The Intellectual Structure and Thematic Evolution of Trust in AI Research: A BERTopic Analysis

Authors’ abstract

As AI is increasingly used across social and organizational domains, trust in AI has emerged as an important research topic. This study analyzes 605 articles from 2016 to 2025 to examine the intellectual structure and thematic evolution of the field. Bibliometric and BERTopic analyses were used to identify transitions before and after generative AI (GenAI). The results show that AI trust research expanded rapidly after GenAI, with notable growth in publications and citations. International collaboration analysis reveals a hub-oriented structure in which the United Kingdom holds the most extensive collaborative ties and the United States plays a prominent bridging role across country groups, while Asian countries have grown more visible in recent years. Topic modeling identified 13 topics across six research areas: AI governance and trustworthiness, organizational and service contexts, XAI and trust, medical AI and trust, AI trust in education, and AI trust in media. Prior to GenAI, research was concentrated around governance, human-AI collaboration, and XAI. Since then, scholarly attention has broadened toward education, media, and the public sector. These findings suggest that AI trust research has expanded toward broader social application domains in the post-GenAI period. INDEX TERMS BERTopic modeling, bibliometric analysis, generative AI, trust in AI, trustworthy AI. I. INTRODUCTION As AI technology is increasingly applied across various social domains, including healthcare, finance, and public adminis-tration, concerns about risks such as algorithmic errors, bias, and privacy violations have also grown [1], [2]. In this con-text, trust in AI carries significant importance: at the technical level, it relates to ensuring the accuracy and transparency of AI systems; at the level of human-AI interaction, it functions as a psychological basis through which users accept uncer-tainty and adopt the technology; and at the societal level, it serves as an institutional foundation for the safe deploy-ment of AI through regulation and ethical standards [3], [4]. Accordingly, trust in AI has emerged as a central topic in AI The associate editor coordinating the review of this manuscript and approving it for publication was Majdi Mansouri. research, and the volume of related studies has continued to grow. The release of ChatGPT in late 2022 marked a turning point, as generative AI (GenAI) rapidly gained widespread use and brought new dimensions to AI trust research. While GenAI is capable of performing a wide range of tasks, includ-ing natural language generation, code writing, and image synthesis, it also introduces new risks such as hallucination, biased outputs, and privacy concerns [5], [6]. These char-acteristics have intensified discussions around the reliability and accountability of AI systems, and it has been suggested that the emergence of GenAI may have brought changes not only in the volume of AI trust research but also in the thematic scope and areas of scholarly interest. As the body of AI trust research has expanded, review studies aimed at mapping the research landscape and identi-fying knowledge structures have also been conducted. These studies fall broadly into two types. The first comprises works that provide comprehensive overviews of AI trust concepts and research trends, examining the components and con-ditions of trust formation with a focus on explainability, transparency, fairness, and accountability [7], [8], [9]. The second includes studies that analyze AI trust within spe-cific application domains such as healthcare, education, and autonomous systems, identifying domain-specific factors and contextual conditions that influence trust formation [10], [11], [12]. These studies have made important contributions to clarifying the conceptual foundations of AI trust and under-standing research trends across different fields. However, existing studies have tended to focus on par-ticular technical factors or application domains, or have emphasized bibliometric analysis of research productivity and collaboration structures. Despite the rapid growth of AI trust research following the emergence of GenAI, little is known about how this technological shift has affected the thematic structure and development of the field. There is therefore a need to examine how the major research topics in AI trust have changed before and after the emergence of GenAI. Against this background, this study analyzes 605 peer-reviewed articles on AI trust published between 2016 and 2025, combining bibliometric analysis with BERTopic-based topic modeling to investigate the intellectual structure and thematic evolution of the field. By comparing the periods before (2016∼2022) and after (2023∼2025) the emergence of GenAI, this study examines how the technological shift has influenced the thematic structure and research trends in AI trust scholarship. The findings are intended to provide a comprehensive account of how AI trust research has devel-oped and to offer implications for future research directions. II. RESEARCH BACKGROUND A. THE CHARACTERISTICS, AND SIGNIFICANCE OF TRUST IN AI Trust is generally defined as a psychological state in which one is willing to accept vulnerability to another’s actions, based on the expectation that the other party will act in one’s interest [13]. In the context of AI, trust refers to users’ willingness to accept the suggestions or decisions of a system, share tasks with it, and support its use [14]. Prior research has described trust in AI as a multidimensional concept com-prising ability, benevolence, and integrity, and