,
January 2008, pp. 1-8.
[71] J. M. Logg et al., “Algorithm appreciation: People
prefer algorithmic to human judgment”. Organizational
Behavior and Human Decision Processes, 151, 2019,
90-103.
[72] A. Waytz, et al., “The mind in the machine:
Anthropomorphism increases trust in an autonomous
vehicle”. Journal of Experimental Social Psychology,
52, 2014, pp. 113-117.
Page 5471

[73] K. E Culley and P. Madhavan, “A note of caution
regarding anthropomorphism in HCI agents”.
Computers in Human Behavior, 29(3), 2013, pp. 577-
579.
[74] A. L. Baker et al., “Toward an Understanding of Trust
Repair in Human-Robot Interaction: Current Research
and Future Directions”. ACM Transactions on
Interactive Intelligent Systems, 8(4), 2018, Artilce 30.
[75] A. Salles et al., “Anthropomorphism in AI”, AJOB
Neuroscience, 11(2), 2020, pp. 88-95.
[76] F.M. Verberne et al., “Trusting a virtual driver that
looks, acts, and thinks like you”, Human factors, 57(5),
2015, pp. 895-909.
[77] K. Kim et al., “Does a digital assistant need a body?
The influence of visual embodiment and social
behavior on the perception of intelligent virtual agents
in AR”, In 2018 IEEE International Symposium on
Mixed and Augmented Reality (ISMAR), October
2018, pp 105-114. IEEE.
[78] W.A. Bainbridge et al., “The effect of presence on
human-robot interaction’, InRO-MAN 2008 – The 17th
IEEE Internationhal Symposium on Robot and Human
Interactive Communication, August 2008, pp. 701-706.
IEEE.
[79] B. Huynh et al., “A study of situated product
recommendations in augmented reality”, In 2018 IEEE
International Conference on Artificial Intelligence and
Virtual Reality (AIVR), December 2018, pp. 35-43.
IEEE
[80] J. Złotowski et al., “Appearance of a robot affects the
impact of its behaviour on perceived trustworthiness
and empathy”, Paladyn, Journal of Behavioral
Robotics, 7, 2016, 55-66.
[81] S. Gupta et al., “Big data with cognitive computing: A
review for the future”, International Journal of
Information Management, 42, 2018, pp. 78-89.
[82] S. Zuboff, “Big other: surveillance capitalism and the
prospects of an information civilization”, Journal of
Information Technology, 30(1), 2015, pp. 75-89.
[83] E. Osnos, “Can Mark Zuckerberg fix Facebook before
it breaks democracy?”, The New Yorker, 10 September
2018, Retrieved from
https://www.newyorker.com/magazine/2018/09/17/can-
mark-zuckerberg-fix-facebook-before-it-breaks-
democracy.
[84] S. Kokolakis, “Privacy attitudes and privacy
behaviour: A review of current research on the privacy
paradox phenomenon. Computers & Security, 64, 2017,
122-134.
[85] J. Andrew & M. Baker, “The General Data Protection
Regulation in the Age of Surveillance Capitalism”,
Journal of Business Ethics, in press.
[86] J. Vallverdú and D. Casacuberta, “Ethical and
technical aspects of emotions to create empathy in
medical machines”, In Machine Medical Ethics, 2015,
pp. 341-362. Springer, Cham.
[87] I.R. Kerr and M. Bornfreund, “Buddy bots: How
turing's fast friends are undermining consumer
privacy”, Presence: Teleoperators & Virtual
Environments, 14(6), 2005, pp. 647-655.
[88] T. Braun, B.C. Fung, F. Iqbal and B. Shah, “Security
and privacy challenges in smart cities”, Sustainable
cities and society, 39, 2018, pp. 499-507.
[89] K. Kaur and G. Rampersad, “Trust in driverless cars:
Investigating key factors influencing the adoption of
driverless cars”. Journal of Engineering and
Technology Management, 48, 2018, pp. 87-96.
[90] S. Reig et al., “A field study of pedestrians and
autonomous vehicles”, In Proceedings of the 10th
international conference on automotive user interfaces
and interactive vehicular applications. September 2018,
pp. 198-209.
[91] N.L. Tenhundfeld et al.,”Calibrating trust in
automation through familiarity with the autoparking
feature of a Tesla Model X”, Journal of Cognitive
Engineering and Decision Making, 13(4), 2019, pp.
279-294.
[92] C.W. Ho et al., “Governance of automated image
analysis and artificial intelligence analytics in
healthcare”, Clinical Radiology, 74(5), pp. 329-337.
[93] A.F. Winfield and M. Jirotka, “Ethical governance is
essential to building trust in robotics and artificial
intelligence systems”, Philosophical Transactions of the
Royal Society A: Mathematical, Physical and
Engineering Sciences, 376(2133), 2018, 20180085.
[94] S. Lockey et al., “Trust in Artificial Intelligence:
Australian Insights”, The University of Queensland and
KPMG Australia, October 2020.
[95] M. Guihot et al., “Nudging robots: Innovative
solutions to regulate artificial intelligence”, Vand. J.
Ent. & Tech. L., 20, 2017, p3p. 385-456.
[96] B. Zhang and A. Dafoe, “Artificial intelligence:
American attitudes and trends”. Available at SSRN
3312874. 2019.
[97] Centre for Data Ethics and Innovation, “AI and
Personal Insurance”, CDEI Snapshot Series, September
2019.
[98] X. Xhang, “Information Asymmetry and Uncertainty
in Artificial Intelligence”, Medium, 9 September 2017,
Retrieved from
https://medium.com/@zhxh/information-asymmetry-
and-uncertainty-in-artificial-intelligence-ad8e444c4d9a
[99] P. Nemitz, “Constitutional democracy and technology
in the age of artificial intelligence”, Philosophical
Transactions of the Royal Society A: Mathematical,
Physical and Engineering Sciences, 376(2133), 2018,
20180089.
[100] S.A. Wright and A.E. Schultz, “The rising tide of
artificial intelligence and business automation:
Developing an ethical framework”, Business Horizons,
61(6), 2018, pp. 823-832.
[101] M. Hengstler et al., “Applied artificial intelligence
and trust—The case of autonomous vehicles and
medical assistance devices”. Technological Forecasting
and Social Change, 105, 2016, pp. 105-120.
[102]E.J. de Visser et al., “Towards a Theory of
Longitudinal Trust Calibration in Human-Robot
Teams”, International Journal of Social Robotics, 12,
2020, pp. 459-478.
*Due to space constraints, we present selected references
from our review. Contact SL for a full reference list.
Page 5472