# Understanding AI Trustworthiness: A Scoping Review of AIES & FAccT Articles
**Authors**: Siddharth Mehrotra, Jin Huang, Xuelong Fu, Roel Dobbe, Clara I. SĂĄnchez, Maarten de Rijke
> University of Amsterdam & Delft University of Technology The Netherlands TBS
> University of Cambridge United Kingdom XXX
> Delft University of Technology The Netherlands XX
> University of Amsterdam The Netherlands XXX
Min-Yen Kan Survey Articles
(2025)
## Abstract
Background: Trustworthy AI serves as a foundational pillar for two major AI ethics conferences: AIES and FAccT. However, current research often adopts techno-centric approaches, focusing primarily on technical attributes such as reliability, robustness, and fairness, while overlooking the sociotechnical dimensions critical to understanding AI trustworthiness in real-world contexts.
Objectives: This scoping review aims to examine how the AIES and FAccT communities conceptualize, measure, and validate AI trustworthiness, identifying major gaps and opportunities for advancing a holistic understanding of trustworthy AI systems.
Methods: We conduct a scoping review of AIES and FAccT conference proceedings to date, systematically analyzing how trustworthiness is defined, operationalized, and applied across different research domains. Our analysis focuses on conceptualization approaches, measurement methods, verification and validation techniques, application areas, and underlying values.
Results: While significant progress has been made in defining technical attributes such as transparency, accountability, and robustness, our findings reveal critical gaps. Current research often predominantly emphasizes technical precision at the expense of social and ethical considerations. The sociotechnical nature of AI systems remains less explored and trustworthiness emerges as a contested concept shaped by those with the power to define it.
Conclusions: An interdisciplinary approach combining technical rigor with social, cultural, and institutional considerations is essential for advancing trustworthy AI. We propose actionable measures for the AI ethics community to adopt holistic frameworks that genuinely address the complex interplay between AI systems and society, ultimately promoting responsible technological development that benefits all stakeholders.
copyright: cc doi: 10.1613/jair.1.xxxxx publicationmonth: 10 journalyear: 2025
## 1. Introduction
The rapid advancement and widespread adoption of artificial intelligence (AI) have ushered in a new era of technological innovation, bringing both immense potential and significant challenges. As AI increasingly permeates aspects of our lives, from healthcare to criminal justice, the need for trustworthy AI has become paramount. Trustworthy AI, as a concept, encompasses a multifaceted approach to AI systems that prioritizes safety, transparency, and ethical considerations for all stakeholders (10.1145/3555803, 85). It extends beyond technical proficiency, embracing principles like reliability, fairness, explainability, and accountability. This has given rise to dedicated academic venues like the AAAI/ACM Conference on AI, Ethics, and Society (AIES) and ACM Conference on Fairness, Accountability, and Transparency (FAccT), fostering interdisciplinary discourse on AI ethics. It is now clear that a purely technology-centric view of trustworthiness is not enough. Trustworthy AI requires an interdisciplinary perspective that views AI systems as sociotechnical systems.
The sociotechnical nature of AI systems demands a holistic approach to trustworthiness, one that considers not only the technical aspects but also the complex interplay between AI and the broader social, cultural, and institutional contexts in which it operates. (betram, 95) lays out five elements of trustworthiness as competent, reliable, transparent, benevolent, and having ethical integrity â and calls out to study these elements in a broader sociotechnical setting. This expanded perspective is particularly crucial for the AI ethics community which aims to bridge the gap between computer science and other disciplines in addressing AIâs ethical challenges.
In a study by (laufer, 82), when asked about what values 36 self-selecting FAccT affiliates believe FAccT scholarship should address in the near future, participants reflected on the need for broader conceptions for trustworthiness. Central to this discourse is the recognition that the concept of trustworthiness is fundamental to understanding and predicting trust levels in AI systems. It becomes imperative to critically examine how the AIES & FAccT communities conceptualize and communicate trustworthiness, and to what ends these efforts are directed. This examination raises important questions about the commitments that trustworthy AI research in these venues signifies, or should signify, in the broader context of AI ethics and societal impact.
The study of the trustworthiness of AI systems has been a topic of interest for many years, even before the existence of the AIES & FAccT conferences. Scholars from various disciplines have identified values that determine the attribution of trustworthiness, revealing both similarities and differences across fields. For example, in interpersonal trust, competence, predictability, benevolence, and integrity have been highlighted as crucial values (lahusen2024trust, 79). For public institutions, the list extends to competence and reliability, procedures like transparency and accountability, and results including effectiveness and general welfare (polemi2024challenges, 115). In the context of AI systems, properties such as reliability, robustness, safety, interpretability, explainability, fairness, transparency, and accountability have been identified as trust-relevant (kaur2021requirements, 70, 84).
Trustworthiness has been used to refer to two sides of a coin (e.g., lee2004trust, 84, 86, 98, 146, 122). On the one hand, trustworthiness has been referred to as an objective attribute of the trustee (e.g., zerilli2022transparency, 152, 52, 71, 65, 47). On the other hand, trustworthiness has been referred to as a trustorâs subjective perception of a trusteeâs attributes (e.g., mayer1995integrative, 98, 122, 84). Overall, trustworthiness is multifaceted, comprising several elements regardless of whether it is viewed as an inherent quality of the trusted party or as a subjective assessment made by those extending trust (baer2018people, 9, 31, 65, 84). Therefore, by concentrating on trustworthiness, we can assess the qualities and behaviors of AI systems that contribute to their reliability, safety, and ethical alignment.
Despite the extensive research on trustworthy AI (toney, 141), there remains a critical gap in understanding the sociotechnical nature of these systems. Few studies have adequately addressed the complex interplay between technical capabilities and social contexts in which AI operates. This paper aims to address these gaps by conducting a comprehensive scoping review of articles published in AIES & FAccT conferences to date. Through our analysis, we seek to answer several key questions:
1. How is trustworthiness conceptualized in the context of AI systems within AIES & FAccT proceedings?
1. What methodologies are employed to measure, verify, and validate trustworthiness?
1. Which application areas are most prominently represented?
1. What underlying values and ethical considerations drive this body of work?
We explicitly focus on values because an account of value embodiment in AI aids in assessing whether designed AI systems indeed embody a range of moral values (mehrotraaies, 101), e.g., those articulated by the EU High-Level Expert Group (EU_AI_Ethics_Guidelines, 42). To this end, we examine three key dimensions: intended, embodied and realized values following the framework of (van2020embedding, 114) to do a thematic analysis on our corpus. Our motivation to use this framework is that it helps us overcome the limitations of focusing solely on intentions (which may not manifest in the system) or outcomes (which may be influenced by external factors) by emphasizing embodied valuesâthose intentionally and successfully embedded in the system.
Overall, our core contributions are:
1. A systematic analysis of how trustworthiness is conceptualized, measured, and validated within the AIES & FAccT community.
1. Identification of major gaps in current research, particularly regarding the sociotechnical aspects of AI systems.
1. A critical examination of the values and ethical considerations underpinning trustworthy AI research.
1. A discussion of the intellectual and broader impact of AIES & FAccT conferences in studying AI trustworthiness.
## 2. Related Work
Trust is a much-discussed topic in algorithmic decision-making, especially in the area of AI (mehrotrareview, 99). In the development of trust, the process by which a human assesses the trustworthiness of a system, leading to their perception of trustworthiness, is crucial. Only with an accurate trustworthiness assessment can people base their trust on adequate expectations about a systemâs capabilities and limitations and make informed decisions. Trustworthiness of AI systems has been studied from multiple disciplines such as computer science (liu2023trustworthy, 89), psychology (schlicker2022calibrated, 124), public administration (lahusen2024trust, 79), and medicine (zhang2023ethics, 154). Below, we provide a background on studying trustworthiness in four subject areas, namely: computer science, law, social sciences, and humanities. These areas are common in the AIES & FAccT conferences. AIES includes experts from various disciplines such as ethics, philosophy, economics, sociology, psychology, law, history, and politics (AIES2018, 50), while FAccT brings together scholars from computer science, law, social sciences, and humanities (Facct2018, 48). These areas serve as the foundation for our interdisciplinary approach to understanding the multifaceted nature of AI trustworthiness.
### 2.1. Computer Science
Computer science has predominantly focused on technical perspectives for trustworthy AI applications, emphasizing three characteristics, (i) robustness, fortifying AI models against malicious attacks such as adversarial attacks; (ii) generalization, ensuring maintaining performance on unseen out-of-distribution (OOD) data; and (iii) interpretability, improving understanding of AI model predictions (wang2023trustworthy, 148, 104). (mucsanyi2023trustworthy, 104) list common pitfalls in evaluating trustworthy machine learning models, e.g., inconsistent coding for evaluation metrics despite being the same mathematically, confounding multiple factors in method comparisons, training and test samples overlap, and lack of validation set. Building on the foundational principles of trustworthy machine learning, recent research has begun to investigate the characteristics that apply to large language models and mechanisms for evaluating their trustworthiness. (liu2023trustworthy, 89) survey key dimensions for assessing LLM trustworthiness. These include reliability, safety, fairness, resistance to misuse, explainability and reasoning, adherence to social norms, and robustness. (liu2023trustworthy, 89) highlight that a key principle for evaluating an LLM âs trustworthiness is the generation of proper test data across the aforementioned dimensions. Finally, (toney, 141) review similarities and differences between governmentsâ and researchersâ definitions and frameworks on trustworthy AI. The authors find inconsistencies between policy and research term frequencies, highlighting the different focuses of each group on trustworthy AI, distinct from our work focusing on AI trustworthiness rather than the use of the terminology within the AIES & FAccT community.
Studies in computer science are often tech-centred, focusing primarily on technical methodologies but overlooking socio-cultural, ethical, and legal dimensions of trustworthiness. But achieving trustworthiness demands interdisciplinary collaboration. Thus, we consider AI systems as socio-technical and explore how the AIES & FAccT community is deepening our insight into trustworthiness of AI.
### 2.2. Social Sciences
Social science provides a broad perspective on AI trustworthiness focusing on its societal, institutional, economic, and political implications and dynamics. E.g., (dacon2023you, 28) investigates the impact of AI trustworthiness on various aspects of society, like the environment, human society, and societal values. These implications are abstracted into three social principles: (i) harm prevention, which focuses on ensuring safety, security, reliability, and privacy; (ii) explicability, which emphasizes explainability and transparency of the system; and (iii) fairness, which consists of accountability and the well-being of society and environment. (thiebes2021trustworthy, 138) evaluate AI trustworthiness frameworks and synthesize the recurring themes in five principles: beneficence, non-maleficence, autonomy, justice, and explicability. They also conceptualize the DaRe4TAI framework, which structures and directs data-driven research in trustworthy AI while addressing tensions between the five principles. Similar framework-oriented approaches are pursued by other social science scholars, as they form a crucial connection with policymaking (polemi2024challenges, 115, 78). We adopt a similar lens in this review as (polemi2024challenges, 115, 78) to discuss the results of our corpus analysis with a socio-technical focus by thinking about the entire ecosystem in which our AI systems operate.
### 2.3. Law
The question of AI trustworthiness is crucial in the legal domain, where even minor inaccuracies can have significant consequences for individuals navigating the complexities of legal processes. Studies have demonstrated the capabilities of LLMs in understanding and responding to natural language queries. However, the tendency of LLMs to generate inaccurate and incomplete information raises serious concerns about their trustworthiness in providing legal guidance (kattnig2024assessing, 69). E.g., (tan2023chatgpt, 134) investigate ChatGPTâs ability to provide legal information using several simulated cases and compared with the performance to that of and find that often leads laypeople to over-trust it. More generally looking at use of AI systems in law, (steenhuis2024ai, 130) raises important questions of exploring trustworthiness of such systems for methods of service delivery, replacement of routine legal tasks and essential legal assistance to those who might otherwise go without. Potential answers to these explorations can be derived from (hagan2020legal, 55) âs work who provides legal design testing metrics to explore trustworthiness of AI systems plus methodology and âdesign deliverablesâ based on these metrics with California state courtsâ Self Help Centers (hagan2018human, 54). Overall, AI systems hold significant promise for improving accessibility and efficiency in the legal domain but face challenges related to inaccuracies, hallucinations, and over-reliance, raising concerns about trustworthiness.
### 2.4. Humanities
The study of AI trustworthiness within the humanities emphasizes ethical, philosophical, and cultural aspects, addressing critical issues such as moral agency, responsibility, and the societal impact of AI decision-making (noller2024extended, 107, 117, 25). (balmer2023sociological, 10) critically examines AIâs role in society through creative methodologies, while others explore how AI reshapes creativity, authorship, and cultural interpretation (lim2018ai, 87). Historical and philosophical analyses (coeckelbergh2022political, 26, e.g.,) provide valuable context for understanding contemporary AI technologies, alongside inquiries into AI consciousness and religious perspectives that highlight cultural and ethical considerations (IsDeweys, 46, 2, 59). These diverse approaches underline the importance of contextualizing AI systems within broader humanistic and historical frameworks to assess their trustworthiness effectively. The integration of ethical, cultural, and philosophical dimensions in evaluating AI trustworthiness offers a more comprehensive understanding of its implications, enabling more informed and responsible development of AI systems.
## 3. Methodology and Corpus Overview
### 3.1. Methodology
We employ a Scoping Literature Review (SLR) following the guidelines by (arksey2005scoping, 8) to identify articles studying trustworthiness of AI systems. SLRs guide the gathering and identification of papers in a topic area for scrutiny (kastner2012most, 68), which enables us to perform our thematic investigation of AIES & FAccT scholarship. We use qualitative manual coding and computational corpus analysis, including topic modeling, to extract themes and patterns from the data.
Data Collection. The data consists of peer-reviewed articles published by the AIES & FAccT conferences between 2018â2025, downloaded from the ACM website on September 05, 2025 Note that the proceedings of AIES 2025 are not available when we have submitted this manuscript, hence AIES 2025 articles are excluded from this review.. In total, 241 articles (83 AIES, 158 FAccT) are obtained that include the keyword âtrustworthinessâ in the full-text excluding references; 6 non-archival extended abstracts are excluded, leaving us 235 articles for screening.
Screening and Selection Criteria. We note several challenges in reviewing the AIES & FAccT papers. The concept of trustworthiness is often embedded throughout papers instead of being the core theme. For example, auditing algorithmic systems or exploring potential biases in AI systems eventually help in understanding their trustworthiness, but this is often a point of discussion rather than the core contribution. Therefore, it is difficult to conclude that these articles engage with the core concept. Hence, we aim for a balanced approach studying trustworthiness, by including two types of article in our analysis: (i) articles that directly address trustworthiness as a central theme, like (liao2022designing, 86, 45, 65, 44), and (ii) articles where trustworthiness is not the main focus, but the authors explain how their findings contribute to a broader understanding of system trustworthiness and, ultimately, user trust.
Therefore, keeping our balanced approach, following prior literature reviews on trust (mehrotrareview, 99, 145, 12) and best research practices to study a particular topic (ajmani, 1, 116), we define our inclusion criteria as:
1. One of the contributions of the article is linked to understanding trustworthiness of an algorithmic system.
1. The article engages with trustworthiness as a component of their measure, directly influenced by the results, or as a key discussion theme.
Our exclusion criteria are:
1. The article discusses the need for AI trustworthiness without directly defining, measuring, or modeling it.
1. The article is classified as a survey, scoping review, or literature review.
The final corpus consists of 43 papers after applying our inclusion and exclusion criteria. An overview of our review process following the PRISMA protocol (page2021prisma, 110) is shown in Figure 1.
<details>
<summary>x1.png Details</summary>
### Visual Description
## Diagram: Patient Data Collection and Selection Flow
### Overview
This image presents a flowchart detailing the process of collecting and selecting patient data for a study. It illustrates the initial number of patients, subsequent filtering steps based on data availability, and the final count of patients included and excluded from the study. The diagram uses rectangular nodes to represent stages or groups of patients and directed arrows to show the flow or progression.
### Components/Axes
The diagram consists of six rectangular nodes, each containing descriptive text and a patient count (N), connected by five directed arrows indicating the flow of the process. There are no traditional axes, legends, or specific color coding beyond black text on a white background within black-bordered rectangles.