more recently, predictability has also been identified as an important dimen-sion [13], [15]. It is also important to distinguish between trust and trustworthiness. Trust refers to the subjective psy-chological state of the user, whereas trustworthiness refers to the objective properties of the system, such as accuracy, safety, and ethical alignment [16]. As a result, users may distrust a system with high trustworthiness, while excessive trust may be placed in a system with low trustworthiness [17]. Trust in AI has several characteristics that distinguish it from interpersonal trust or trust in conventional technologies. First, trust in AI is not a fixed state but a dynamic process that evolves continuously through actual use experience, begin-ning with initial interactions [3], [18]. Second, trust in AI is multidimensional in nature, shaped by both technical indi-cators such as accuracy, safety, and robustness, and ethical indicators such as fairness, transparency, and accountabil-ity [19], [20]. Third, the way trust forms may vary across cultural contexts, with reported differences between cultures that prioritize transparency and explanation and those that place greater weight on system capability and social pres-ence [21], [22]. The significance of trust in AI is evident across tech-nical, human-AI relational, and institutional dimensions. At the technical level, trust functions as a psychological basis that allows users to engage with AI systems even with-out fully understanding their internal operations, given the non-transparent character of many AI systems [23], [24]. In response to this challenge, explainability and transparency have been emphasized as key means of supporting trust for-mation, with system designs that help users understand how decisions are made gaining increasing attention [4], [16]. At the level of human-AI relations, insufficient trust can lead to the underuse of capable systems, while excessive trust may result in uncritical reliance on flawed ones [19], [21]. Trust calibration, the alignment between a system’s actual performance and the user’s level of trust, is therefore consid-ered a prerequisite for safe and effective AI use [25]. As AI increasingly functions as a collaborative agent, human factors such as technical literacy, risk perception, and anthropomor-phism have also been recognized as important influences on trust [20], [26]. At the institutional level, regulatory frame-works, ethical standards, and data protection policies serve as the foundation for societal trust in AI. In high-stakes domains such as healthcare and public administration, institutional trust plays a particularly important role in addressing risks such as algorithmic bias, unclear accountability, and privacy violations [20], [21]. B. A REVIEW OF BIBLIOMETRIC STUDIES ON AI TRUST As AI technology has spread across various social and indus-trial domains, trust in AI systems has become an important research topic. In response to this growth, bibliometric anal-yses and systematic literature reviews aimed at mapping research trends and identifying knowledge structures have also been conducted [27]. For this study, a search of Scopus for bibliometric analyses and literature reviews on AI trust yielded a total of 48 articles. These studies were found to have been conducted predominantly in recent years, indicating that as AI trust research has grown in volume, the need to organize its trends in a structured way has also increased. The studies can be broadly divided into two types: those that address the concept and structure of AI trust, and those that analyze AI trust within specific application domains. Several studies have organized the conceptual structure and research trends of AI trust. These works examine the conditions for trust formation with a focus on explainability, transparency, fairness, and accountability. Studies providing overviews of the conceptual structure and research trends in AI trust have been reported [7], [28], and research discussing the importance of trust formation in the design and deploy-ment of AI systems has also been presented [29], [30]. Studies analyzing the role of trust in AI-based decision-making environments [8] and research examining the conditions for trustworthy AI in organizational and societal contexts [9], [31] have also been reported. These studies contribute to clarifying the key concepts and components of AI trust and identifying the factors that influence trust formation. Review studies analyzing AI trust within specific applica-tion domains have also been conducted. In healthcare, as the use of AI-based diagnostic systems and clinical decision support systems has expanded, studies examining trust for-mation factors in medical settings have been reported [10], [32], [33]. Research analyzing the effects of explainability and transparency on user trust [34], [35] and studies exam-ining AI trust issues in automated and autonomous system environments [11] have also been presented. In education, research analyzing trust issues in GenAI-based learning envi-ronments has been reported [12], and systematic literature reviews examining user trust formation across various service contexts have been conducted [36]. These existing studies have contributed to organizing con-ceptual discussions on AI trust and research trends, and to analyzing scholarly productivity and collaboration structures through bibliometric methods. However, they tend to focus on particular technical factors or application domains, and studies that systematically examine how the rapid growth of AI trust research following the emergence of GenAI has affected the thematic structure and development of the field remain limited. There