Source-linked AI summary
A Systematic Literature Review of User Trust in AI-Enabled Systems: An HCI Perspective
Tita Alissa Bach, Amna Khan, Harry Hallock, Gabriela Beltrão, Sonia Sousa
TL;DR
Definitions and measurement of user trust in AI-enabled systems remain insufficiently agreed on and studied. This review synthesizes empirical research, finding that most included articles examined trust factors and surveys were the most common measurement method.
Problem
The definition, importance, and measurement of user trust in AI-enabled systems are not yet well agreed on or studied.
Source: §1.1 · p. 3Method
A systematic literature review synthesized empirical evidence on user trust in AI-enabled systems across 23 included studies.
Source: §2.1 · p. 3, §3 · p. 4Results
Almost 78.36% of included articles focused on trust antecedents, predictors, dimensions, or factors; surveys were the most common measurement method (59.56%).
Source: §3 · p. 4Takeaways & Limitations
The review positions fostering and maintaining user trust as important to calibrating the user–AI relationship.
Source: §1.1 · p. 3The included studies varied in context, so generalizing the findings may not be feasible and should be done with caution.
Source: §4.4 · p. 13
Abstract
from arXiv · showhide
User trust in Artificial Intelligence (AI) enabled systems has been increasingly recognized and proven as a key element to fostering adoption. It has been suggested that AI-enabled systems must go beyond technical-centric approaches and towards embracing a more human centric approach, a core principle of the human-computer interaction (HCI) field. This review aims to provide an overview of the user trust definitions, influencing factors, and measurement methods from 23 empirical studies to gather insight for future technical and design strategies, research, and initiatives to calibrate the user AI relationship. The findings confirm that there is more than one way to define trust. Selecting the most appropriate trust definition to depict user trust in a specific context should be the focus instead of comparing definitions. User trust in AI-enabled systems is found to be influenced by three main themes, namely socio-ethical considerations, technical and design features, and user characteristics. User characteristics dominate the findings, reinforcing the importance of user involvement from development through to monitoring of AI enabled systems. In conclusion, user trust needs to be addressed directly in every context where AI-enabled systems are being used or discussed. In addition, calibrating the user-AI relationship requires finding the optimal balance that works for not only the user but also the system.
1. Introduction
AI-enabled systems offer real-world applications but raise concerns about opacity, bias, and harm. This review examines how HCI research defines, explains, and measures user trust in the user-AI relationship.
1. Introduction
- Opaque, uninterpretable, biased, and uncertain AI can impede prediction and error tracing; poor adaptation may mislead decisions, with failed systems causing discrimination or death.Source: §1 · p. 2
- Potential harms undermine user trust and raise accountability concerns; AI ethics guidelines diverge in interpreting and implementing shared principles, while trustworthy AI adds legality and robustness (European Commission, 2019).Source: §1 · pp. 2–3
- The review frames trustworthiness as risk management through human-centric HCI, an interdisciplinary approach that develops technology around user needs and informs strategies to foster user trust.Source: §1 · p. 3
1.1. Research scope
- Trust definitions, importance, and measurement remain unsettled, while complex AI can create a trust gap; the review centers the user-AI relationship as a sociotechnical collaboration.Source: §1.1 · p. 3
1.2. Research questions
- The review asks how user trust in AI-enabled systems is defined, what factors influence it, and how it can be measured.Source: §1.2 · p. 3
2. Methodology
The authors conducted a PRISMA-aligned systematic literature review to identify empirical studies of factors influencing user trust in AI-enabled systems, then extracted and synthesized findings from 23 included articles.
2.1. Literature search and strategy
- The review followed PRISMA standards and searched EBSCO Discovery Service and Web of Science using terms refined through research questions, a pilot search, and the ACM classification system.The systematic review was selected to provide a rigorous overview of empirical evidence while aiming to minimize bias and support replication.Source: §2.1 · pp. 3–4
2.2. The study selection
- The screening process reduced 493 search results to 23 articles for analysis, using group consensus and an independent eligibility check by the first authors.Eligible studies were English-language scientific articles or conference proceedings published from January 1, 2011 to May 15, 2021; they had to be empirical, address factors influencing trust in AI-enabled systems, provide a clear methodology, and have an available full text. The team discussed 67 articles at full-text screening before accepting the final 23.Source: §2.2 · p. 4
2.3. Analysis and synthesis
- The authors independently extracted study and trust-related information, clustered influencing factors through content analysis, and resolved disagreements by group consensus.The extracted information included study context and domain, the AI-enabled system, trust definition and measurement, methodology, participant or dataset size, and trust-influencing factors; first authors then checked extraction consistency.Source: §2.3 · p. 4
3. Results
Across 23 studies, the reviewed evidence commonly examined trust factors and used surveys, with many studies published after 2018 or focused on Robotics and E-commerce.