### Detailed Analysis
The flow proceeds from top to bottom, with a branching point in the middle.
1. **Top Node (Center-Top):**
* Text: "Data collected from\nPatients\n(N=211)"
* This node represents the initial pool of data collected from 211 patients.
* An arrow points downwards from this node.
2. **Second Node (Center, below the Top Node):**
* Text: "Data collected from the\nmedical records of patients\n(N=179)"
* This node indicates a subset of 179 patients from whom data was collected specifically from medical records. This suggests 32 patients (211 - 179) were not included in this medical record data collection step or were excluded prior to it.
* Two arrows originate from this node, branching downwards and outwards to the left and right.
3. **Third Node (Bottom-Left, below the Second Node):**
* Text: "Patients with all the\nnecessary data\n(N=117)"
* This node represents 117 patients who had all the required data. This group is a direct outcome of the "Data collected from medical records" step.
* No arrows originate from this node, indicating it's an endpoint or a group not further processed in this specific flow.
4. **Fourth Node (Bottom-Right, below the Second Node):**
* Text: "Patients with missing\nmedical record data\n(N=62)"
* This node represents 62 patients who had missing data in their medical records. This group, along with the "Patients with all the necessary data" (N=117), sums up to the 179 patients from the previous step (117 + 62 = 179).
* Two arrows originate from this node, branching downwards and outwards to the left and right.
5. **Fifth Node (Bottom-Most Left, below the Fourth Node):**
* Text: "Patients excluded from\nthe study due to missing\ndata\n(N=11)"
* This node represents 11 patients who were explicitly excluded from the study because of missing data. This group is a subset of the "Patients with missing medical record data" (N=62).
* No arrows originate from this node.
6. **Sixth Node (Bottom-Most Right, below the Fourth Node):**
* Text: "Patients included\nin the study\n(N=51)"
* This node represents 51 patients who were ultimately included in the study. This group is also a subset of the "Patients with missing medical record data" (N=62), implying that despite some initial missing data, 51 patients were still deemed suitable for inclusion (perhaps after further review or imputation). The sum of this group and the excluded group (11 + 51 = 62) matches the "Patients with missing medical record data" count.
* No arrows originate from this node.
### Key Observations
* The diagram clearly illustrates a patient selection funnel, starting with a larger group and narrowing down to a final study population.
* The numbers (N values) are consistent at each branching point, demonstrating a clear accounting of patients as they move through the selection process.
* The primary reason for exclusion explicitly stated is "missing data."
* The process distinguishes between patients with complete data and those with missing data, and then further processes the latter group.
### Interpretation
This flowchart describes a common methodology in research studies, particularly in medical or clinical fields, for patient recruitment and data management. It demonstrates a systematic approach to handling patient data, starting from an initial collection and moving through stages of refinement and exclusion.
The initial step of "Data collected from Patients (N=211)" suggests a broader initial contact or data source. The subsequent step, "Data collected from the medical records of patients (N=179)," indicates a specific focus on medical record data, leading to an initial reduction of 32 patients (211 - 179). This reduction could be due to patients not having medical records, or their records not being accessible, or other initial exclusion criteria not detailed here.
The core of the selection process lies in the split of the 179 patients into those with "all the necessary data (N=117)" and those with "missing medical record data (N=62)." This highlights a critical decision point based on data completeness.
Interestingly, the "Patients with missing medical record data (N=62)" group is further processed, resulting in "Patients excluded from the study due to missing data (N=11)" and "Patients included in the study (N=51)." This implies that not all patients with *some* missing data were automatically excluded. There might have been a secondary review, a threshold for acceptable missingness, or specific types of missing data that were deemed non-critical for the study's objectives, allowing 51 of these patients to be included.
The final outcome shows two main groups of patients contributing to the study: the 117 patients who initially had all necessary data, and the 51 patients who, despite some missing data, were ultimately included. The total number of patients included in the study would be 117 + 51 = 168. The diagram effectively communicates the transparency and rigor of the patient selection process, accounting for every patient at each stage.
</details>
Figure 1. Flowchart of the articles reviewing process following the PRISMA protocol (page2021prisma, 110).
Free-form Questions. Following the Goal-Question-Metric (GQM) approach (caldiera1994goal, 11), we analyze the articles along the GQM dimensions that are phrased as questions: (i) what is the goal of the paper, (ii) what are different research questions related to this goal, and (iii) what are metrics to measure some aspects/factors of this goal? This framework provides us with a way to categorize the understanding of trustworthiness in the AIES & FAccT community in the form of generic free-form questions closely tied to the research questions from Section 1.
We employ a thematic analysis to analyze the 43 articles and address our research questions. First, we develop a coding scheme to extract relevant information from each paper. For defining trustworthiness (RQ1), we identify key terms, phrases, and concepts used across papers to formulate common definitional themes. We further link the identified themes with the International Organization for Standardization (ISO) standard 5723 on trustworthiness (ISO_Trustworthiness_Vocabulary_2022, 63), allowing us to track frequency and depth of coverage. The ISO standard is chosen as a reference framework as it represents an internationally recognized benchmark for trustworthiness in technical systems (toney, 141). Furthermore, to understand the drivers of studying trustworthiness, we use both inductive and deductive coding approaches, first allowing themes to emerge naturally from the text then mapping these against common clusters.
For trustworthiness measures (RQ2), we develop a hierarchical coding structure to categorize assessment methods and validation & verification techniques. For application scenarios (RQ3), we employ open coding to identify emerging categories of use cases, followed by axial coding to establish relationships between different application contexts. Finally, to better understand the role of values in AI trustworthiness (RQ4), we analyze intended, embodied, and realized values. We followed (van2020embedding, 114) âs work to study the role of values. According to his work, intended values refer to the ethical principles and societal benefits that AI systems aim to achieve, embodied values are those integrated into their design and implementation, and realized values are the actual outcomes observed in practice. We use a two-stage coding process where we first identify explicitly stated intended values related to trustworthiness and then compare these against realized outcomes and embodied values. This involves creating pairs of intended-realized values and analyzing any gaps or alignments between them.
Our coding process is iterative, with initial codes refined over multiple analysis rounds. To ensure reliability, three researchers independently code a subset of papers (25%) and calculate an average Cohenâs Kappa coefficient ( $Îș=0.62$ ). Through iterative discussions, they update their understanding, recode previous papers to include emergent concepts, include new papers in small batche of five and achieve improved reliability. The team collaborated for about two months to discuss findings and potential discrepancies in coding of the articles. Through multiple discussions, discrepancies were reevaluated and corrected, resulting in the final coefficient of $Îș=0.91$ .
### 3.2. Corpus Overview
First, we perform a metadata analysis on the final corpus of 43 articles, focusing on understanding the trustorâs background, the object of trustworthiness, variables, and study type. Our meta analysis reveals that there are six interlinked trustors: users of a specific AI system (20), The number in brackets denote the count of articles corresponding to a specific dimension. citizens (12), developers (4), designers (3), practitioners (2), and public administrators (2). As to the object of trustworthiness, the corpus revolves around studying trustworthiness of AI linked with data sources (19) and institutions developing the AI system (12). The role of trustworthiness as a study variable is almost equally divided among the papers as independent variable (21) and dependent variable (22). Finally, trustworthiness is studied normatively in only 1/3rd of the studies while the remaining 2/3rd studies it in empirical study design.
Second, Figure 2 shows a heatmap of the log-normalized, min-max scaled word occurrences of trustworthiness dimensions by (ISO_Trustworthiness_Vocabulary_2022, 63), in the full texts of the included corpus. Dimensions like quality, transparency, and accuracy occur frequently throughout the full texts, while authenticity, controllability, and resilience are mentioned relatively less. Some articles, like (paraschou2025ties, 112) and (toney, 141), refer to a wide array of dimensions.
<details>
<summary>x2.png Details</summary>
### Visual Description
## Chart Type: Heatmap of AI/ML System Attributes Addressed by Publications
### Overview
This image displays a heatmap illustrating the extent to which various attributes of AI/ML systems are addressed by a collection of academic publications. The Y-axis lists 15 different attributes, while the X-axis lists 39 publications, primarily identified by the first author's name and publication year. The intensity of the blue color in each cell indicates the degree to which a specific attribute is addressed by a particular publication, ranging from 0.0 (white/very light blue) to 1.0 (darkest blue), as indicated by the color bar on the right.
### Components/Axes
**Main Chart Area:** A grid of approximately 15 rows by 39 columns, with each cell colored according to a blue gradient.
**Y-axis (Left Side):**
This axis lists 15 attributes, from top to bottom:
* quality
* transparency
* privacy
* accuracy
* accountability
* safety
* integrity
* reliability
* security
* robustness
* availability
* usability
* authenticity
* controllability
* resilience
**X-axis (Bottom Side):**
This axis lists 39 publications, rotated for readability, from left to right:
* Shailya et al. 2025
* Magana & Ferrario 2025
* Paraschou 2025
* Kinahan et al. 2024
* Pareek et al. 2024
* Kim et al. 2024
* Toney et al. 2024
* Casper et al. 2024
* Maeda & Quanhaase 2024
* Manzini et al. 2024
* Inie et al. 2024
* Wang et al. 2024
* Alpherts et al. 2024
* Scharowski et al. 2024
* Knowles et al. 2023
* Shelby et al. 2023
* Williams & Haring 2023
* Brand 2023
* Lawrence et al. 2023
* Schmitz et al. 2023
* Panigutti et al. 2023
* Kim et al. 2023
* Nannini et al. 2023
* Ferrario et al. 2023
* Zilka et al. 2022
* Engelmann et al. 2022
* Thornton et al. 2022
* Schoeffer et al. 2022
* Liao & Sundar 2022
* Knowles & Richards 2021
* Andrus et al. 2021
* Kim et al. 2021
* Loi & Spielkamp 2021
* Jacovi et al. 2021
* Thornton et al. 2021
* Lakkaraju & Bastani 2020
* Zhang & Dafoe 2020
* Bhatt et al. 2020
* Toreini et al. 2020
* Engelmann et al. 2020
* Mohseni 2019
**Color Bar/Legend (Right Side):**
A vertical color bar is positioned on the far right, indicating the mapping of color intensity to numerical values. The gradient ranges from white (bottom) to dark blue (top).
* 1.0 (darkest blue)
* 0.8
* 0.6
* 0.4
* 0.2
* 0.0 (white/lightest blue)
### Detailed Analysis
The heatmap shows the distribution of focus on different AI/ML attributes across a range of publications from 2019 to 2025 (projected). Darker blue cells indicate a higher degree of focus (closer to 1.0), while lighter blue or white cells indicate less or no focus (closer to 0.0).
**General Trends by Attribute (Rows):**
* **quality:** Shows moderate to high focus in several papers, notably Shailya et al. 2025 (~0.9), Kinahan et al. 2024 (~0.7), Wang et al. 2024 (~0.7), Schmitz et al. 2023 (~0.7), Engelmann et al. 2022 (~0.7), and Jacovi et al. 2021 (~0.7). Many other papers show some focus (~0.2-0.5).
* **transparency:** Widely addressed with moderate to high values. Strongest in Magana & Ferrario 2025 (~0.8), Pareek et al. 2024 (~0.7), Manzini et al. 2024 (~0.7), Alpherts et al. 2024 (~0.7), Shelby et al. 2023 (~0.7), Panigutti et al. 2023 (~0.7), Nannini et al. 2023 (~0.7), Ferrario et al. 2023 (~0.7), Zilka et al. 2022 (~0.7), Schoeffer et al. 2022 (~0.7), Liao & Sundar 2022 (~0.7), Knowles & Richards 2021 (~0.7), Andrus et al. 2021 (~0.7), Kim et al. 2021 (~0.7), Loi & Spielkamp 2021 (~0.7), Zhang & Dafoe 2020 (~0.7), Bhatt et al. 2020 (~0.7), Toreini et al. 2020 (~0.7), Engelmann et al. 2020 (~0.7), and Mohseni 2019 (~0.7).
* **privacy:** Shows high focus in Parashou 2025 (~0.9), Toney et al. 2024 (~0.9), Inie et al. 2024 (~0.9), Brand 2023 (~0.9), Kim et al. 2023 (~0.9), and Lakkaraju & Bastani 2020 (~0.9). Several others show moderate focus.
* **accuracy:** Very widely addressed, with many papers showing high focus. Notably, Kim et al. 2024 (~0.9), Maeda & Quanhaase 2024 (~0.9), Scharowski et al. 2024 (~0.9), Lawrence et al. 2023 (~0.9), Thornton et al. 2022 (~0.9), and Thornton et al. 2021 (~0.9).
* **accountability:** Shows high focus in Casper et al. 2024 (~0.9), Williams & Haring 2023 (~0.9), and Loi & Spielkamp 2021 (~0.9). Many other papers show moderate focus.
* **safety:** Shows high focus in Toney et al. 2024 (~0.9) and Zhang & Dafoe 2020 (~0.9).
* **integrity:** Shows high focus in Williams & Haring 2023 (~0.9) and Engelmann et al. 2020 (~0.9).
* **reliability:** Shows high focus in Maeda & Quanhaase 2024 (~0.9) and Mohseni 2019 (~0.9).
* **security:** Shows high focus in Inie et al. 2024 (~0.9), Brand 2023 (~0.9), and Lakkaraju & Bastani 2020 (~0.9).
* **robustness:** Shows high focus in Shailya et al. 2025 (~0.9), Alpherts et al. 2024 (~0.9), and Kim et al. 2023 (~0.9).
* **availability:** Shows high focus in Panigutti et al. 2023 (~0.9) and Jacovi et al. 2021 (~0.9).
* **usability:** Shows high focus in Knowles et al. 2023 (~0.9) and Schoeffer et al. 2022 (~0.9).
* **authenticity:** Shows high focus in Manzini et al. 2024 (~0.9) and Zilka et al. 2022 (~0.9).
* **controllability:** Shows high focus in Toney et al. 2024 (~0.9) and Loi & Spielkamp 2021 (~0.9).
* **resilience:** Shows high focus in Casper et al. 2024 (~0.9) and Mohseni 2019 (~0.9).
**General Trends by Publication (Columns):**
* **Shailya et al. 2025:** High focus on quality and robustness (~0.9 each).
* **Magana & Ferrario 2025:** High focus on transparency (~0.8).
* **Paraschou 2025:** High focus on privacy (~0.9).
* **Kinahan et al. 2024:** High focus on quality (~0.7).
* **Pareek et al. 2024:** High focus on transparency (~0.7).
* **Kim et al. 2024:** High focus on accuracy (~0.9).
* **Toney et al. 2024:** High focus on privacy, safety, and controllability (~0.9 each).
* **Casper et al. 2024:** High focus on accountability and resilience (~0.9 each).
* **Maeda & Quanhaase 2024:** High focus on accuracy and reliability (~0.9 each).
* **Manzini et al. 2024:** High focus on transparency and authenticity (~0.7 and ~0.9 respectively).
* **Inie et al. 2024:** High focus on privacy and security (~0.9 each).
* **Wang et al. 2024:** High focus on quality (~0.7).
* **Alpherts et al. 2024:** High focus on transparency and robustness (~0.7 and ~0.9 respectively).
* **Scharowski et al. 2024:** High focus on accuracy (~0.9).
* **Knowles et al. 2023:** High focus on usability (~0.9).
* **Shelby et al. 2023:** High focus on transparency (~0.7).
* **Williams & Haring 2023:** High focus on accountability and integrity (~0.9 each).
* **Brand 2023:** High focus on privacy and security (~0.9 each).
* **Lawrence et al. 2023:** High focus on accuracy (~0.9).
* **Schmitz et al. 2023:** High focus on quality (~0.7).
* **Panigutti et al. 2023:** High focus on transparency and availability (~0.7 and ~0.9 respectively).
* **Kim et al. 2023:** High focus on privacy and robustness (~0.9 each).
* **Nannini et al. 2023:** High focus on transparency (~0.7).
* **Ferrario et al. 2023:** High focus on transparency (~0.7).
* **Zilka et al. 2022:** High focus on transparency and authenticity (~0.7 and ~0.9 respectively).
