=Paper=
{{Paper
|id=Vol-3906/paper1_SCRITA
|storemode=property
|title=Old wine in new bottles: Are agent-specific trustworthiness measures necessary? (short paper)
|pdfUrl=https://ceur-ws.org/Vol-3906/paper1_SCRITA.pdf
|volume=Vol-3906
|authors=Connor Esterwood,Samia Cornelius Bhatti,Lionel P. Robert
|dblpUrl=https://dblp.org/rec/conf/ro-man/EsterwoodBR24
}}
==Old wine in new bottles: Are agent-specific trustworthiness measures necessary? (short paper)==
https://ceur-ws.org/Vol-3906/paper1_SCRITA.pdf
Old Wine in New Bottles: Are Agent-Specific
Trustworthiness Measures Necessary?
Connor Esterwood*, Samia Cornelius Bhatti and Lionel P. Robert
University of Michigan School of Information, Ann Arbor, United States of America.
Abstract
A vital aspect of human–robot interaction (HRI) is trustworthiness. Measuring trustworthiness in the context of
HRI, however, presents new challenges. One such challenge is the ongoing debate between whether the type
of agent examined requires the use of a trustworthiness measure specific to that agent. This paper presents
both sides of this debate and argues that there is no compelling reason to consider one of these measures more
appropriate to a specific agent over another until evidence suggests otherwise.
Keywords
Trust, Measures, Human–Robot Interaction, Methods and Metrics
1. Introduction
Trust in robots is central to their successful integration into our daily lives [1]. Humans rely on
trust to negotiate social interactions with such technologies [2]. Understa the role of trust in social
relationships with robots is crucial today, where the deployment of social robots in organizations is
becoming increasingly common.
To date, no universally accepted definition of trust exists in the field of human-robot interaction
(HRI). One forming consensus, however, is that trust is based on one’s positive expectations or attitudes
towards a given trustee [3, 4, 5]. The higher order concept of trustworthiness often represents attributes
that determine how much someone or something should be trusted. In this way, trustworthiness
precedes and ultimately determines trust [3, 6, 1].
Measuring the trustworthiness of agents, such as robots, presents new challenges. Previous research
often distinguishes between trustworthiness in humans and technology [7, 8]. However, advanced
automation is designed to blur these boundaries [2]. Consequently, many studies have adapted measures
validated for assessing humans’ trustworthiness in robots [1]. These measures, however, were originally
designed solely for humans, casting doubt on their reliability when applied to robots.
In response, measures have been developed specifically for advanced automation, such as robots
[8, 4]. Yet these measures seem to capture similar, if not the same, attributes found in measures of
human trustworthiness. This begs the question of whether such specific measures for robots are needed.
No clear empirical evidence suggests that measures for humans or robots are less effective when used
on the other. Without such empirical evidence, the debate remains speculative and driven by intuition
rather than solid data.
This paper argues that there is no compelling reason to prefer one set of trustworthiness measures
(for humans versus automation) over another until evidence suggests otherwise. This debate risks
repackaging old ideas in new forms, with significant implications for research practices and theories. If
distinct measures are necessary, it becomes crucial to adopt them and assess the extent to which prior
research may have been compromised. Conversely, if distinct measures are not needed, researchers can
confidently proceed with meta-analyses. From a theoretical perspective, resolving this question will
ALTRUIST, BAILAR, SCRITA, WARN 2024: Workshop on sociAL roboTs for peRsonalized, continUous and adaptIve aSsisTance,
Workshop on Behavior Adaptation and Learning for Assistive Robotics, Workshop on Trust, Acceptance and Social Cues in
Human-Robot Interaction, and Workshop on Weighing the benefits of Autonomous Robot persoNalisation. August 26, 2024,
Pasadena, USA
*
Corresponding author.
$ cte@umich.edu (C. Esterwood*)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
deepen our understanding of potential philosophical differences. If trustworthiness is similar across
agents, it would be insightful to explore why. If trustworthiness differs, understanding those differences
is equally important [9].