3. Results
- 52.17% of the 23 included articles were published after 2018, 56.52% of studies were conducted in the USA and Germany, and 52.17% focused on Robotics and E-commerce.Table 3 is cited for the publication-year distribution.Source: §3 · p. 4
- General AI/ML and automated algorithms were the most common AI-enabled system types, at 30.43% of included articles.Source: §3 · p. 4
- Almost 78.36% of articles assessed, predicted, or augmented trust antecedents, predictors, critical dimensions, or factors, while surveys measured trust in 59.56%.Surveys were used alone or with interviews or focus groups.Source: §3 · p. 4
3.1. RQ1: How is user trust in AI-enabled systems defined?
Among the 23 empirical studies, seven defined trust, while eight conceptualized it without defining it and eight did neither.
3.1. RQ1: How is user trust in AI-enabled systems defined?
- Seven articles defined trust, eight conceptualized it without a definition, and eight neither defined nor conceptualized it.The eight articles that conceptualized trust without defining it were Corritore et al. (2012), Duffy (2017), Elkins & Derrick (2013), H€oddinghaus et al. (2021), Law et al. (2021), Lee et al. (2021), Sharma (2015), and Smith (2016).Source: §3.1 · p. 4
- Four articles used Mayer’s trust definition.These were Foehr & Germelmann (2020), Glikson & Woolley (2020), Lin et al. (2019), and Thielsch et al. (2018).Source: §3.1 · p. 4
- Two articles used Lee and See’s trust definition (2004), while Yan et al. (2013) developed its own definition with trustworthy characteristics from Avizienis et al. (2004).Source: §3.1 · p. 5
3.2. RQ2: What factors influence user trust in AI-enabled systems?
Across 23 studies, trust was influenced by socio-ethical considerations, technical and design features, and user characteristics, with user characteristics identified most often. Factors ranged from system features and user attitudes to prior experience, interactions, and readiness.
3.2. RQ2: What factors influence user trust in AI-enabled systems?
- User characteristics were identified in 22 articles, compared with 12 for technical and design features and 8 for socio-ethical considerations.The user-characteristic findings were grouped as inherent characteristics (3 articles), acquired characteristics (4), attitudes (10), and external variables (6).Source: §3.2 · p. 5, §3.2.3 · p. 7
3.2.1. Socio-ethical considerations influencing user trust
- Trust depends on user readiness and social-ethical safeguards, including data protection, ongoing feedback, and mechanisms to foster, maintain, or recover trust; accountability for harm can be difficult to assign.Mismatched readiness between users and system development may undermine trust, while open communication and feedback were suggested as determinants. Accountability may be unclear when harm occurs.Source: §3.2.1 · p. 5
3.2.2. Technical and design features influencing user trust
- For virtual agents, trust-related features included human-like benevolence, social presence, immediacy, integrity, and communication choices; for algorithms, explanations, prediction correctness, and integrity mattered.Other identified features included lower vocal pitch or text rather than synthetic voice, and explanations of algorithm actions, reliability, performance, risk factors, and contextual information. Complex information-system reliability, information credibility, and website features also influenced trust; dependence could leave users with no choice but to trust.Source: §3.2.2 · pp. 5–6
3.2.3.1. User inherent characteristics (i.e., personality traits, gender, and self-trust).
- Personality traits, gender, and self-trust were among inherent characteristics linked to trust in AI-enabled systems.Low Openness was associated with the highest trust among the Big Five traits; women were more likely to report higher trust, and users could rely on their own skills to assess a system.Source: §3.2.3.1 · p. 7
3.2.3.2. User acquired characteristics (i.e., user experiences and educational levels).
- Prior experiences and education shaped trust, while familiarity over time could increase it even when initial trust was low.Positive experience could transfer trust to other systems from the same provider, whereas need or dependency could override negative experience. Users without college education were less likely to trust, but trust increased with familiarity.Source: §3.2.3.2 · p. 7
3.2.3.3. User attitudes (i.e., user acceptance and readiness, needs and expectations, judgment and perceptions).
- Acceptance and readiness, needs and expectations, and perceptions of credibility, risk, ease of use, benevolence, integrity, transparency, and relatability all influenced trust.Addressing uncertainty and the artificial divide—including low motivation or technical competence—was suggested to foster acceptance and readiness; improving understandability, control, and information accuracy could reduce uncertainty (Klumpp & Zijm, 2019; Hoffmann & S€ollner, 2014). Mismatches between expectations and experience could threaten trust, especially under dependency.Source: §3.2.3.3 · pp. 7, 9
3.2.3.4. User external variables (i.e., initial interactions, user interactions, cognitive load levels, time and usage).
- Introductions through close contacts, high-quality user interactions, cognitive load, time spent interacting, and system usage were identified as external influences on trust.Trust increased across successive question blocks in one embodied-agent study regardless of initial trust, and greater usage was suggested as a predictor of greater trust.Source: §3.2.3.4 · p. 10
3.3. RQ3: How is user trust in AI-enabled systems measured?
User trust in AI-enabled systems was measured most often with surveys, while qualitative methods were the next most common approach. Studies also varied in participant counts and in the demographic information they reported.