* **Engelmann et al. 2022:** High focus on quality (~0.7).
* **Thornton et al. 2022:** High focus on accuracy (~0.9).
* **Schoeffer et al. 2022:** High focus on transparency and usability (~0.7 and ~0.9 respectively).
* **Liao & Sundar 2022:** High focus on transparency (~0.7).
* **Knowles & Richards 2021:** High focus on transparency (~0.7).
* **Andrus et al. 2021:** High focus on transparency (~0.7).
* **Kim et al. 2021:** High focus on transparency (~0.7).
* **Loi & Spielkamp 2021:** High focus on transparency, accountability, and controllability (~0.7, ~0.9, and ~0.9 respectively).
* **Jacovi et al. 2021:** High focus on quality and availability (~0.7 and ~0.9 respectively).
* **Thornton et al. 2021:** High focus on accuracy (~0.9).
* **Lakkaraju & Bastani 2020:** High focus on privacy and security (~0.9 each).
* **Zhang & Dafoe 2020:** High focus on transparency and safety (~0.7 and ~0.9 respectively).
* **Bhatt et al. 2020:** High focus on transparency (~0.7).
* **Toreini et al. 2020:** High focus on transparency (~0.7).
* **Engelmann et al. 2020:** High focus on transparency and integrity (~0.7 and ~0.9 respectively).
* **Mohseni 2019:** High focus on transparency, reliability, and resilience (~0.7, ~0.9, and ~0.9 respectively).
### Key Observations
* **Transparency and Accuracy are Most Addressed:** The rows for "transparency" and "accuracy" show the highest density of blue cells, indicating these attributes are frequently and significantly addressed across a large number of publications. "Transparency" has a particularly consistent moderate-to-high focus across almost all papers.
* **Privacy, Security, Accountability, and Reliability are also prominent:** These attributes also show a strong presence, with several papers dedicating high focus (dark blue cells).
* **Less Consistently Addressed Attributes:** Attributes like "quality," "safety," "integrity," "robustness," "availability," "usability," "authenticity," "controllability," and "resilience" are addressed by specific papers with high intensity, but not as broadly across the entire set of publications as "transparency" or "accuracy." Many cells for these attributes are white or very light blue, indicating less or no focus in many papers.
* **No Single Publication Covers All Attributes Extensively:** While some publications address multiple attributes with high intensity (e.g., Toney et al. 2024, Loi & Spielkamp 2021), no single paper shows dark blue cells across all 15 attributes. This suggests specialization or a focused scope in individual research efforts.
* **Recent Publications (2024-2025) Show Diverse Focus:** The most recent publications (e.g., Shailya et al. 2025, Toney et al. 2024, Casper et al. 2024) are addressing a wide range of attributes, including some of the less common ones like robustness, safety, accountability, controllability, and resilience.
* **Older Publications (2019-2020) also show diversity:** Mohseni 2019, Engelmann et al. 2020, and Lakkaraju & Bastani 2020 also show high focus on a diverse set of attributes, indicating that these concerns are not entirely new.
### Interpretation
This heatmap provides a valuable overview of the research landscape concerning AI/ML system attributes. The data suggests that:
1. **Core Concerns:** "Transparency" and "accuracy" appear to be foundational or consistently critical concerns in AI/ML research, as almost every publication touches upon them to some degree. This could be due to their direct impact on system performance, trustworthiness, and interpretability.
2. **Emerging or Specialized Concerns:** Attributes like "safety," "integrity," "robustness," "availability," "usability," "authenticity," "controllability," and "resilience" are addressed more selectively. This might indicate that these are either more niche areas of research, or they are gaining prominence more recently, or they are highly context-dependent and thus not universally applicable to all AI/ML systems. The presence of dark blue cells for these attributes in various papers, especially recent ones, suggests a growing recognition of their importance.
3. **Interdisciplinary Nature:** The diverse set of attributes reflects the multifaceted challenges and considerations in developing and deploying AI/ML systems. It highlights that a holistic approach requires addressing not just technical performance (accuracy, robustness) but also ethical, social, and human-centric aspects (privacy, accountability, safety, usability, transparency).
4. **Potential Research Gaps:** The numerous white cells, particularly for attributes beyond transparency and accuracy, point to potential areas where more research is needed or where existing research is not yet widely integrated. For instance, while "resilience" is highly addressed by Casper et al. 2024 and Mohseni 2019, many other papers do not seem to focus on it, suggesting a potential gap in comprehensive coverage across the field.
5. **Evolution of Focus:** While the heatmap doesn't explicitly show trends over time for *all* attributes, the presence of high values for diverse attributes in both older (e.g., Mohseni 2019) and newer (e.g., Shailya et al. 2025) papers suggests that while some concerns are perennial, the specific emphasis and combinations of attributes addressed by researchers evolve. The projected 2025 papers indicate continued attention to a broad spectrum of attributes.
In essence, the heatmap demonstrates that while certain attributes like transparency and accuracy are almost universally considered, the broader landscape of AI/ML system attributes is addressed with varying degrees of intensity and specialization across different research publications. This highlights both the progress and the ongoing challenges in building comprehensive and responsible AI systems.
</details>
Figure 2. Heatmap of the normalized frequencies of trustworthiness dimensions in each paper of the final corpus ( $N=43$ ).
Third, we broaden the analysis and discover themes across the initial keyword-based selection of 235 articles, without extended abstracts, through topic modeling. We apply BERTopic (grootendorst2022bertopic, 53), which leverages c-TF-IDF and the capabilities of the transformer-based language model BERT (devlin2018bert, 29), to analyze topics that are both coherent and contextually meaningful. The generated topics are analyzed using two visualizations as shown in Figure 3. On the one hand, we present the normalized topic distributions over time using a heatmap to capture trends and shifts across publication years. We observe a consistent prevalence and recent increase of topics relating to âexplanationsâ and âaccountabilityâ, while there is a relative decline in topics referring to ethics and norms. On the other hand, we visualize the representations and relative distance of the topics in a 2D plot. Notably, topics surrounding âfairness,â âtransparency,â and âethicsâ lie in closer proximity to each other, while they are more distanced from the topic around âexplanations.â Also, we find that although fairness and bias are key ethical concerns, the considerable distance between fairness-related topics and ethical themes suggests that papers tend to emphasize one while giving limited attention to the other.
Topic 1 (72): explanations, human, decision, language, study
Topic 2 (85): transparency, accountability, public, human, technologies
Topic 3 (49): fairness, bias, algorithmic, fair, discrimination
Topic 4 (29): ethics, ethical, moral, responsible, norms
<details>
<summary>x3.png Details</summary>
### Visual Description
## Combined Chart: Topic Evolution and Spatial Distribution
### Overview
The image presents two distinct but related visualizations side-by-side. On the left is a heatmap illustrating the temporal evolution or prevalence of four different topics (Topic 1 to Topic 4) across eight years (2018 to 2025). On the right is a 2D scatter plot, or bubble chart, showing the spatial distribution and relative importance (indicated by size) of these same four topics within a dimensional space defined by D1 and D2.
### Components/Axes
**Left Chart: Heatmap (Topic Prevalence Over Time)**
* **Type:** Heatmap
* **Y-axis (Vertical):** Labeled with years, from bottom to top: 2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025.
* **X-axis (Horizontal):** Labeled with topics, from left to right, rotated for readability: Topic 1, Topic 2, Topic 3, Topic 4.
* **Color Bar Legend (Right side of heatmap):** A vertical color gradient bar indicating the value associated with each shade of blue.
* 0.6 (darkest blue)
* 0.4 (medium-dark blue)
* 0.2 (medium-light blue)
* 0.0 (lightest blue/white)
* The gradient transitions smoothly from white/very light blue at 0.0 to a deep, dark blue at 0.6.
**Right Chart: Bubble Chart (Topic Spatial Distribution)**
* **Type:** Bubble Chart / Scatter Plot
* **Horizontal Axis (X-axis):** Labeled "D1". This axis spans from negative values on the left to positive values on the right, with the origin (0) at the center.
* **Vertical Axis (Y-axis):** Labeled "D2". This axis spans from negative values at the bottom to positive values at the top, with the origin (0) at the center.
* **Data Points:** Four dark blue circular bubbles, each labeled with a topic name (Topic 1, Topic 2, Topic 3, Topic 4). The size of the bubbles varies.
### Detailed Analysis
**Heatmap Data (Approximate values based on color intensity):**
| Year | Topic 1 (Value) | Topic 2 (Value) | Topic 3 (Value) | Topic 4 (Value) |
| :--- | :-------------- | :-------------- | :-------------- | :-------------- |
| 2018 | ~0.10 (Light) | ~0.10 (Light) | ~0.10 (Light) | ~0.55 (Dark) |
| 2019 | ~0.10 (Light) | ~0.10 (Light) | ~0.10 (Light) | ~0.55 (Dark) |
| 2020 | ~0.10 (Light) | ~0.10 (Light) | ~0.25 (Medium) | ~0.10 (Light) |
| 2021 | ~0.10 (Light) | ~0.50 (Dark) | ~0.10 (Light) | ~0.10 (Light) |
| 2022 | ~0.25 (Medium) | ~0.10 (Light) | ~0.10 (Light) | ~0.10 (Light) |
| 2023 | ~0.25 (Medium) | ~0.25 (Medium) | ~0.25 (Medium) | ~0.10 (Light) |
| 2024 | ~0.35 (Med-Dark)| ~0.50 (Dark) | ~0.10 (Light) | ~0.10 (Light) |
| 2025 | ~0.50 (Dark) | ~0.50 (Dark) | ~0.10 (Light) | ~0.10 (Light) |
**Heatmap Trends:**
* **Topic 1:** Shows a clear upward trend. It remains low (~0.10) from 2018-2021, then increases to medium (~0.25) in 2022-2023, and becomes significantly high (~0.35 to ~0.50) in 2024-2025.
* **Topic 2:** Exhibits a fluctuating pattern with distinct peaks. It is low (~0.10) for 2018-2020, then spikes to high (~0.50) in 2021, drops back to low (~0.10) in 2022, rises to medium (~0.25) in 2023, and reaches high levels again (~0.50) in 2024-2025.
* **Topic 3:** Generally maintains a low prevalence (~0.10) throughout the period, with slight increases to medium (~0.25) in 2020 and 2023.
* **Topic 4:** Shows a strong downward trend. It is very high (~0.55) in 2018-2019, then drops sharply to low levels (~0.10) from 2020 onwards and remains low.
**Bubble Chart Data (Approximate coordinates and relative sizes):**
* **Topic 1:**
* Position: Top-left quadrant.
* Coordinates (D1, D2): (~-0.7, ~0.6)
* Size: Medium-large.
* **Topic 2:**
* Position: Bottom-right quadrant.
* Coordinates (D1, D2): (~0.7, ~-0.4)
* Size: Largest.
* **Topic 3:**
* Position: Top-right quadrant.
* Coordinates (D1, D2): (~0.6, ~0.6)
* Size: Medium.
* **Topic 4:**
* Position: Bottom-left quadrant, closest to the origin.
* Coordinates (D1, D2): (~-0.3, ~-0.5)
* Size: Smallest.
### Key Observations
* **Temporal Dynamics:** Topics show diverse temporal patterns. Topic 4 is a declining topic, Topic 3 is consistently low, while Topic 1 is an emerging topic, and Topic 2 is a recurring/fluctuating topic with high recent prevalence.
* **Topic Importance (Bubble Size):** The bubble chart suggests Topic 2 is the most prominent overall, followed by Topic 1, then Topic 3, and Topic 4 is the least prominent. This generally aligns with their recent or overall prevalence in the heatmap.
* **Spatial Grouping:**
* Topics 1 and 3 are both in the positive D2 region (upper half), suggesting some shared characteristic along Dimension 2.
* Topics 2 and 4 are both in the negative D2 region (lower half), suggesting another shared characteristic along Dimension 2.
* Topics 1 and 4 are in the negative D1 region (left half).
* Topics 2 and 3 are in the positive D1 region (right half).
* **Opposing Topics:** Topic 1 and Topic 2 appear diagonally opposite, as do Topic 3 and Topic 4, suggesting contrasting characteristics across both D1 and D2 dimensions.
### Interpretation
The two charts provide a comprehensive view of topic dynamics and relationships. The heatmap reveals the "life cycle" of each topic over an eight-year span, while the bubble chart offers a static "snapshot" of their relative positions and importance in a conceptual space.
* **Topic 4** appears to be a "legacy" topic, highly prevalent in earlier years (2018-2019) but quickly fading into insignificance by 2020. Its small bubble size in the D1/D2 space reflects this diminished importance. Its position in the bottom-left quadrant (negative D1, negative D2) might indicate it represents an older, perhaps less innovative or less relevant theme compared to others.
* **Topic 3** maintains a consistently low profile throughout the years, with only minor fluctuations. Its medium-sized bubble, despite low prevalence, might suggest a foundational or niche topic that is always present but never dominant. Its position in the top-right quadrant (positive D1, positive D2) suggests it shares characteristics with Topic 1 along D2 but is distinct along D1.
* **Topic 2** is a highly dynamic topic, experiencing significant surges in prevalence (2021, 2024-2025). Its large bubble size indicates its overall high importance or impact. Its position in the bottom-right quadrant (positive D1, negative D2) suggests it is distinct from Topic 1 and 3, possibly representing a different domain or approach. Its recurring high prevalence suggests it's a persistent or cyclical theme.
* **Topic 1** is an "emerging" topic, showing a clear and strong upward trend in prevalence, especially in the later years (2022-2025). Its medium-large bubble size, while not the largest, suggests growing importance, potentially reflecting its future relevance. Its position in the top-left quadrant (negative D1, positive D2) places it in opposition to Topic 2, but sharing some D2 characteristics with Topic 3. This could imply it's a new direction or innovation.
The D1 and D2 dimensions likely represent underlying semantic or thematic characteristics that differentiate the topics. For instance, D1 could represent a spectrum from "traditional" (negative D1) to "innovative" (positive D1), while D2 could represent "technical" (positive D2) to "social" (negative D2).
* If D1 is "Traditional (-) to Innovative (+)", then Topics 1 and 4 are more traditional, while Topics 2 and 3 are more innovative.
* If D2 is "Technical (+) to Social (-)", then Topics 1 and 3 are more technical, while Topics 2 and 4 are more social.
This interpretation suggests a shift in focus over time, with Topic 4 (perhaps an older, traditional-social theme) declining, and Topic 1 (an emerging, traditional-technical theme) gaining prominence, while Topic 2 (an innovative-social theme) remains highly impactful. The interplay between these temporal trends and spatial relationships provides a rich understanding of the topic landscape.
</details>
Figure 3. Visualizations of topics generated using BERTopic applied to the initial keyword-based selection of articles ( $N=235$ ). First, the heatmap (left) shows normalized topic distributions, grouped by year. Values refer to the proportion of documents associated with the topic. Second, the intertopic distance plot (right) shows the relative positions of topics in a 2D space. The representations of topics are reduced to two dimensions using the uniform manifold approximation and projection dimensionality reduction technique, enabling the generation of a 2D plot to visualize the topics.
## 4. Understanding of AI Trustworthiness in the AIES & FAccT Community
### 4.1. Definition or Conceptualization
To understand how to understand an AI systemâs trustworthiness, it is crucial to examine how we define or conceptualize it. Definition refers to how trustworthiness is formally described and articulated in the selected papers and conceptualization encompasses the broader understanding of trustworthiness, e.g., considering its role in systems.
First, the AIES & FAccT community presents a rich and multifaceted understanding of AI trustworthiness, as evidenced across numerous studies. Through our coding and thematic analysis, we identify seven key conceptualization themes, summarized in Table 1.