2. Background
Trustworthiness is the degree to which humans believe a robot is worthy of their trust [10, 3, 1, 11].
Trustworthiness is subdivided into 2 or 3 sub-dimensions. However, the number and nature of these sub-
dimensions often vary based on whether one sees a robot as more like a human, more like automation,
or unique. As such, measures of trustworthiness in HRI have diverged with the formation of various
agent-specific trustworthiness measures. In HRI, the most popular of these trustworthiness measures
are Mayer’s trust measure [3], Jian et al. trust checklist [8], and the multidimensional measure of trust
in HRI (MDMT) [4].
Mayer’s trust measure was developed as part of a larger trust model in the psychology literature.
This measure considers trustworthiness a composite of ability, benevolence, and integrity. Ability is a
trustee’s perceived competence and skillfulness [3]. Integrity is the extent to which the trustee is viewed
as honest and adherent to principles [12, Pg.2]. Lastly, benevolence captures the trustee’s intention to
act in the best interest of the trustor [3, Pg.718].
Jian et al.’s trust checklist [8] adopts a different approach by targeting human–automation trust.
Specifically, [8] divided trust into human–human trust, human–machine trust, and “trust in general” [8,
Pg.12]. Through multiple cluster analyses [8] established a measure comprised of 12 unique items that
specifically applied to human–machine trust. This checklist, however did not organize these items into
higher order constructs implying a uni-dimensional structure. Subsequent work, however, has argued
that the checklist for trust “has questions that closely map to ability, integrity, and benevolence” [13,
Pg.4] implying the presence of similar sub-dimensions to that of Mayer’s trust measure [3].
The MDMT examines trustworthiness by dividing it into two 3rd order constructs. At the highest
level these are moral trust and performance trust [4]. These dimensions of trustworthiness can be
further subdivided where performance trust is a composite of reliability and capability while moral trust
is a composite of sincerity and ethics [4]. At this level the MDMT also begins to mirror Mayer’s trust
measure, however, [4] specifically argue that these dimensions are not fully compatible with Mayer’s
trust model.
The variety of trustworthiness measures in Human-Robot Interaction (HRI) complicates the field’s
ability to discuss findings and compare results across studies. Each measure is often based on different
theoretical perspectives, depending on the type of agent it applies to. Despite this, many scholars
treat these diverse measures as interchangeable [14]. On the surface, this might seem misguided, like
comparing apples to oranges. However, it may be justified given these measures’ lack of substantial
agent-specific differences. Below, we outline the arguments for and against each perspective.
2.1. The Argument for Agent Specific Trustworthiness
Extant research posits that trust development in technology differs from trust development in humans
[7]. Specifically, trust in humans hinges on factors such as the trustee’s ability, integrity, and benevolence
[3]. In contrast, trust in technology typically depends on the automated system’s performance, purpose,
and process [15]. However, as technologies like robots increasingly incorporate human-like attributes,
there is a growing effort to develop standardized measures for trust in robots that integrate both human
and technical factors [16]. This is evident in measures proposed by [8] and [4], which include integrity,
ethics, and benevolence to address the relational aspect of trust in robots. Yet, the research also suggests
that some of these factors, for example, "genuine" [4], are deemed to be relevant to robots [16].
The various trust measures and the efforts to integrate social and technical dimensions highlight
that perceptions of a technology’s human-like attributes play a critical role in shaping trust in human-
technology interactions. If perceptions significantly impact trust, choosing between human or technical
trust factors in measuring trust can be crucial empirically. For instance, using human-like trust measures
for non-human-like technology might confuse respondents, making it difficult for them to provide
accurate answers. Conversely, applying system-like trust measures to human-like technology could
lead to respondents struggling to relate to the measures. In both scenarios, mismatched trust measures
might result in lower path coefficients between trust variables and outcomes than if the appropriate
measures were used. Hence, there is a need for nuanced trust measures that account for the perceived
humanness of the technological agent to predict trust behaviors more accurately.