3.3. RQ3: How is user trust in AI-enabled systems measured?
- 16 studies (69.56%) used surveys, either alone or combined with interviews or focus groups, to measure user trust.Of these, 12 studies (75%) developed their own questionnaires, two (12.5%) used both their own and previously developed questionnaires, and two (12.5%) used previously developed questionnaires.Source: §3.3 · p. 10
- Six studies (26.09%) used interviews and/or focus groups as standalone or combined methods to measure trust.Source: §3.3 · p. 10
- 20 articles included participants, with study sizes ranging from 21 to 3423 participants (M = 326.80).Fourteen articles reported gender: 23.08–61.86% female (M = 45.62%); two reported gender other than male or female, at 4% (Zhang, Genc, et al., 2021), 0.51% and 0.71%.Source: §3.3 · p. 10
4. Discussion
Across 23 empirical studies, the review finds that trust definitions and measures vary, while user characteristics, interactions, system features, and socio-ethical conditions shape trust. It recommends context-sensitive trust definitions and user involvement throughout AI system development and use.
4.1. RQ1: How is user trust in AI-enabled systems defined
- Only seven of 23 studies explicitly define trust, while eight conceptualize it and nine do neither; the review recommends choosing definitions suited to context.Only one study developed its own definition, while six used Mayer’s or Lee and See’s definitions. The review suggests risk and personal impact can guide the choice between definitions, rather than comparing which is better.Source: §4.1 · p. 10
4.2. RQ2: What factors influence user trust in AI-enabled systems
- User characteristics dominate trust influences, while interactions, context-tailored system features, and socio-ethical conditions shape trust and support its development.The review highlights user involvement from development through monitoring, and suggests repeated interactions can build trust as users adjust expectations and become familiar with a system. It also recommends tailoring design and interventions to user groups and contexts, and notes the need for explanations and multidisciplinary work to address gaps between AI regulation and practice.Source: §4.2 · pp. 10, 13
4.3. RQ3: How user trust in AI-enabled systems is measured
- 69.56% of included studies developed and used their own trust questionnaires, making surveys the most common measurement method; qualitative methods were second most used.Qualitative methods can explore trust’s complexity but are harder to compare and open to varied interpretations. The review points to validated cross-context tools such as Gulati et al.’s 12-item Human-Computer Trust Scale and calls for more studies measuring trust.Source: §4.3 · p. 13
4.4. Limitations
- The review cautions that its findings may not generalize because included studies vary in context, and that its search terms and exclusion of grey literature may have omitted relevant work.Other trust definitions may also exist beyond those identified in the 23 included studies.Source: §4.4 · p. 13
5. Conclusions and future aims
The conclusion calls for context-sensitive study of how trust and its drivers change over time, broader and more objective measurement, and trust calibration that accounts for both users and systems.
Conclusions and future aims
- Future research should test trust concepts in specific contexts and examine how influencing factors vary by user, context, and time as user-AI relationships evolve.The review identifies time as a particular gap for understanding how trust factors operate at different points in an interaction.Source: §5 · pp. 13–14
- Future studies should complement perception-based surveys, interviews, and focus groups with more objective measures such as psychophysiological signals.Source: §5 · p. 14
- Research should extend beyond the USA and Germany to diverse geographical locations to examine how cultural factors influence trust (Jobin et al., 2019).Approximately half of the reviewed studies were conducted in the USA and Germany.Source: §5 · p. 14
- Trust-fostering work should involve multidisciplinary teams from early ideation, including developers, designers, target users, and ethics, legal, behavioral, social-science, and domain experts.Development should account for inherent, acquired, attitudinal, and external user characteristics: systems may need adjustment for less changeable inherent traits, while attitudes and external variables may respond to user-experience improvements (Zhang, Genc, et al., 2021).Source: §5 · p. 14
- Quality interactions require environments with mechanisms to foster, maintain, and recover trust, while calibration seeks an optimal balance for both users and systems (Shneiderman, 2020b).Source: §5 · p. 14
About the authors
The author biographies describe research and professional work spanning human factors, trustworthy and transparent systems, healthcare AI, cross-cultural trust in technology, and interaction design.
About the authors
- Tita Alissa Bach holds a doctorate in behavioural and social sciences and applies Human Factors principles in organizations using technology.Her doctorate is from Groningen University, with a focus on Human Factors and Organisational Psychology.Source: About the authors · p. 17
- Amna Khan is a Junior Researcher and PhD student developing a framework for adaptive knowledge reuse systems designed to be trustworthy and transparent.She studies Information Society Technologies at Tallinn University and earned a master’s degree in Mechatronics Engineering in 2017.Source: About the authors · p. 17
- Harry Hallock’s current work addresses trustworthy AI in healthcare, federated health-data networks, and health-data governance, management, and quality assurance.He holds a doctorate in Cognitive Neurosciences from the University of Sydney and has experience in product development and large-scale project management.Source: About the authors · p. 17
- Gabriela Beltr~ao is a candidate in Information Society Technologies whose research examines cross-cultural differences in trust in technology and their design implications.Source: About the authors · p. 17
- Sonia Sousa is an Associate Professor of Interaction Design at Tallinn University and heads its Joint Online MSc in Interaction Design.Her funded projects include NGI-Trust, CHIST-ERA, Horizon 2020, and AFOSR; she was nominated for AcademiaNet.Source: About the authors · p. 17