Analysis of the conceptual landscape reveals several critical insights about how AI trustworthiness is understood in contemporary scholarship. First, there exists a fundamental tension between transparency-focused approaches (T1) and technical robustness perspectives (T6), suggesting competing philosophies about whether trustworthiness derives primarily from system interpretability or from demonstrated reliability regardless of internal opacity. Second, the prominence of anthropomorphization themes (T3) alongside ethical considerations (T4) indicates growing recognition that trust in AI cannot be divorced from human psychological tendencies to perceive social relationships with technology, a finding that challenges purely technical or procedural definitions. Third, the emphasis on broader societal implications (T5) and regulatory ecosystems signals an important shift from treating trustworthiness as an individual system property to understanding it as embedded within institutional contexts where harms, accountability, and public opinion are distributed unevenly across populations. Finally, the reliance on established frameworks like Mayerâs model and NIST/ISO definitions (T7) alongside emerging concepts of perceived parasocial relationships suggests the field is simultaneously borrowing from organizational trust theory while grappling with AI-specific phenomena that existing models may not adequately capture. These insights collectively reveal that AI trustworthiness is conceptualized not as a unified construct but as a contested terrain where technical, psychological, ethical, and societal dimensions intersect and sometimes conflict.
Table 1. Understanding of AI Trustworthiness: Key themes in definitions and conceptualizations.
| T1: Emphasis on transparency as a foundational aspect of trustworthiness | Trustworthiness cues (liao2022designing, 86), Good explanation (rayyan-64504518, 105), Open communication (brandj, 16), Transparency, Provenance & connections between them (rayyan-64504562, 139), Computational reliabilism & anti-monitoring (Ferrario, 45), Expressing Uncertainty (Sunnie2024a, 73), Transparency with human oversight (rayyan-64504523, 111), Explanation Trustworthiness (rayyan-64504630, 103, 128), Transparency of algorithmic tools deployed in the criminal justice system (rayyan-64504664, 155) |
| --- | --- |
| T2: The importance of benchmarks, rigorous auditing and compliance mechanisms | White- and outside-the box audits (Stephen2024, 21), Certification Labels (nicolas, 121), Performance benchmarks and evaluation measures (rayyan-64504620, 83), Mapping regulatory guidelines to philosophical account of accountability (rayyan-64504624, 90) |
| T3: Perceived Anthropomorphization | (de)Anthropomorphized description of AI system (Nanna2024, 62), Perceived Parasocial Relationships (rayyan-64504506, 92), Facial AI inferences (Severin2022, 40) |
| T4: Ethical considerations, including fairness and respect for user intent | Inclusive & intersectional algorithms (kime, 72), Data leakage & reproducibility (Sean2024, 75), Model interpretability (rayyan-64504617, 81), Understanding of harms and how (and by whom) this has been informed (Bran2023, 76), Interaction between risks and the dispositions of the trustee (rayyan-64504660, 151), Informational fairness (10.1145/3531146.3533218, 126), Achieving fairness (rayyan-64504644, 125), Fairness by design (rayyan-64504611, 61), Fair data procurement (mckane, 4) |
| T5: Broader societal implications: Regulatory Ecosystem and Accountability caused by harms | Social trustworthiness score (Severin2019, 39), Trustworthiness criteria for policy-making (rayyan-64504432, 3), Public trust in AI & institutional trust (Bran2021, 77), AI Assistant design & Organizational practices and third-party governance (rayyan-64504509, 97), sociotechnical harms (rayyan-64504649, 129), Public opinion (rayyan-64504662, 153), Safe, reliable, and acceptable to users and public (magana2025frameworks, 93) |
| T6: Justified reliance on AIâs reliability and technical robustness | Reasonable trust in modelâs output (Umang2020, 13), Warranted Trust (jacovi, 65), AI system must maintain to function correctly within its context of application (ferrario2025trustworthiness, 44) |
| T7: Mayerâs (mayer1995integrative, 98) and Lee & Seeâs model (lee2004trust, 84) & NIST/ISO definition | Ability, Benevolence & Integrity (Sunnie2023, 74, 86, 73, 149, 140, 113, 112), ABI+ (ABI, Predictability) framework (toreini, 142), NIST & ISO Definition (toney, 141) |
We also study the trustworthiness dimensions defined by the ISO trustworthiness standard (ISO_Trustworthiness_Vocabulary_2022, 63) to ensure that our analysis is based on internationally recognized standards. âTransparencyâ is the most frequently mentioned ISO dimension, appearing in 21 out of 43 papers, followed by âaccountabilityâ and âexplainabilityâ in 12 papers. The less frequently mentioned ISO dimensions are âcontrollabilityâ (5), âresilienceâ (4), ârobustnessâ (4), âsecurityâ (4), âsafetyâ (3), âreliabilityâ (3), âusabilityâ (3), and âprovenanceâ (2).
While many studies implicitly link trustworthiness to accuracy, reliability, transparency, and fairness, increasing attention is being given to more nuanced conceptualizations, such as stakeholder-centric approaches and systemic perspectives. Frequent references to ISO dimensions, particularly transparency, accountability, and explainability, underscore their importance in building trustworthy AI systems. However, a more comprehensive and standardized approach to defining and measuring AI trustworthiness is needed for the responsible development and deployment of AI systems.
### 4.2. Drivers
(anjomshoae2019explainable, 7) have underlined the importance of considering the intended purpose when studying the trustworthiness of AI systems. To this end, this scoping review seeks to understand why the reviewed studies focus on understanding trustworthiness. Most of these works stated their motivation or intended purpose for studying trustworthiness. Table 1 lists the drivers of the 43 papers included in the review, with some papers having more than one driver.
Table 1. The drivers of AI trustworthiness of the primary studies covered by the review.
| D1: Conformity with regulations & Guidelines and ethical considerations | (Stephen2024, 21, 40, 75, 105, 111, 126, 121, 141, 16, 61, 83, 90, 125, 129, 139, 5, 77, 103, 61, 155, 141, 4, 128, 44) |
| --- | --- |
| D2: Promoting honesty | (Severin2019, 39, 151, 3) |
| D3: Trust | (Umang2020, 13, 45, 62, 65, 73, 74, 86, 92, 97, 139, 126, 113, 140, 149, 81, 103, 76, 65, 151, 3, 77, 142, 153, 93, 112) |
| D4: Economic growth & Cultural values | (Severin2019, 39, 16) |
| D5: Reproducibility | (Sean2024, 75) |
| D6: Accuracy | (Aida2023, 94, 72, 3) |
Increasing usersâ trust in the system, ethical considerations, and conformity with regulations such as those of (EU_AI_Ethics_Guidelines, 42, 127, 136, 137, 66, 144, 43, 147), and accuracy are among the listed motivations for the explanations. The table reveals that trust and ethical considerations are the most prominent drivers of studying AI trustworthiness. Naturally, trust and trustworthiness go hand in hand to increase usersâ confidence in the system by understanding how its reasoning mechanism works (anjomshoae2019explainable, 7). In applications requiring human-AI interaction, honesty and accuracy are among the main drivers for trustworthiness to ensure that AIâs decision-making is reliable and fair (10.1145/3531146.3533218, 126, 113). For public AI systems, trustworthiness drivers are often linked with conformity with regulations and guidelines, and economic growth and cultural values (Severin2019, 39, 16, 77, 90, 153). Finally, (Sean2024, 75) identify reproducibility of their results as a driver for studying trustworthiness for their system. Overall, these studies provide a holistic overview of the key drivers for studying AI trustworthiness.
### 4.3. Measurement and Verification & Validation
(jacobs2021measurement, 64) have introduce measurement theory for AI/ML systems, where validation and verification are key to ensure we are measuring what we are trying to measure. In the context of AI trustworthiness, measurement refers to quantifying specific attributes or characteristics of the AI system, typically aligned with ISO trustworthiness dimensions. Following ISO 9001 (dnv2015iso, 32), verification involves confirming through examination and provision of objective evidence that the specified trustworthiness dimensions are fulfilled. Validation is similar to verification, but must be confirmed under real-world usage conditions.
Measurement. To analyze the measurement techniques for AI trustworthiness, we follow (schlicker2025we, 122) âs distinction of actual and perceived trustworthiness. (schlicker2025we, 122) define a systemâs actual trustworthiness as a latent construct that indicates the true value of a systemâs trustworthiness (in alignment with Realistic Accuracy Model (funder1995accuracy, 49)), e.g., benevolence, integrity, and ability (mehrotrareview, 99, 84, 98). Similarly, perceived trustworthiness reflects the result of a trustorâs assessment of the trusteeâs actual trustworthiness.
Table 1. The measurement of AI trustworthiness of the primary studies covered by the review.
| Likert scale | McKnight et al.[(HARRISONMCKNIGHT2002297, 58)] (Sunnie2024a, 73), Jian et al. [(Jian01032000, 67)] (rayyan-64504525, 113), Perceived trustworthiness (Own scale) (10.1145/3531146.3533218, 126) |
| --- | --- |
| Reliance behavior | Change in peopleâs trust judgment (liao2022designing, 86), Agreement between a participantâs answer and that of the AI system (Sunnie2024a, 73, 74, 93) (rayyan-64504525, 113, 149, 83) |
| Actual cues (e.g., precision, recall, and f1-scores) | Positive predictive value (kime, 72), LDA topic modelling (Severin2019, 39), Stochastic model with surrogates (rayyan-64504518, 105), (Aida2023, 94, 142, 75, 21, 111), Metrics for plausibility and faithfulness scores (shailya2025lext, 128) |
| Assessment of impact and alignment with values and concerns | Ethical charters & internal review bodies (rayyan-64504509, 97), sociotechnical harms taxonomy, Moral groundings for public transparency (rayyan-64504624, 90) (rayyan-64504649, 129) |
| Levels of monitoring activities | Monitoring-avoiding relation between trustor & trustee (Ferrario, 45) (toreini, 142), Trustworthiness level across system accuracy ranges (ferrario2025trustworthiness, 44) |
| Communication style and anthropomorphic features | Roleplaying & reciprocal engagement (rayyan-64504506, 92) |
| Qualitative assessment | Interviews & Surveys (nicolas, 121, 13, 139, 62, 81, 149, 3, 153, 40, 4), User satisfaction (rayyan-64504664, 155) |
| Model Cards and AI Assessment Catalog | Information about data leakage & selective inference for EEG (Sean2024, 75), AI Assessment Catalog (rayyan-64504644, 125), NIST AI Risk Management Framework & other national frameworks (toney, 141) |
| Mediator & Trustee role | Social, Institutional and Temporal Embeddedness, Situational Normality, Structural Assurance, Motivation and ABI (lauren, 140) |
Analysis of measurement strategies reveals several important insights about the operationalization of AI trustworthiness. First, the field exhibits a problematic conflation between measuring trust (a psychological state) and measuring trustworthiness (system properties that warrant trust). Reliance behavior and Likert scales capture user perceptions while technical metrics assess objective capabilities, yet the relationship between these remains theoretically underdeveloped. Second, the diversity of approaches, from LDA topic modeling to ethical charters to role-playing exercises, suggests that trustworthiness is inherently multi-faceted and resists reduction to single metrics, yet this proliferation also indicates a lack of consensus on what actually constitutes valid measurement. Third, the emergence of documentation artifacts like model cards (Sean2024, 75) and AI assessment catalogs (rayyan-64504644, 125) represents an important shift toward longitudinal, context-aware evaluation rather than snapshot assessments, implicitly acknowledging that trustworthiness is not static but emerges through ongoing scrutiny. Finally, the inclusion of mediator and trustee role factors (social embeddedness, situational normality, structural assurance) indicates growing sophistication in recognizing that measurement must account for the relational and contextual dimensions of trust rather than treating it as residing solely in the AI system. Together, these developments suggest the field is moving toward more holistic measurement paradigms, though integration across approaches remains elusive.
Verification and/or Validation. More than half (34 out of 43) of the papers did not (explicitly) provide or adopt verification or validation approaches. Among the remaining 9 papers, surveys are the most widely used verification methodology, appearing in 5 papers (Severin2022, 40, 62, 126, 113, 153). E.g., (rayyan-64504525, 113) conduct a survey-based between-subject experiments involving 300 participants to investigate trust development, erosion, and recovery during AI-assisted decision-making. Additionally, (rayyan-64504662, 153) pre-register their experiments to ensure methodological rigor and transparency. (rayyan-64504617, 81) and (rayyan-64504432, 3) verify trustworthiness measurement through user studies with domain experts. Finally, only one paper employs a mixed method consisting of qualitative interviews and a subsequent survey to collect quantitative data (nicolas, 121).
The gap in current research, i.e., the insufficient attention to validation under real-world constraints and the lack of alignment between verification practice and trustworthiness dimensions, can be attributed to two primary factors: the lack of real-world data or representative scenarios (rayyan-64504434, 5) and the absence of multi-disciplinary collaboration (brundage2020toward, 17). The 9 papers that do so represent important but limited work in controlled settings, not diverse real-world environments. Our thoughts resonate with (rayyan-64504434, 5) and (brandj, 16) who have demonstrated challenges in obtaining representative datasets and how siloed approaches can miss critical sociotechnical dimensions for verification or validation.
### 4.4. Application Scenarios
AI techniques are applied across domains, including high-stakes ones, making trustworthiness vital to their success. The AIES & FAccT community demonstrates a comprehensive and diverse exploration of trustworthiness across various application scenarios; see Table 2. High-stake domains, including healthcare and medical applications, financial and economic systems, human resources, governance, law, and justice, are the most prominent application scenarios of studying trustworthiness of AI systems (22 out of 43). 12 out of 43 papers do not centre on a specific application scenario but rather generally discuss AI applications and their impact on society. Additionally, (rayyan-64504518, 105) and (rayyan-64504523, 111) examine AI policies with a focus on explainability of AI systems. (rayyan-64504518, 105) perform a thematic and gap analysis of policies and standards on explainability in EU, US, and UK and provide a set of recommendations on how to address explainability in regulations for AI systems. (rayyan-64504523, 111) analyze the EU AI Act (european2021proposal, 41) and argue that it neither mandates explainable AI nor bans the use of black-box AI systems; instead, it seeks to achieve its stated policy objectives with a focus on transparency and human oversight.
Table 2. The application scenarios of AI trustworthiness of the primary studies covered by the review.
| Healthcare and medical applications | (liao2022designing, 86, 121, 142, 73, 75, 80, 5, 4) |
| --- | --- |
| Financial and economic systems | (Umang2020, 13, 16, 5, 126, 121, 4) |
| Human resource | (Severin2022, 40, 121, 103, 5, 4) |
| Governance, law, and justice | (rayyan-64504620, 83, 90, 103, 153, 61, 81, 39, 155, 151, 125, 77) |
| Content moderation and information systems | (Umang2020, 13, 91) |
| Personal assistance and interaction | (Lydia2024, 96, 97, 121) |
| Environmental science | (rayyan-64504432, 3, 74, 139) |
| Domain agnostic | (Stephen2024, 21, 62, 65, 76, 77, 113, 141, 72, 129, 92, 140, 142, 103, 40) |
| AI policies and standards | (rayyan-64504518, 105, 111) |
| Facial recognition | (kime, 72) |
| Software engineering and programming | (wangr, 149) |
### 4.5. Intended, Embedded, and Realized Values
Our analysis reveals a significant gap between the intended, embodied, and realized values of AI systems. While many are developed with good intentions, e.g., promoting transparency, fairness, and accountability, these values are not always reflected in the final product or its impact on society.
Intended Values. The most frequently cited intended values were transparency (26), fairness (19), and accountability (12) 39 articles had more than one intended value for studying AI trustworthiness.. Other important values included privacy (8), security (5), and human oversight (5). E.g., (Umang2020, 13) emphasize the importance of transparency and trustworthiness as the intended values of explainable AI. Similarly, (Stephen2024, 21) highlight the role of AI audits in identifying risks and improving transparency, and (Severin2019, 39) examine the intended value of the social credit system to promote honesty and trust in Chinese society. In abidance with the AI act, (rayyan-64504523, 111) examine the role of explainable AI and state the intended value to ensure AI is trustworthy by focusing on transparency and human oversight. This includes enabling users to interpret outputs and manage risks associated with AI systems.
Embodied Values. The embodied values in studying AI trustworthiness are often shaped by the techniques used, the stakeholders involved, and the context of deployment. (rayyan-64504518, 105) examine the embodied value of explainable AI, which is determined by how well it can fulfill its intended purpose without compromising other desiderata, such as accuracy and privacy. (nicolas, 121) examine the embodied value of certification labels, which is determined by how well they can fulfill their intended purpose by reflecting the AI systemâs compliance with trustworthiness criteria. (rayyan-64504432, 3) examine the embodied value of AI models, determined by how well they can generate explanations that help understand the relations between visual elements in street view imagery and socio-economic variables such as housing pricing.