The variations in human-like design in technology and the subjective nature of perceived humanness
underscore the importance of having agent-specific trust measures. Historically, people have found it
easy to categorize technology as an object versus a human, hence using human-like trust measures
for technology including robots may not effectively predict trust behavior [17]. However, technology
can now display more or less human-like attributes, leading to varying perceptions of human-likeness
[18, 17]. For most human-like robots, human-like trust measures may predict behavior. On the
other hand, for less human-like robots, a combination of human and technical measures might be
more appropriate. This variability highlights that a one-size-fits-all approach to trust measurement
is inadequate. Tailoring trust measures to fit the technological agent’s specific design and perceived
humanness is crucial for accurately predicting trust behaviors.
2.2. The Argument Against Agent Specific Trustworthiness
Generally, the justification for agent-specific trustworthiness measures argues that different agents
possess unique characteristics, thus justifying the need for agent-specific measurements reflective of
these differences. It is, therefore, reasonable to expect that each of these measures of trustworthiness
would possess significantly different items. A closer examination of these items, however, shows that
these measures appear more similar than different. For example, of the three most popular measures of
trustworthiness used in HRI, each measure generally contains individual items that can be mapped
onto one of the three sub-dimensions of trust proposed by [3] as visible in table A.
Given the relative similarities across the trust measures [8, 4, 3], it is, therefore, unclear if these
existing measures are as distinct as one might expect. As a result, scholars do not yet know if they
should strictly adhere to the use of agent-specific trustworthiness measures or if these measures are
truly measuring agent-specific forms of trustworthiness.
If there were support for an agent-independent trustworthiness measure, it would allow for more
meaningful comparisons across different studies [14]. In particular, using a standardized measurement
instrument would ensure that scholars examine the same thing –i.e., trustworthiness– rather than
permutations or variations. This could accelerate research on trustworthiness in HRI and allow for
more robust consolidations and replications of existing work.
3. A Call to Action
There has yet to be a robust empirical examination comparing the efficacy of trustworthiness measures
across different agents. This prevents us from knowing if the field of HRI should strictly adhere to
these discrete measures or if a more generalizable measure can be used across different agent types.
This paper’s position is that this is a critical gap in the literature. By examining if agent-specific
trustworthiness measures are warranted, we can re-examine the current practice of creating and using
agent-specific trustworthiness measures.
Re-examining agent-specific trustworthiness measures could enable the use of more generalizable
metrics. This would, in turn, facilitate more robust comparisons of empirical results across studies.
Additionally, comparing these measures can show the similarities and differences between robots and
humans. This approach has the potential to offer new insights into human-machine relationships, not
only in terms of trustworthiness but also in aspects like robot acceptance and usage.
References
[1] C. Esterwood, L. P. Robert Jr, Three strikes and you are out!: The impacts of multiple human–robot
trust violations and repairs on robot trustworthiness, Computers in Human behavior 142 (2023)
107658.
[2] S. You, L. Robert, Trusting and working with robots: A relational demography theory of preference
for robotic over human co-workers, You, S. and Robert, LP (2023). Trusting and Working with
Robots: A Relational Demography Theory of Preference for Robotic over Human Co-Workers,
MIS Quarterly,(conditionally accepted) (2023).
[3] R. C. Mayer, J. H. Davis, F. D. Schoorman, An integrative model of organizational trust, Academy
of management review 20 (1995) 709–734.
[4] B. F. Malle, D. Ullman, A multidimensional conception and measure of human-robot trust, in:
Trust in human-robot interaction, Elsevier, 2021, pp. 3–25.
[5] J. D. Lee, K. A. See, Trust in automation: Designing for appropriate reliance, Human factors 46
(2004) 50–80.
[6] L. Robert, S. You, Are you satisfied yet? shared leadership, trust and individual satisfaction in
virtual teams, in: Proceedings of the iConference, 2013.
[7] K. A. Hoff, M. Bashir, Trust in automation: Integrating empirical evidence on factors that influence
trust, Human factors 57 (2015) 407–434.