Realized Values. The realized values of AI systems can differ significantly from their intended values due to various factors, including design flaws, unintended consequences, and societal biases. Some examples of how realized values can deviate from intended values: (Ferrario, 45) find that the realized value of explainability in AI systems may differ from its intended value, particularly when explainability does not lead to a reduction in monitoring or when it fosters an unjustified belief in the AIâs trustworthiness and (Sunnie2023, 74) find that the realized value of AI systems may differ from their intended value due to factors such as malicious manipulation of trustworthiness cues or ill-designed cues that mislead users.
The gap between intended, embodied, and realized values in AI systems raises important questions about the effectiveness of current approaches to embedding values in AI design for studying AI trustworthiness. It highlights the need for more robust methods to evaluate AIâs societal impact.
## 5. Discussion: Challenges, Open Questions, and Future Directions
This review has examined the multifaceted nature of AI trustworthiness as articulated within the AIES & FAccT community. This section examines the challenges, open questions, and future directions that arise from our analysis.
### 5.1. Why Do We Need to Rethink AI Trustworthiness?
Several critical gaps and limitations emerge from our analysis in Section 4.1. Despite this rich conceptual terrain, the current literature inadequately address: who defines trustworthiness criteria, whose harms count as significant, and how user intent is interpreted remain largely unexamined despite ethical considerations being prominent. Additionally, while there is convergence around components like accuracy and fairness linking to ISO dimensions, the literature largely neglects the power dynamics and structural inequalities that fundamentally shape trust relationships in AI systems (dobbe2021hard, 33).
Additionally, while benchmarks and auditing mechanisms (T2) are emphasized, there is insufficient attention to how compliance-focused approaches might incentivize âethics washingâ where organizations meet procedural requirements without substantively addressing underlying trust deficits. The conceptualizations also struggle with context-dependency: trustworthiness defined through computational reliability may be necessary but insufficient in domains like criminal justice (as noted in T1) where procedural justice and legitimacy matter as much as accuracy, yet the frameworks provide limited guidance on how to weight these competing demands situationally.
Current conceptualizations, despite their evolution from purely technical metrics to sociotechnical constructs, fail to fully account for how trust can erode or strengthen over time, particularly in response to system failures or successful interventions (dzhelyova2012temporal, 37). An emphasis on system-level characteristics and user perceptions, while important, has come at the expense of examining broader institutional and systemic factors that influence trustworthiness, such as regulatory frameworks, corporate governance structures, and market incentives. Without addressing these gaps, current approaches to building trustworthy AI risk perpetuating existing biases and power imbalances while failing to establish sustainable trust relationships across user communities (osasona2024reviewing, 109).
Based on our analysis of the corpus, we propose the following measures:
1. Future research must implement longitudinal studies that track trust dynamics over time, and design systems with mechanisms to rebuild trust after failures rather than focusing solely on initial trust establishment.
1. At the heart of AI trustworthiness lies a fundamental distinction between perceived and actual trustworthiness (schlicker2022calibrated, 124). Future research should clearly distinguish them to foster a more unified understanding of AI trustworthiness.
1. Develop comprehensive trustworthiness frameworks that explicitly incorporate institutional accountability measures, regulatory compliance mechanisms, and transparent corporate governance structures.
### 5.2. Is AI Trustworthiness Just a Checklist?
A critical analysis of the drivers of AI trustworthiness revealed oversights in how the field currently conceptualizes motivations for studying this crucial aspect. While the reviewed literature identifies several key drivers (see Section 4.2), the understanding appears superficial and fails to address several fundamental concerns.
First, the heavy emphasis on regulatory compliance and ethical guidelines, while important, suggests a potentially reactive rather than proactive approach to trustworthiness (tyler2001trust, 143). Many studies treat regulations as a checklist (see representative articles for D3 in Table 1) rather than engaging with the deeper philosophical and societal implications of AI trustworthiness. This compliance-driven approach risks creating a false sense of security while potentially missing emerging challenges not yet addressed by current regulations, such as the evolving nature of biases in dynamic systems, the challenge of ensuring explainability in complex models (saeed2023explainable, 118), attributing accountability in autonomous decision-making (busuioc2021accountable, 18), and impacts on human agency and societal power structures (santoni, 119). In real-world industry settings, compliance requirements often act as crucial entry points and motivators for responsible AI work (diaz2023connecting, 30). When organizational resources and attention are limited, framing trustworthiness as a compliance issue can secure support and budget allocations (caldwell2003organizational, 19).
Second, the classification of drivers reveals a bias toward accuracy and promotion of trust in AI, with a limited focus on end-user needs and societal impacts. While âtrustâ is frequently cited as a driver, there is an inadequate exploration of building an appropriate level of trust in AI, resulting in avoidance of over- and under-trust (10.1145/3610578, 102, 99).
Third, we found vagueness in debates about the safety behavior of AI systems and identifying & diagnosing safety risks in complex social contexts. Safety has been designated as a key component of AI trustworthiness both in the EU AI Act (european2021proposal, 41) and ISO (ISO_Trustworthiness_Vocabulary_2022, 63) & NIST (NIST_AI_Risk_Management_Framework, 106) guidelines. However, in our analysis, system safety does not appear as a prominent measure for trustworthiness except for the mathematical formalism (toney, 141). This raises two key questions: (i) Is safety marred by an underlying vagueness where it is hard to establish whether a system is safe or not (dobbe2021hard, 33)? And (ii) does it even make sense to study trustworthiness of an AI system without exploring its safety especially from a sociotechnical lens?
Finally, drivers related to long-term sustainability, environmental impact, and social justice are limited in our analyzed papers. While accuracy is listed as a driver, there is limited attention to the complex relationship between technical accuracy and real-world effectiveness. This limited understanding of drivers profoundly impacts how trustworthiness is studied and implemented in AI systems. Without a more nuanced understanding of why trustworthiness matters, the field risks producing solutions that address surface concerns while failing to engage with deeper systemic challenges. Hence, we propose the following measures:
1. The ideal approach could involve using compliance frameworks as practical starting points while simultaneously cultivating organizational cultures that recognize the broader societal implications of AI systems beyond regulatory requirements.
1. Develop frameworks and metrics that address the appropriateness of trust, ensuring users understand both the capabilities and limitations of AI systems they interact with.
1. Future research must expand beyond these conventional drivers of trustworthiness to include societal concerns, power dynamics, and long-term implications for human-AI interaction.
### 5.3. Measurement Problem in AI Trustworthiness
The critical analysis of measurement, verification, and validation approaches in AI trustworthiness research in Section 4.3 reveals gaps and methodological shortcomings to establish robust frameworks for evaluating trustworthiness of AI systems and build appropriate level of trust. More precisely, this necessitates: (i) conducting evaluations of both actual and perceived trustworthiness, i.e., establishing valid and reliable measurements, (ii) identifying methods to make valid comparisons between these two dimensions, and also (iii) to achieve alignment between them.
First, current research presents a significant gap concerning evaluation of trustworthiness. Prior studies have described the general challenge of aligning actual and perceived trustworthiness (mehrotrareview, 99, 86). They have also identified relevant components that constitute actual and perceived trustworthiness (schlicker_trustworthy_2025, 123). Additionally, previous work has emphasized the necessity for a communication process that conveys information about an AI system and its actual trustworthiness to shape and align trustworthiness perceptions (liao2022designing, 86). However, we still lack understanding of how the evaluation of actual trustworthiness and the successful alignment of this actual trustworthiness with perceived trustworthiness would manifest in practice, based on concrete measures and quantitative values for trustworthiness characteristics both for the AI system and user perceptions.
Second, when trustworthiness is measured using only questionnaires or reliance behaviour, there is still uncertainty about the trusteeâs actual trustworthiness (schlicker2025we, 122). Robust methodologies that incorporate objective measures of trustworthiness alongside subjective assessments are essential for developing a comprehensive understanding of trustworthiness in human-AI interaction. However, it is still complex to consider that actual and perceived trustworthiness combined together provides a clear overview of the AI systemâs overall trustworthiness, perhaps the $2+2=4$ analogy is not a fit here. We believe the issue is there is no single correct way of doing this, i.e., making individual characteristics commensurable. There is no ground truth to questions like what is the objective trustworthiness value for a given performance or explainability score of an AI system (Aida2023, 94).
The third problem differs from the first two in a fundamental way. Rather than dealing with how to combine or compare trustworthiness measures and individual traits, it addresses a more basic question: which specific characteristics should be evaluated when assessing trustworthiness in the first place? The challenge here isnât about creating compatible measurement scales. Instead, it involves establishing comparable sets of characteristics between two sidesâthe AI system itself and the users who perceive it. Critical questions arise from this challenge: Which performance metrics should be considered? What aspects of transparency matter? How should explainability be evaluated? What privacy measures and safety standards should factor into the assessment of AI system trustworthiness? These questions are all fundamentally connected to this obstacle.
The verification and validation issues are even more concerning, with nearly half of the papers failing to address these crucial aspects. For example, what is the source of cues that inform users about an AI systemâs trustworthiness and how can we verify or validate it? One perspective holds that these cues should originate directly from the systems and their creators. Some researchers advocate for defining good cues that technology developers should implement (liao2022designing, 86). These same scholars acknowledge that such indicators must be accurate and honest to be effective. When examining practical applications, they assume this honesty requirement is already met by creators and focus their analysis on additional necessary conditions. This assumption proves problematic in real-world scenarios. System providers cannot simply be expected to present entirely accurate cues about their productsâ limitations. Commercial entities operate with sales objectives as their primary driver. When users place excessive confidence in a product beyond whatâs justified, achieving proper confidence calibration would require the provider to communicate information that reduces trust. Yet commercial logic creates a strong disincentive for broadcasting unfavorable information; doing so would directly harm revenue and competitive positioning. It seems unlikely, therefore, that providers would intentionally engineer their systems to broadcast cues indicating mediocre or poor trustworthiness to potential users.
We propose the following measures for effective measurement, validation and verification:
1. Our analysis suggests a pressing need for AI system designers and developers to must explain their rationale for choosing particular trustworthiness attributes and specific metrics to measure those attributes. They also need to justify their methodology for prioritizing and combining these elements to generate a composite trustworthiness score. In doing so, they should be required to acknowledge their inherent perspectives and the criteria they apply when judging an AI systemâs trustworthiness.
1. Greater focus on trustworthiness assessment mechanisms within the trust development process would likely enhance future research efforts to clarify how perceived trustworthiness transforms into actual trust and subsequent trusting behaviors.
1. Those who provide the cues to the AI system users need to elaborate on and give reasons why certain characteristics of the AI system and trustworthiness cues warrant trust. Also, focus on understanding how cues are relevant for the users, how they detect and utlize it forms a key element for verification and validation of the trustworthiness assessment.
### 5.4. Values Gap in AI Trustworthiness
Our critical examination of the values in AI trustworthiness research reveals significant oversights and problematic assumptions that limit our understanding of how values materialize in practice. The analysis of intended values demonstrates a concerning preoccupation with surface-level technical attributes such as transparency and fairness, while giving insufficient attention to deeper systemic values such as social justice, environmental sustainability, and cultural preservation. This narrow focus reflects a persistent bias toward engineering solutions rather than addressing fundamental societal challenges (olteanu2019social, 108). The heavy emphasis on transparency risks becoming a performative gesture rather than a meaningful commitment to openness and accountability, thereby risking to entrench problematic capture of the institutions that are meant to support public values (10.1145/3531146.3533241, 51, 150).
The discussion of embodied values highlights weaknesses in current research. The literature often oversimplifies how values are embedded in technical systems, reducing complex social and ethical considerations to fit technical implementations. This simplification risks a false sense of progress while neglecting the deeper challenges of value implementation (stern2002eva, 131). Particularly concerning is the treatment of realized values, where the analysis lacks a systematic framework for understanding value divergence. The literature notes discrepancies between intended and realized values but inadequately theorizes these gaps, attributing them to superficial issues like âdesign flawsâ or âunintended consequencesâ without addressing structural barriers that hinder the realization of intended values.
The conclusion that a âmore holistic approachâ is required appears insufficient given the complexity of the challenges at hand. We need rigorous, dynamic frameworks reflecting the contested and evolving nature of values in AI systems. Such frameworks must move beyond static, universal notions of values to address how they are negotiated, implemented, and evaluated in real-world contexts, confronting structures of power. E.g., these power structures manifest in multiple dimensions that trustworthiness research must address: (i) corporate concentrations of AI development capacity that privilege certain stakeholdersâ definitions of âtrustworthiness,â (ii) institutional hierarchies within organizations that determine whose values and concerns shape AI systems, (iii) socio-economic disparities that affect who benefits from or bears risks of AI deployment, and (iv) geopolitical imbalances that enable dominant nations to set global AI governance norms; and professional authority structures that privilege technical expertise over lived experience.
We recommend the following measures to reduce the values gap in AI trustworthiness:
1. Implementing participatory design methodologies that meaningfully involve marginalized communities in defining trustworthiness criteria as informed by (harrington2019deconstructing, 57).
1. Developing evaluation frameworks that assess differential impacts across diverse populations rather than assuming universal benefits (carey2017glossary, 20).
1. Establishing transparency mechanisms that reveal how power influences AI system development decisions and provide mechanisms to contest them as showcased by (ehsan2021expanding, 38, 100).
1. Creating accountability structures that redistribute decision-making authority beyond technical experts (busuioc2021accountable, 18) and recognizing that trustworthiness is a contested concept shaped by those with the power to define it (hardin2002trust, 56).
### 5.5. Operationalizing Recommendations: Use Cases and Transformations
In the above four subsections, we propose general recommendations and measures for future directions from a sociotechnical perspective. Concretely, these recommendations can be deployed across multiple AI application domains. To make them more actionable for researchers working from technical perspectives, such as machine learning practitioners, we provide several illustrative examples across diverse applications: AI-powered clinical decision support systems and conversational agents for caregiver mental well-being support (AI in healthcare), AI-powered hiring/recruitment systems (AI in recruitment), criminal risk assessment algorithms (AI in criminal justice), and AI-powered adaptive learning platforms and automated assessment systems (AI in education). Table 2 demonstrates how current AI systems are defined, built, and evaluated in these use cases, and illustrates the transformative changes that our proposed recommendations and measures would bring to each domain.