[8] J.-Y. Jian, A. M. Bisantz, C. G. Drury, Foundations for an empirically determined scale of trust in
automated systems, International journal of cognitive ergonomics 4 (2000) 53–71.
[9] J. Kraus, I. Valori, M. Fairhurst, The social bridge: An interdisciplinary view on trust in technology,
Computers in Human Behavior 149 (2023).
[10] S. Ye, G. Neville, M. Schrum, M. Gombolay, S. Chernova, A. Howard, Human trust after robot
mistakes: Study of the effects of different forms of robot communication, in: 2019 28th IEEE
International Conference on Robot and Human Interactive Communication (RO-MAN), IEEE, 2019,
pp. 1–7.
[11] L. P. Robert, A. R. Denis, Y.-T. C. Hung, Individual swift trust and knowledge-based trust in
face-to-face and virtual team members, Journal of management information systems 26 (2009)
241–279.
[12] W. Kim, N. Kim, J. B. Lyons, C. S. Nam, Factors affecting trust in high-vulnerability human-robot
interaction contexts: A structural equation modelling approach, Applied ergonomics 85 (2020)
103056.
[13] S. C. Kohn, E. J. De Visser, E. Wiese, Y.-C. Lee, T. H. Shaw, Measurement of trust in automation: A
narrative review and reference guide, Frontiers in psychology 12 (2021) 604977.
[14] T. Law, M. Scheutz, Trust: Recent concepts and evaluations in human-robot interaction, Trust in
human-robot interaction (2021) 27–57.
[15] J. Lee, N. Moray, Trust, control strategies and allocation of function in human-machine systems,
Ergonomics 35 (1992) 1243–1270.
[16] M. Chita-Tegmark, T. Law, N. Rabb, M. Scheutz, Can you trust your trust measure?, in: Proceedings
of the 2021 ACM/IEEE international conference on human-robot interaction, 2021, pp. 92–100.
[17] N. K. Lankton, D. H. McKnight, J. Tripp, Technology, humanness, and trust: Rethinking trust in
technology, Journal of the Association for Information Systems 16 (2015) 1.
[18] C. Nass, Y. Moon, Machines and mindlessness: Social responses to computers, Journal of social
issues 56 (2000) 81–103.
A. Supplemental Table
Common Sub-Dimension Mayer’s Trust Measure [H-H] Items [3] Checklist for Trust [H-A] Items [8] Multidimensional Model of Trust [H-R] Items [4]
The agent is very capable of performing its job. I have confidence in the agent. The agent is capable.
The agent is known to be successful at the things it tries to do. The agent is dependable. The agent is skilled.
The agent has much knowledge about the work that needs to be done. The agent is reliable. The agent is competent.
I feel very confident about the agent’s skills. I can trust the agent. The agent is meticulous.
Ability
The agent has specialized capabilities that can increase our performance. I am familiar with the agent. The agent is reliable.
The agent is well-qualified to perform the job. The agent is predictable.
The agent is dependable.
The agent is consistent.
The agent is very concerned about my welfare. The agent behaves in an underhanded manner.* The agent is benevolent.
My needs and desires are very important to the agent. I am suspicious of the agent’s intent, action, or output.* The agent is kind.
Benevolence The agent would not knowingly do anything to hurt me. The agent’s actions will have a harmful or injurious outcome.* The agent is considerate.
The agent really looks out for what is important to me. The agent has goodwill.
The agent will go out of its way to help me.
The agent has a strong sense of justice. The agent is deceptive.* The agent is sincere.
I never have to wonder whether the agent will stick to its word. I am wary of the agent.* The agent is genuine.
The agent tries hard to be fair in dealings with others. The agent provides security. The agent is candid.
The agent’s actions and behaviors are not very consistent. The agent has integrity. The agent is transparent.
Integrity
I like the agent’s values. The agent is ethical.
Sound principles seem to guide the agent’s behavior. The agent is moral.
The agent is principled.
The agent has integrity.
Table 1
Comparison of trust measures across common sub-dimensions based on [3]. * indicates reverse coded item.