Table 2. Comparison of current AI system practices and practices informed by our proposed recommendations: Illustrative use cases shows how our recommendations address limitations in current practices.
| AI-powered clinical decision support system | Evaluate overall accuracy metrics (sensitivity/specificity) on aggregated test datasets (hicks2022evaluation, 60); Measure trust through post-deployment user surveys with general practitioners (chen2022acceptance, 22); Apply fairness metrics (e.g., demographic parity, equalized odds) to measure performance disparities across protected attributes like race and gender (chen2023algorithmic, 23); Assess system performance using standardized benchmark datasets primarily from well-resourced healthcare settings (blagec2023benchmark, 15). | Participatory measurement design: Co-develop trustworthiness criteria with patients from underserved communities (e.g., rural populations, minorities with higher disease prevalence) to identify what âtrustworthy screeningâ means beyond accuracyâincluding accessibility, cultural sensitivity, and explanation comprehensibility; Differential impact evaluation: Stratify performance metrics across race, socioeconomic status, age, and geographic location to reveal disparities (e.g., higher false negative rates in darker skin tones due to training data bias); Power-aware transparency: Document and disclose how decisions about training data selection, feature prioritization, and deployment locations were made, revealing potential bias toward well-resourced hospitals; Redistributed evaluation authority: Include community health workers, patients, and advocacy groups in defining evaluation success criteria alongside ophthalmologists and data scientists; Context-sensitive trustworthiness: Recognize that trustworthiness criteria differârural clinic practitioners may prioritize offline functionality and low-bandwidth operation, while medical centers focus on integration with their existing systems. |
| --- | --- | --- |
| Conversational agents for caregiver mental well-being support | Maintain system safety mainly through content moderation filters (sarkar2023review, 120); Measure user engagement metrics (session duration, return rate, sentiment scores) (limpanopparat2024user, 88); Assess conversational quality using expert annotations and standardized mental health screening tools (dosovitsky2021psychometric, 34); Conduct one-time user satisfaction surveys post-interaction (limpanopparat2024user, 88); Focus on clinical efficacy benchmarks and harm prevention protocols (sarkar2023review, 120). | Longitudinal trust dynamics: Track how caregiver trust evolves across crisis moments (e.g., after receiving unhelpful advice during acute stress) and design explicit trust repair mechanisms, such as human handoff protocols when the system fails to provide adequate support; Perceived vs. actual trustworthiness: Distinguish between caregiversâ confidence in the system (perceived) and its actual clinical appropriatenessâmeasuring whether users over-rely on the agent; Institutional accountability frameworks: Establish clear governance structures defining who is responsible when the system provides harmful adviceâincluding regulatory compliance mechanisms, clinical oversight boards, and transparent corporate policies on data usage and care escalation; Appropriateness of trust calibration: Develop frameworks that help caregivers understand when to trust the AI (e.g., emotional validation, stress management techniques) versus when to seek human professionals (e.g., suicidal ideation, severe depression), with explicit system boundaries |
| AI-powered hiring/recruitment systems | Evaluate prediction accuracy by measuring how well AI selections match historical hiring decisions and subsequent employee performance (black2020ai, 14); Assess system efficiency in filtering large applicant pools (chen2023collaboration, 24); Apply minimal fairness audits primarily to meet legal compliance requirements (equal2023assessing, 27). | Participatory measurement design: Co-develop trustworthiness criteria with job seekers from underrepresented groups to identify what âfair screeningâ means beyond demographic parity; Power-aware transparency: Document and disclose whose definition of âqualified candidateâ is encoded in the systemârevealing whether criteria reflect genuine job requirements or historical biases of previous hiring managers, and how employer preferences versus candidate needs are balanced; Context-sensitive trustworthiness: Recognize that trustworthiness criteria differ by stakeholderâemployers may prioritize efficiency and cultural fit, while candidates prioritize fairness and transparency in rejection reasons, and marginalized communities may prioritize systems that challenge rather than perpetuate historical exclusion patterns. |
| Criminal risk assessment algorithms | Evaluate prediction accuracy by measuring how well risk scores correlate with actual re-arrest or re-offense rates (dressel2018accuracy, 35); Conduct post-hoc bias and fairness audits when systems face legal challenges or public scrutiny (angwin2022machine, 6); Measure system adoption rates and judicial satisfaction with decision support (stevenson2024algorithmic, 132). | Differential impact evaluation: Assess not just classification accuracy but downstream consequences across racial and socioeconomic groupsâincluding cumulative disadvantage from pretrial detention, access to rehabilitation programs, employment prospects post-release, and intergenerational community impacts of mass incarceration; Power-aware transparency: Document and disclose whose conception of âriskâ and âpublic safetyâ is embedded in the systemârevealing whether models perpetuate historical policing patterns that over-surveilled marginalized communities, how feature selection reflects prosecutorial versus defense perspectives, and which stakeholder interests (law enforcement efficiency vs. defendant rights) dominate design decisions; Redistributed evaluation authority: Include criminal justice reform advocates, affected community members, defense attorneys, and social workers in defining evaluation success criteria alongside prosecutors. |
| AI-powered adaptive learning platforms and automated assessment systems | Assess prediction accuracy of student performance and dropout risk models; Measure engagement metrics (time-on-task, interaction frequency, content completion) (tan2025artificial, 135); Conduct user satisfaction surveys with teachers and administrators (strielkowski2025ai, 133); Optimize for institutional metrics like pass rates and graduation rates (du2024personalized, 36). | Longitudinal trust dynamics: Track how student and teacher trust in the system evolves across academic failures or successesâdesigning trust repair mechanisms when the system provides ineffective learning paths or inaccurate assessments; Differential impact evaluation: Assess not just average learning gains but stratify outcomes across socioeconomic status, disability status, language background, and prior educational opportunityârevealing whether the system narrows or widens achievement gaps, and whether it channels certain populations toward vocational tracks while others receive enrichment; Power-aware transparency: Document and disclose whose pedagogical philosophy is embedded in the systemârevealing whether learning objectives reflect standardized testing priorities versus holistic education; Appropriateness of trust calibration: Develop frameworks helping students and teachers understand when to trust AI recommendations (e.g., practice problem selection, pacing suggestions) versus when to override them (e.g., when algorithms misinterpret struggle as inability rather than productive challenge, or when standardized paths ignore individual passions and strengths). |
Our proposed recommendations represent the first step rather than the end of achieving trustworthiness of AI systems in real-world applications. Realizing this vision requires sustained, collaborative efforts from researchers, practitioners, and stakeholders across multiple disciplines.
### 5.6. Limitations
First, our focus on AIES & FAccT conferences reflects a particular scholarly perspective. To address this and build a more comprehensive understanding of how trustworthiness is conceptualized across AI research landscape, we propose extending our analysis to major technical AI conferences such as NeurIPS, ICML, IJCAI, and AAAI. This expanded corpus would enable comparison between how trustworthiness is framed in technically-focused venues versus interdisciplinary ones, potentially revealing important differences in priorities, metrics, evaluation methods, and theoretical frameworks.
Second, our focus on academic discourse may not adequately represent how trustworthiness is understood by other key stakeholders, including industry practitioners, policymakers, and diverse user communities. These groups may prioritize different aspects of trustworthiness based on their contexts and concerns.
Finally, cultural and geographical limitations exist in our analysis, as both AIES & FAccT have historically featured stronger representation from Western institutions. This may result in overlooking important cultural variations in how trustworthiness is conceptualized across societies and knowledge traditions.
## 6. Conclusion
This study underscores the growing importance of trustworthiness in AI systems as a foundation for ethical and reliable technology. Through a review of AIES & FAccT conference proceedings, we analyzed how trustworthiness is conceptualized, measured, and validated. While progress has been made in defining attributes such as transparency, accountability, and robustness, significant gaps remain, particularly in addressing the sociotechnical nature of AI and integrating safety into trustworthiness discussions.
Our findings highlight the need for an interdisciplinary approach that combines technical precision with social and ethical considerations. Current research often prioritizes technical attributes, overlooking the complex interplay between AI systems and the broader social, cultural, and institutional contexts in which they operate. Addressing these gaps is crucial for the AIES & FAccT community to drive a paradigm shift, advancing holistic approaches to trustworthy AI that genuinely benefit society and promote responsible technological development.
Acknowledgements. This research was (partially) supported by the Dutch Research Council (NWO), under project numbers 024.004.022, NWA.1389.20.183, and KICH3.LTP.20.006, and the European Union under grant agreements No. 101070212 (FINDHR) and No. 101201510 (UNITE). Views and opinions expressed are those of the author(s) only and do not necessarily reflect those of their respective employers, funders and/or granting authorities.
## References
- (1) Leah Hope Ajmani et al. âA Systematic Review of Ethics Disclosures in Predictive Mental Health Researchâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT â23 Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 1311â1323 DOI: 10.1145/3593013.3594082
- (2) Khader I Alkhouri âThe Role of Artificial Intelligence in the Study of the Psychology of Religionâ In Religions 15.3 MDPI, 2024, pp. 290
- (3) Tim Alpherts, Sennay Ghebreab, Yen-Chia Hsu and Nanne Noord âPerceptive Visual Urban Analytics is Not (Yet) Suitable for Municipalitiesâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 1341â1354 DOI: 10.1145/3630106.3658976
- (4) McKane Andrus, Elena Spitzer, Jeffrey Brown and Alice Xiang âWhat We Canât Measure, We Canât Understand: Challenges to Demographic Data Procurement in the Pursuit of Fairnessâ In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT â21 Virtual Event, Canada: Association for Computing Machinery, 2021, pp. 249â260 DOI: 10.1145/3442188.3445888
- (5) McKane Andrus, Elena Spitzer, Jeffrey Brown and Alice Xiang âWhat We Canât Measure, We Canât Understand: Challenges to Demographic Data Procurement in the Pursuit of Fairnessâ In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency Virtual Event, Canada: Association for Computing Machinery, 2021, pp. 249â260 DOI: 10.1145/3442188.3445888
- (6) Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner âMachine biasâ In Ethics of data and analytics Auerbach Publications, 2022, pp. 254â264
- (7) Sule Anjomshoae, Amro Najjar, Davide Calvaresi and Kary FrĂ€mling âExplainable Agents and Robots: Results from a Systematic Literature Reviewâ In 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), Montreal, Canada, May 13â17, 2019, 2019, pp. 1078â1088 International Foundation for Autonomous AgentsMultiagent Systems
- (8) Hilary Arksey and Lisa OâMalley âScoping Studies: Towards a Methodological Frameworkâ In International journal of social research methodology 8.1 Taylor & Francis, 2005, pp. 19â32
- (9) Michael D Baer and Jason A Colquitt âWhy Do People Trust?: Moving toward a More Comprehensive Consideration of the Antecedents of Trustâ In The Routledge companion to trust Routledge, 2018, pp. 163â182
- (10) Andrew Balmer âA Sociological Conversation with ChatGPT about AI Ethics, Affect and Reflexivityâ In Sociology 57.5 SAGE Publications Sage UK: London, England, 2023, pp. 1249â1258
- (11) Victor R. Basili, Gianluigi Caldiera and H. Rombach âThe Goal Question Metric Approachâ In Encyclopedia of Software Engineering, 1994, pp. 528â532
- (12) Michaela Benk, Sophie Kerstan, Florian Wangenheim and Andrea Ferrario âTwenty-four Years of Empirical Research on Trust in AI: A Bibliometric Review of Trends, Overlooked Issues, and Future Directionsâ In AI & society Springer, 2024, pp. 1â24
- (13) Umang Bhatt et al. âExplainable Machine Learning in Deploymentâ In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency Barcelona, Spain: Association for Computing Machinery, 2020, pp. 648â657 DOI: 10.1145/3351095.3375624
- (14) J Stewart Black and Patrick Esch âAI-enabled recruiting: What is it and how should a manager use it?â In Business horizons 63.2 Elsevier, 2020, pp. 215â226
- (15) Kathrin Blagec, Jakob Kraiger, Wolfgang FrĂŒhwirt and Matthias Samwald âBenchmark datasets driving artificial intelligence development fail to capture the needs of medical professionalsâ In Journal of Biomedical Informatics 137 Elsevier, 2023, pp. 104274
- (16) Joshua Brand âExploring the Moral Value of Explainable Artificial Intelligence Through Public Service Postal Banksâ In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, AIES â23 MontrĂ©al, QC, Canada: Association for Computing Machinery, 2023, pp. 990â992 DOI: 10.1145/3600211.3604741
- (17) Miles Brundage et al. âToward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claimsâ In arXiv preprint arXiv:2004.07213, 2020
- (18) Madalina Busuioc âAccountable Artificial Intelligence: Holding Algorithms to Accountâ In Public administration review 81.5 Wiley Online Library, 2021, pp. 825â836
- (19) Cam Caldwell and Stephen E Clapham âOrganizational Trustworthiness: An International Perspectiveâ In Journal of business ethics 47 Springer, 2003, pp. 349â364
- (20) Gemma Carey and Brad Crammond âA glossary of Policy Frameworks: The Many Forms of âUniversalismâand Policy âTargetingââ In J Epidemiol Community Health 71.3 BMJ Publishing Group Ltd, 2017, pp. 303â307
- (21) Stephen Casper et al. âBlack-Box Access is Insufficient for Rigorous AI Auditsâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 2254â2272 DOI: 10.1145/3630106.3659037
- (22) Mingyang Chen et al. âAcceptance of clinical artificial intelligence among physicians and medical students: a systematic review with cross-sectional surveyâ In Frontiers in medicine 9 Frontiers Media SA, 2022, pp. 990604
- (23) Richard J Chen et al. âAlgorithmic fairness in artificial intelligence for medicine and healthcareâ In Nature biomedical engineering 7.6 Nature Publishing Group UK London, 2023, pp. 719â742
- (24) Zhisheng Chen âCollaboration among recruiters and artificial intelligence: removing human prejudices in employmentâ In Cognition, Technology & Work 25.1 Springer, 2023, pp. 135â149
- (25) Jon Chun and Katherine Elkins âThe Crisis of Artificial Intelligence: A New Digital Humanities Curriculum for Human-Centred AIâ In International Journal of Humanities and Arts Computing 17.2 Edinburgh University Press, 2023, pp. 147â167
- (26) Mark Coeckelbergh âThe Political Philosophy of AI: An Introductionâ John Wiley & Sons, 2022
- (27) Equal Employment Opportunity Commission âAssessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964.(2023)â L, 2023
- (28) Jamell Dacon âAre You Worthy of My Trust?: A Socioethical Perspective on the Impacts of Trustworthy AI Systems on the Environment and Human Societyâ In arXiv preprint arXiv:2309.09450, 2023
- (29) Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova âBERT: Pre-training of Deep Bidirectional Transformers for Language Understandingâ In arXiv preprint arXiv:1810.04805, 2018
- (30) Natalia DĂaz-RodrĂguez et al. âConnecting the Dots in Trustworthy Artificial Intelligence: From AI Principles, Ethics, and Key Requirements to Responsible AI Systems and Regulationâ In Information Fusion 99 Elsevier, 2023, pp. 101896
- (31) Graham Dietz and Deanne N Den Hartog âMeasuring Trust inside Organisationsâ In Personnel Review 35.5 Emerald Group Publishing Limited, 2006, pp. 557â588
- (32) GL DNV âISO 9001: 2015 Quality Management Systems-Requirementsâ In Guidance Document, 2015
- (33) Roel Dobbe, Thomas Krendl Gilbert and Yonatan Mintz âHard Choices in Artificial Intelligenceâ In Artificial Intelligence 300 Elsevier, 2021, pp. 103555
- (34) Gilly Dosovitsky, Erick Kim and Eduardo L Bunge âPsychometric properties of a chatbot version of the PHQ-9 with adults and older adultsâ In Frontiers in Digital Health 3 Frontiers Media SA, 2021, pp. 645805
- (35) Julia Dressel and Hany Farid âThe accuracy, fairness, and limits of predicting recidivismâ In Science advances 4.1 American Association for the Advancement of Science, 2018, pp. eaao5580
- (36) Eileen Du Plooy, Daleen Casteleijn and Denise Franzsen âPersonalized adaptive learning in higher education: A scoping review of key characteristics and impact on academic performance and engagementâ In Heliyon 10.21 Elsevier, 2024
- (37) Milena Dzhelyova, David I Perrett and Ines Jentzsch âTemporal Dynamics of Trustworthiness Perceptionâ In Brain research 1435 Elsevier, 2012, pp. 81â90
- (38) Upol Ehsan et al. âExpanding Explainability: Towards Social Transparency in AI Systemsâ In Proceedings of the 2021 CHI conference on human factors in computing systems, 2021, pp. 1â19
- (39) Severin Engelmann et al. âClear Sanctions, Vague Rewards: How Chinaâs Social Credit System Currently Defines âGoodâ and âBadâ Behaviorâ In Proceedings of the Conference on Fairness, Accountability, and Transparency Atlanta, GA, USA: Association for Computing Machinery, 2019, pp. 69â78 DOI: 10.1145/3287560.3287585
- (40) Severin Engelmann, Chiara Ullstein, Orestis Papakyriakopoulos and Jens Grossklags âWhat People Think AI Should Infer From Facesâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency Seoul, Republic of Korea: Association for Computing Machinery, 2022, pp. 128â141 DOI: 10.1145/3531146.3533080
- (41) European Commission âProposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Actsâ ESO-European Sources Online, 2021 URL: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A52021PC0206
- (42) European Union HLEG âEthics Guidelines for Trustworthy Artificial Intelligence (AI)â Accessed: 2024-01-11, 2019 URL: https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=60419
- (43) Executive Office of the President âPresidential Executive Order 13960â Accessed: 2024-01-11, 2020 URL: https://www.federalregister.gov/documents/2020/02/14/2020-02544/e-o-13960-maintaining-american-leadership-in-artificial-intelligence
- (44) Andrea Ferrario âA Trustworthiness-based Metaphysics of Artificial Intelligence Systemsâ In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 2025, pp. 1360â1370
- (45) Andrea Ferrario and Michele Loi âHow Explainability Contributes to Trust in AIâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT â22 Seoul, Republic of Korea: Association for Computing Machinery, 2022, pp. 1457â1466 DOI: 10.1145/3531146.3533202
- (46) Andrea Fiore âIs Deweyâs Aesthetics Critical? A Reflection on the Relationship Between Artificial Intelligence and Art from a Deweyan Perspectiveâ In Contemporary Pragmatism 21.4 Leiden, The Netherlands: Brill, 2024, pp. 381â398 DOI: 10.1163/18758185-bja10096
- (47) Luciano Floridi et al. âAI4PeopleâAn Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendationsâ In Minds and Machines 28 Springer, 2018, pp. 689â707
- (48) âFAT* â18: Proceedings of the Conference on Fairness, Accountability, and Transparencyâ New York, NY, USA: Association for Computing Machinery, 2018 URL: https://facctconference.org/2018/
- (49) David C. Funder âOn the Accuracy of Personality Judgment: A Realistic Approach.â In Psychological Review 102.4 American Psychological Association, 1995, pp. 652
- (50) âAIES â18: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Societyâ New York, NY, USA: Association for Computing Machinery, 2018 URL: https://dl.acm.org/doi/proceedings/10.1145/3278721
- (51) Ben Gansky and Sean McDonald âCounterFAccTual: How FAccT Undermines Its Organizing Principlesâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT â22 Seoul, Republic of Korea: Association for Computing Machinery, 2022, pp. 1982â1992 DOI: 10.1145/3531146.3533241
- (52) Rory Gillis, Johann Laux and Brent Mittelstadt âTrust and Trustworthiness in Artificial Intelligenceâ In Handbook on Public Policy and Artificial Intelligence Edward Elgar Publishing, 2024, pp. 181â193
- (53) Maarten Grootendorst âBERTopic: Neural Topic Modeling with a Class-based TF-IDF Procedureâ In arXiv preprint arXiv:2203.05794, 2022
- (54) Margaret Hagan âA Human-centered Design Approach to Access to Justice: Generating New Prototypes and Hypotheses for Interventions to Make Courts User-friendlyâ In ndiana Journal of Law and Social Equality 6 HeinOnline, 2018, pp. 199
- (55) Margaret Hagan âLegal Design as a Thing: A Theory of Change and a Set of Methods to Craft a Human-centered Legal Systemâ In Design Issues 36.3 MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info âŠ, 2020, pp. 3â15
- (56) Russell Hardin âTrust and Trustworthinessâ Russell Sage Foundation, 2002
- (57) Christina Harrington, Sheena Erete and Anne Marie Piper âDeconstructing Community-based Collaborative Design: Towards More Equitable Participatory Design Engagementsâ In Proceedings of the ACM on Human-Computer Interaction 3.CSCW ACM New York, NY, USA, 2019, pp. 1â25
- (58) D. Harrison McKnight, Vivek Choudhury and Charles Kacmar âThe impact of initial consumer trust on intentions to transact with a web site: a trust building modelâ In The Journal of Strategic Information Systems 11.3, 2002, pp. 297â323 DOI: https://doi.org/10.1016/S0963-8687(02)00020-3
- (59) Yugang He âArtificial Intelligence and Socioeconomic Forces: Transforming the Landscape of Religionâ In Humanities and Social Sciences Communications 11.1 Palgrave, 2024, pp. 1â10
- (60) Steven A Hicks et al. âOn evaluation metrics for medical applications of artificial intelligenceâ In Scientific reports 12.1 Nature Publishing Group UK London, 2022, pp. 5979
- (61) Lingxiao Huang, Julia Wei and Elisa Celis âTowards Just, Fair and Interpretable Methods for Judicial Subset Selectionâ In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society New York, NY, USA: Association for Computing Machinery, 2020, pp. 293â299 DOI: 10.1145/3375627.3375848
- (62) Nanna Inie, Stefania Druga, Peter Zukerman and Emily M. Bender âFrom âAIâ to Probabilistic Automation: How Does Anthropomorphization of Technical Systems Descriptions Influence Trust?â In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 2322â2347 DOI: 10.1145/3630106.3659040
- (63) ISO âTrustworthiness â Vocabularyâ Accessed 2023-11-20, 2022 URL: https://www.iso.org/obp/ui/##iso:std:iso-iec:ts:5723:ed-1:v1:en
- (64) Abigail Z Jacobs and Hanna Wallach âMeasurement and Fairnessâ In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2021, pp. 375â385
- (65) Alon Jacovi, Ana MarasoviÄ, Tim Miller and Yoav Goldberg âFormalizing Trust in Artificial Intelligence: Prerequisites, Causes and Goals of Human Trust in AIâ In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, 2021, pp. 624â635
- (66) Japan Expert Group on How AI Principles Should be Implemented âGovernance Guidelines for Implementation of AI Principlesâ Accessed: 2024-01-11, 2022 URL: https://www.meti.go.jp/shingikai/mono_info_service/ai_shakai_jisso/pdf/20220128_2.pdf
- (67) Jiun-Yin Jian, Ann M. Bisantz and Colin G. Drury âFoundations for an Empirically Determined Scale of Trust in Automated Systemsâ In International Journal of Cognitive Ergonomics 4.1 Routledge, 2000, pp. 53â71 DOI: 10.1207/S15327566IJCE0401âË04
- (68) Monika Kastner et al. âWhat is the Most Appropriate Knowledge Synthesis Method to Conduct a Review? Protocol for a Scoping Reviewâ In BMC medical research methodology 12 Springer, 2012, pp. 1â10
- (69) Markus Kattnig, Alessa Angerschmid, Thomas Reichel and Roman Kern âAssessing Trustworthy AI: Technical and Legal Perspectives of Fairness in AIâ In Computer Law & Security Review 55 Elsevier, 2024, pp. 106053
- (70) Davinder Kaur, Suleyman Uslu and Arjan Durresi âRequirements for Trustworthy Artificial IntelligenceâA Reviewâ In Advances in Networked-Based Information Systems: The 23rd International Conference on Network-Based Information Systems (NBiS-2020) 23, 2021, pp. 105â115 Springer
- (71) Christoph Kelp and Mona Simion âWhat is Trustworthiness?â In NoĂ»s 57.3 Wiley Online Library, 2023, pp. 667â683
- (72) Eugenia Kim, DeâAira Bryant, Deepak Srikanth and Ayanna Howard âAge Bias in Emotion Detection: An Analysis of Facial Emotion Recognition Performance on Young, Middle-Aged, and Older Adultsâ In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, AIES â21 Virtual Event, USA: Association for Computing Machinery, 2021, pp. 638â644 DOI: 10.1145/3461702.3462609
- (73) Sunnie S.. Kim et al. ââIâm Not Sure, ButâŠâ: Examining the Impact of Large Language Modelsâ Uncertainty Expression on User Reliance and Trustâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 822â835 DOI: 10.1145/3630106.3658941
- (74) Sunnie S.. Kim et al. âHumans, AI, and Context: Understanding End-Usersâ Trust in a Real-World Computer Vision Applicationâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 77â88 DOI: 10.1145/3593013.3593978
- (75) Sean Kinahan et al. âAchieving Reproducibility in EEG-Based Machine Learningâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 1464â1474 DOI: 10.1145/3630106.3658983
- (76) Bran Knowles, Jasmine Fledderjohann, John T. Richards and Kush R. Varshney âTrustworthy AI and the Logics of Intersectional Resistanceâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 172â182 DOI: 10.1145/3593013.3593986
- (77) Bran Knowles and John T. Richards âThe Sanction of Authority: Promoting Public Trust in AIâ In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency Virtual Event, Canada: Association for Computing Machinery, 2021, pp. 262â271 DOI: 10.1145/3442188.3445890
- (78) Isabel Kusche âPossible Harms of Artificial Intelligence and the EU AI act: Fundamental Rights and Riskâ In Journal of Risk Research Routledge, 2024, pp. 1â14 DOI: 10.1080/13669877.2024.2350720
- (79) Christian Lahusen, Martino Maggetti and Marija Slavkovik âTrust, Trustworthiness and AI Governanceâ In Scientific Reports 14.1 Nature Publishing Group UK London, 2024, pp. 20752
- (80) Vivian Lai et al. âTowards a Science of Human-AI Decision Making: An Overview of Design Space in Empirical Human-Subject Studiesâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 1369â1385 DOI: 10.1145/3593013.3594087
- (81) Himabindu Lakkaraju and Osbert Bastani ââHow do I fool you?â: Manipulating User Trust via Misleading Black Box Explanationsâ In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society New York, NY, USA: Association for Computing Machinery, 2020, pp. 79â85 DOI: 10.1145/3375627.3375833
- (82) Benjamin Laufer et al. âFour Years of FAccT: A Reflexive, Mixed-Methods Analysis of Research Contributions, Shortcomings, and Future Prospectsâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT â22, 2022, pp. 401â426
- (83) Christie Lawrence, Isaac Cui and Daniel Ho âThe Bureaucratic Challenge to AI Governance: An Empirical Assessment of Implementation at U.S. Federal Agenciesâ In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society MontrĂ©al, QC, Canada: Association for Computing Machinery, 2023, pp. 606â652 DOI: 10.1145/3600211.3604701
- (84) John D Lee and Katrina A See âTrust in Automation: Designing for Appropriate Relianceâ In Human Factors 46.1 SAGE Publications Sage UK: London, England, 2004, pp. 50â80
- (85) Bo Li et al. âTrustworthy AI: From Principles to Practicesâ In ACM Comput. Surv. 55.9 New York, NY, USA: Association for Computing Machinery, 2023 DOI: 10.1145/3555803
- (86) Q.Vera Liao and S. Sundar âDesigning for Responsible Trust in AI Systems: A Communication Perspectiveâ In 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT â22 Seoul, Republic of Korea: Association for Computing Machinery, 2022, pp. 1257â1268 DOI: 10.1145/3531146.3533182
- (87) Daryl Lim âAI & IP: Innovation & Creativity in an Age of Accelerated Changeâ In Akron Law Review 52 HeinOnline, 2018, pp. 813
- (88) Sucharat Limpanopparat, Erin Gibson and Andrew Harris âUser engagement, attitudes, and the effectiveness of chatbots as a mental health intervention: A systematic reviewâ In Computers in Human Behavior: Artificial Humans 2.2 Elsevier, 2024, pp. 100081
- (89) Yang Liu et al. âTrustworthy LLMs: A Survey and Guideline for Evaluating Large Language Modelsâ Alignmentâ In arXiv preprint arXiv:2308.05374, 2023
- (90) Michele Loi and Matthias Spielkamp âTowards Accountability in the Use of Artificial Intelligence for Public Administrationsâ In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society Virtual Event, USA: Association for Computing Machinery, 2021, pp. 757â766 DOI: 10.1145/3461702.3462631
- (91) Chiara Longoni, Andrey Fradkin, Luca Cian and Gordon Pennycook âNews from Generative Artificial Intelligence Is Believed Lessâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency Seoul, Republic of Korea: Association for Computing Machinery, 2022, pp. 97â106 DOI: 10.1145/3531146.3533077
- (92) Takuya Maeda and Anabel Quan-Haase âWhen Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Designâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 1068â1077 DOI: 10.1145/3630106.3658956
- (93) Maria Isabel Magaña and Katie Shilton âFrameworks, Methods and Shared Tasks: Connecting Participatory AI to Trustworthy AI Through a Systematic Review of Global Projectsâ In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 2025, pp. 2166â2179
- (94) Jakob Mainz, Lauritz Munch and Jens Christian Bjerring âTwo Reasons for Subjecting Medical AI Systems to Lower Standards than Humansâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 44â49 DOI: 10.1145/3593013.3593975
- (95) Bertram F. Malle âBeyond Fairness and Explanation: Foundations of Trustworthiness of Artificial Agentsâ In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society, AIES â22 Oxford, United Kingdom: Association for Computing Machinery, 2022, pp. 4 DOI: 10.1145/3514094.3539570
- (96) Lydia Manikonda, Christopher Fosco and Eric Gilbert âWhen Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Designâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 1068â1081 DOI: 10.1145/3630106.3658956
- (97) Arianna Manzini et al. âShould Users Trust Advanced AI Assistants? Justified Trust As a Function of Competence and Alignmentâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 1174â1186 URL: https://doi.org/10.1145/3630106.3658964
- (98) Roger C. Mayer, James H. Davis and F. Schoorman âAn Integrative Model of Organizational Trustâ In Academy of Management Review 20.3, 1995, pp. 709â734
- (99) Siddharth Mehrotra et al. âA Systematic Review on Fostering Appropriate Trust in Human-AI Interaction: Trends, Opportunities and Challengesâ In ACM J. Responsib. Comput. 1.4 New York, NY, USA: Association for Computing Machinery, 2024 DOI: 10.1145/3696449
- (100) Siddharth Mehrotra et al. ââEven explanations will not help in trusting [this] fundamentally biased systemâ: A Predictive Policing Case-Studyâ In Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization, UMAP â25 NY, USA: Association for Computing Machinery, 2025, pp. 51â62 DOI: 10.1145/3699682.3728343
- (101) Siddharth Mehrotra, Catholijn M. Jonker and Myrthe L. Tielman âMore Similar Values, More Trust? - the Effect of Value Similarity on Trust in Human-Agent Interactionâ In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, AIES â21 Virtual Event, USA: Association for Computing Machinery, 2021, pp. 777â783 DOI: 10.1145/3461702.3462576
- (102) Siddharth Mehrotra, Carolina Centeio Jorge, Catholijn M. Jonker and Myrthe L. Tielman âIntegrity-based Explanations for Fostering Appropriate Trust in AI Agentsâ In ACM Trans. Interact. Intell. Syst. 14.1 New York, NY, USA: Association for Computing Machinery, 2024 DOI: 10.1145/3610578
- (103) Sina Mohseni âToward Design and Evaluation Framework for Interpretable Machine Learning Systemsâ In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society Honolulu, HI, USA: Association for Computing Machinery, 2019, pp. 553â554 DOI: 10.1145/3306618.3314322
- (104) BĂĄlint MucsĂĄnyi et al. âTrustworthy Machine Learningâ In arXiv preprint arXiv:2310.08215, 2023
- (105) Luca Nannini, Agathe Balayn and Adam Leon Smith âExplainability in AI Policies: A Critical Review of Communications, Reports, Regulations, and Standards in the EU, US, and UKâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 1198â1212 DOI: 10.1145/3593013.3594074
- (106) National Institute of Standards and Technology âArtificial Intelligence Risk Management Framework (AI RMF 1.0)â, https://csrc.nist.gov/publications/detail/sp/1270/final, 2023
- (107) Jörg Noller âExtended Human Agency: Towards a Teleological Account of AIâ In Humanities and Social Sciences Communications 11.1 Palgrave, 2024, pp. 1â7
- (108) Alexandra Olteanu, Carlos Castillo, Fernando Diaz and Emre Kıcıman âSocial Data: Biases, Methodological pitfalls, and Ethical Boundariesâ In Frontiers in big data 2 Frontiers Media SA, 2019, pp. 13
- (109) Femi Osasona et al. âReviewing the Ethical Implications of AI in Decision Making Processesâ In International Journal of Management & Entrepreneurship Research 6.2, 2024, pp. 322â335
- (110) Matthew J Page et al. âThe PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviewsâ In BMJ 372 British Medical Journal Publishing Group, 2021
- (111) Cecilia Panigutti et al. âThe Role of Explainable AI in the Context of the AI Actâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency Chicago, IL, USA: Association for Computing Machinery, 2023, pp. 1139â1150 DOI: 10.1145/3593013.3594069
- (112) Eva Paraschou et al. âTies of Trust: a bowtie model to uncover trustor-trustee relationships in LLMsâ In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 2025, pp. 1715â1728
- (113) Saumya Pareek, Eduardo Velloso and Jorge Goncalves âTrust Development and Repair in AI-Assisted Decision-Making during Complementary Expertiseâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency Rio de Janeiro, Brazil: Association for Computing Machinery, 2024, pp. 546â561 DOI: 10.1145/3630106.3658924
- (114) Ibo Poel âEmbedding Values in Artificial Intelligence (AI) Systemsâ In Minds and Machines 30.3 Springer, 2020, pp. 385â409
- (115) Nineta Polemi, Isabel Praça, Kitty Kioskli and Adrien BĂ©cue âChallenges and Efforts in Managing AI Trustworthiness Risks: A State of Knowledgeâ In Frontiers in Big Data 7 Frontiers Media SA, 2024, pp. 1381163
- (116) Nicholas Proferes et al. âStudying Reddit: A Systematic Overview of Disciplines, Approaches, Methods, and Ethicsâ In Social Media+ Society 7.2 SAGE Publications Sage UK: London, England, 2021, pp. 20563051211019004
- (117) Soha Rawas âAI: The Future of Humanityâ In Discover Artificial Intelligence 4.1 Springer, 2024, pp. 25
- (118) Waddah Saeed and Christian Omlin âExplainable AI (XAI): A Systematic Meta-survey of Current Challenges and Future Opportunitiesâ In Knowledge-Based Systems 263 Elsevier, 2023, pp. 110273
- (119) Filippo Santoni de Sio and Giulio Mecacci âFour Responsibility Gaps with Artificial Intelligence: Why They Matter and How to Address Themâ In Philosophy & Technology 34.4 Springer, 2021, pp. 1057â1084
- (120) Surjodeep Sarkar et al. âA review of the explainability and safety of conversational agents for mental health to identify avenues for improvementâ In Frontiers in Artificial Intelligence 6 Frontiers Media SA, 2023, pp. 1229805
- (121) Nicolas Scharowski et al. âCertification Labels for Trustworthy AI: Insights From an Empirical Mixed-Method Studyâ In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT â23, 2023, pp. 248â260
- (122) Nadine Schlicker et al. âHow Do We Assess the Trustworthiness of AI? Introducing the Trustworthiness Assessment Model (TrAM)â In Computers in Human Behavior Elsevier, 2025, pp. 108671
- (123) Nadine Schlicker et al. âTrustworthy enough? Examining trustworthiness assessments of large language model-based medical agents.â In Technology, Mind, and Behavior 6.2, 2025, pp. 1â29 DOI: 10.1037/tmb0000164
- (124) Nadine Schlicker et al. âCalibrated Trust as a Result of Accurate Trustworthiness AssessmentâIntroducing the Trustworthiness Assessment Modelâ In PsyArXiv PPR:PPR547033, 2022
- (125) Anna Schmitz âTowards Formalizing and Assessing AI Fairnessâ In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society MontrĂ©al, QC, Canada: Association for Computing Machinery, 2023, pp. 999â1001 DOI: 10.1145/3600211.3604762
- (126) Jakob Schoeffer, Niklas Kuehl and Yvette Machowski ââThere Is Not Enough Informationâ: On the Effects of Explanations on Perceptions of Informational Fairness and Trustworthiness in Automated Decision-Makingâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT â22, 2022, pp. 1616â1628
- (127) Science Australia Department of Industry and Resources âAI Ethics Principlesâ Accessed: 2024-01-11, https://legalinstruments.oecd.org/en/instruments/oecd-legal0449, 2023
- (128) Krithi Shailya, Shreya Rajpal, Gokul S Krishnan and Balaraman Ravindran âLExT: Towards Evaluating Trustworthiness of Natural Language Explanationsâ In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 2025, pp. 1565â1587
- (129) Renee Shelby et al. âSociotechnical Harms of Algorithmic Systems: Scoping a Taxonomy for Harm Reductionâ In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society MontrĂ©al, QC, Canada: Association for Computing Machinery, 2023, pp. 723â741 DOI: 10.1145/3600211.3604673
- (130) Quinten Steenhuis âAI and Tools for Expanding Access to Justiceâ In SSRN 4876633, 2024
- (131) Joel M Stern, John S Shiely and Irwin Ross âThe EVA Challenge: Implementing Value-added Change in an Organizationâ John Wiley & Sons, 2002
- (132) Megan T Stevenson and Jennifer L Doleac âAlgorithmic risk assessment in the hands of humansâ In American Economic Journal: Economic Policy 16.4 American Economic Association 2014 Broadway, Suite 305, Nashville, TN 37203-2425, 2024, pp. 382â414
- (133) Wadim Strielkowski et al. âAI-driven adaptive learning for sustainable educational transformationâ In Sustainable Development 33.2 Wiley Online Library, 2025, pp. 1921â1947
- (134) Jinzhe Tan, Hannes Westermann and Karim Benyekhlef âChatGPT as an Artificial Lawyer?â In AI4AJ@ ICAIL, 2023
- (135) Le Ying Tan, Shiyu Hu, Darren J Yeo and Kang Hao Cheong âArtificial Intelligence-Enabled Adaptive Learning Platforms: A Reviewâ In Computers and Education: Artificial Intelligence Elsevier, 2025, pp. 100429
- (136) The Government of Canada âDirective on Automated Decision-Makingâ Accessed: 2024-01-11, https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592, 2023
- (137) The Government of Canada âResponsible Use of Artificial Intelligence Guiding Principlesâ Accessed: 2024-01-11, https://www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai.html, 2023
- (138) Scott Thiebes, Sebastian Lins and Ali Sunyaev âTrustworthy Artificial Intelligenceâ In Electronic Markets 31 Springer, 2021, pp. 447â464
- (139) Lauren Thornton, Bran Knowles and Gordon Blair âFifty Shades of Grey: In Praise of a Nuanced Approach Towards Trustworthy Designâ In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency Virtual Event, Canada: Association for Computing Machinery, 2021, pp. 64â76 DOI: 10.1145/3442188.3445871
- (140) Lauren Thornton, Bran Knowles and Gordon Blair âThe Alchemy of Trust: The Creative Act of Designing Trustworthy Socio-Technical Systemsâ In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT â22, 2022, pp. 1387â1398
- (141) Autumn Toney, Kathleen Curlee and Emelia Probasco âTrust Issues: Discrepancies in Trustworthy AI Keywords Use in Policy and Researchâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT â24, 2024, pp. 2222â2233
- (142) Ehsan Toreini et al. âThe Relationship between Trust in AI and Trustworthy Machine Learning Technologiesâ In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* â20, 2020, pp. 272â283
- (143) Tom R Tyler âTrust and Law Abidingness: A Proactive Model of Social Regulationâ In Boston University Law Review 81 HeinOnline, 2001, pp. 361
- (144) United Kingdom Information Commissionerâs Office âGuidance on AI and Data Protectionâ Accessed: 2024-01-11, 2023 URL: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/
- (145) Oleksandra Vereschak, Gilles Bailly and Baptiste Caramiaux âHow to Evaluate Trust in AI-assisted Decision Making? A Survey of Empirical Methodologiesâ In Proceedings of the ACM on Human-Computer Interaction 5.CSCW2 ACM New York, NY, USA, 2021, pp. 1â39
- (146) Ewart J Visser, Marvin Cohen, Amos Freedy and Raja Parasuraman âA Design Methodology for Trust Cue Calibration in Cognitive Agentsâ In Virtual, Augmented and Mixed Reality. Designing and Developing Virtual and Augmented Environments: 6th International Conference, VAMR 2014, Held as Part of HCI International 2014, Heraklion, Crete, Greece, June 22-27, 2014, Proceedings, Part I 6, 2014, pp. 251â262 Springer
- (147) T. Vought âOMB Memorandum M-21-06: Guidance for Regulation of Artificial Intelligence Applicationsâ Accessed: 2024-01-11, 2020 URL: https://www.whitehouse.gov/wp-content/uploads/2020/11/M-21-06.pdf
- (148) Jindong Wang et al. âTrustworthy Machine Learning: Robustness, Generalization, and Interpretabilityâ In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 5827â5828
- (149) Ruotong Wang, Ruijia Cheng, Denae Ford and Thomas Zimmermann âInvestigating and Designing for Trust in AI-powered Code Generation Toolsâ In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT â24, 2024, pp. 1475â1493
- (150) Meredith Whittaker âThe Steep Cost of Captureâ In Interactions 28.6 New York, NY, USA: Association for Computing Machinery, 2021, pp. 50â55 DOI: 10.1145/3488666
- (151) Tom Williams and Kerstin Sophie Haring âNo Justice, No Robots: From the Dispositions of Policing to an Abolitionist Roboticsâ In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society MontrĂ©al, QC, Canada: Association for Computing Machinery, 2023, pp. 566â575 DOI: 10.1145/3600211.3604663
- (152) John Zerilli, Umang Bhatt and Adrian Weller âHow Transparency Modulates Trust in Artificial Intelligenceâ In Patterns 3.4 Elsevier, 2022
- (153) Baobao Zhang and Allan Dafoe âU.S. Public Opinion on the Governance of Artificial Intelligenceâ In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society New York, NY, USA: Association for Computing Machinery, 2020, pp. 187â193 DOI: 10.1145/3375627.3375827
- (154) Jie Zhang and Zong-ming Zhang âEthics and Governance of Trustworthy Medical Artificial Intelligenceâ In BMC Medical Informatics and Decision Making 23.1 Springer, 2023, pp. 7
- (155) Miri Zilka, Holli Sargeant and Adrian Weller âTransparency, Governance and Regulation of Algorithmic Tools Deployed in the Criminal Justice System: a UK Case Studyâ In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society Oxford, United Kingdom: Association for Computing Machinery, 2022, pp. 880â889 DOI: 10.1145/3514094.3534200
## Appendix
## Appendix A List of included articles in the final corpus for analysis
1. [FAccT] Alpherts et al., Perceptive Visual Urban Analytics is Not (Yet) Suitable for Municipalities, Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Andrus et al., What We Canât Measure, We Canât Understand: Challenges to Demographic Data Procurement in the Pursuit of Fairness. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Bhatt et al., Explainable machine learning in deployment. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Casper et al., Black-Box Access is Insufficient for Rigorous AI Audits. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Engelmann et al., Clear Sanctions, Vague Rewards: How Chinaâs Social Credit System Currently Defines âGoodâ and âBadâ Behavior. Proceedings of the Conference on Fairness, Accountability, and Transparency, FAT* 2019.
1. [FAccT] Engelmann et al., What People Think AI Should Infer From Faces. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Ferrario et al., How Explainability Contributes to Trust in AI. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Ferrario, A. A Trustworthiness-based Metaphysics of Artificial Intelligence Systems. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Inie et al., From âAIâ to Probabilistic Automation: How Does Anthropomorphization of Technical Systems Descriptions Influence Trust? Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Jacovi et al., Formalizing Trust in Artificial Intelligence: Prerequisites, Causes and Goals of Human Trust in AI. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Kim et al., âIâm Not Sure, ButâŠâ: Examining the Impact of Large Language Modelsâ Uncertainty Expression on User Reliance and Trust. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Kim et al., Humans, AI, and Context: Understanding End-Users Trust in a Real-World Computer Vision Application. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Kinahan et al., Achieving Reproducibility in EEG-Based Machine Learning. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Knowles et al., Trustworthy AI and the Logics of Intersectional Resistance. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Knowles et al., The Sanction of Authority: Promoting Public Trust in AI. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Liao et al., Designing for Responsible Trust in AI Systems: A Communication Perspective. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Maeda et al., When Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Design. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Magaña et al., Frameworks, Methods and Shared Tasks: Connecting Participatory AI to Trustworthy AI Through a Systematic Review of Global Projects. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Manzini et al., Should Users Trust Advanced AI Assistants? Justified Trust As a Function of Competence and Alignment. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Nannini et al., Explainability in AI Policies: A Critical Review of Communications, Reports, Regulations, and Standards in the EU, US, and UK. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Panigutti et al., The Role of Explainable AI in the Context of the AI Act. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Paraschou et al., Ties of Trust: A Bowtie Model to Uncover Trustor-Trustee Relationships in LLMs. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Pareek et al., Trust Development and Repair in AI-Assisted Decision-Making during Complementary Expertise. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Scharowski et al., Certification Labels for Trustworthy AI: Insights From an Empirical Mixed-Method Study. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Schoeffer et al., There Is Not Enough Information: On the Effects of Explanations on Perceptions of Informational Fairness and Trustworthiness in Automated Decision-Making. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Shailya et al., LExT: Towards Evaluating Trustworthiness of Natural Language Explanations. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Thornton et al., Fifty Shades of Grey: In Praise of a Nuanced Approach Towards Trustworthy Design. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Thornton et al., The Alchemy of Trust: The Creative Act of Designing Trustworthy Socio-Technical Systems. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Toney et al., Trust Issues: Discrepancies in Trustworthy AI Keywords Use in Policy and Research. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Toreini et al., The Relationship between Trust in AI and Trustworthy Machine Learning Technologies. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency.
1. [FAccT] Wang et al., Investigating and Designing for Trust in AI-powered Code Generation Tools. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency
1. [AIES] Brand et al., Exploring the Moral Value of Explainable Artificial Intelligence Through Public Service Postal Banks. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Huang et al., Towards Just, Fair and Interpretable Methods for Judicial Subset Selection. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Kim et al., Age Bias in Emotion Detection: An Analysis of Facial Emotion Recognition Performance on Young, Middle-Aged, and Older Adults. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Lakkaraju et al., âHow do I fool you?â: Manipulating User Trust via Misleading Black Box Explanations. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Lawrence et al., The Bureaucratic Challenge to AI Governance: An Empirical Assessment of Implementation at U.S. Federal Agencies. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Loi et al., Towards Accountability in the Use of Artificial Intelligence for Public Administrations. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Mohseni et al., Toward Design and Evaluation Framework for Interpretable Machine Learning Systems. Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Schmitz et al., Towards Formalizing and Assessing AI Fairness. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Shelby et al., Sociotechnical Harms of Algorithmic Systems: Scoping a Taxonomy for Harm Reduction. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society
1. [AIES] Williams et al., No Justice, No Robots: From the Dispositions of Policing to an Abolitionist Robotics. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Zhang et al., U.S. Public Opinion on the Governance of Artificial Intelligence. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society.
1. [AIES] Zilka et al., Transparency, Governance and Regulation of Algorithmic Tools Deployed in the Criminal Justice System: A UK Case Study. Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society.
## Appendix B Research Ethics and Social Impact
### B.1. Ethical considerations statement
This paper is primarily a literature review and theoretical contribution, thus we did not engage in any human subjects research, systems development or deployment. Our work has been guided by considerable efforts of the AIES & FAccT community in the ethical ramifications of AI systems and their impact on human societies. This scoping review aims to contribute to a better understanding of AI trustworthiness, acknowledging the potential biases and limitations. We are committed to fostering responsible AI development and deployment, and this research is intended to support that goal.
### B.2. Adverse impact statement
With this work, we seek to spark conversations across disciplines about the social and technical inequalities being seen in the literature while studying AI trustworthiness. Rather than âcalling outâ specific articles, we hope this critique serves as a âcall inâ other scholars to join together and work collectively towards critically understanding the trustworthiness of AI systems.
### B.3. Social impact
One of the primary impacts of AIES & FAccT scholarship is the enhancement of public literacy regarding AI technologies. By critically examining the articles focusing on understanding trustworthiness of AI systems, we hope to have fostered a greater understanding of the complexities involved in trustworthy AI deployment, particularly concerning issues of comprehensive measurement frameworks, socio-technical integration and contribution to SDGs.
## Appendix C Reproducibility Checklist for JAIR
### All articles:
1. All claims investigated in this work are clearly stated. [yes]
1. Clear explanations are given how the work reported substantiates the claims. [yes]
1. Limitations or technical assumptions are stated clearly and explicitly. [yes]
1. Conceptual outlines and/or pseudo-code descriptions of the AI methods introduced in this work are provided, and important implementation details are discussed. [NA]
1. Motivation is provided for all design choices, including algorithms, implementation choices, parameters, data sets and experimental protocols beyond metrics. [yes]