is therefore a need to analyze how the major research topics in AI trust have changed before and after the emergence of GenAI. III. RESEARCH METHODS Bibliographic data were collected from the Scopus database on January 3, 2026. Scopus is one of the most widely used academic databases for bibliometric analysis, as it comprehensively indexes peer-reviewed journal articles and offers broad international coverage [27]. The search was conducted using the title search field with the query (‘‘artifi-cial intelligence’’ OR ‘‘AI’’ OR GenAI OR ChatGPT) AND ‘‘trust∗’’, limited to journal articles written in English pub-lished between 2016 and 2025. The starting year of 2016 was selected on the grounds that this period is widely regarded as a turning point at which advances in deep learning drew broad public and scholarly attention to AI capabilities and their societal implications, thereby stimulating research interest in AI trust [1], [37]. The initial search yielded 688 articles, which were then assessed for eligibility through a review of titles and abstracts. This review was conducted by the authors together with one professor and one AI specialist affiliated with a government-funded research institute. Articles were excluded if they did not directly address trust in AI as a primary research theme, for instance, if AI was mentioned only as a peripheral tool or if the concept of trust appeared only incidentally. Where reviewers disagreed on eligibility, a final decision was reached through discussion and consen-sus. Following this process, 83 articles were excluded, and 605 articles were retained for final analysis. The analysis period was divided into two sub-periods based on the release of ChatGPT in late 2022: before the emergence of GenAI (2016 to 2022) and after (2023 to 2025). This division was intended to allow a comparative examination of how the spread of GenAI may have influenced the thematic structure of AI trust research. Bibliometric analysis quantitatively examines scholarly lit-erature to map the knowledge structure and developmental trends of a research field [27]. In this study, it was applied to annual publication and citation trends, productivity indicators for core journals and countries, and international collabora-tion networks. Prior to analysis, the raw data exported from Scopus underwent preprocessing to improve data quality. This included standardizing country names, removing dupli-cate records, and verifying author affiliation information to ensure consistency across the dataset. Productivity indicators included total publication count, total citation count, Citation per paper (CPP), h-index, g-index, and m-index [38], [39]. CPP represents the average number of citations per article and reflects citation efficiency adjusted for publication volume. The h-index simultaneously captures publication productivity and citation impact, the g-index is more sensitive to the influence of highly cited articles, and the m-index reflects the rate of growth in scholarly influence by accounting for years of activity. To identify core journals and countries, an IQR-based robust z-score criterion was applied. Unlike conventional z-scores, which are sensitive to extreme values, the robust z-score is based on the median and IQR, reducing the distorting effect of skewed citation distributions. Journals and countries that fell within the top 20 rankings in both publication count and citation count under this criterion were considered core. Growth rates between the pre- and post-GenAI periods were calculated using a log-difference approach, specifically ln(post-period + 1) minus ln(pre-period + 1), and were combined with total publication or citation counts to construct portfolio matrices. Each matrix was divided into four quadrants using the mean values of each axis as boundaries, distinguishing Leading (high volume, high growth), Established (high volume, low growth), Rising (low volume, high growth), and Peripheral (low volume, low growth). International collaboration was examined through co-authorship network analysis. The network included only countries appearing at least twice and connected through at least two co-authored publications, so as to focus on struc-turally stable collaborative ties. Community detection was performed using the Louvain clustering algorithm proposed by Blondel et al. [40], which identifies communities by iteratively partitioning the network to maximize modular-ity. This algorithm is widely used in scholarly collaboration network analysis due to its computational efficiency and reliable detection of community structures in large networks. To assess the structural position of each country within the network, five centrality measures were calculated: degree centrality (number of direct connections), weighted degree centrality (strength of connections weighted by collaboration frequency), betweenness centrality (bridging role in informa-tion flow across the network), closeness centrality (average proximity to all other nodes), and eigenvector centrality (degree of connection to other influential nodes). Topic modeling was conducted using BERTopic [41], which embeds documents via a pretrained BERT-based language model, reduces dimensionality through UMAP, and identifies topics through density-based clustering using HDBSCAN. Topic keywords are extracted using c-TF-IDF, reflecting how distinctively a term is used within a topic relative to others. Compared to LDA, BERTopic more pre-cisely captures semantic similarity among documents and does not require the number of topics to be specified in advance, making it well suited for exploratory research [42]. Separate models were fitted for the two sub-periods, and each topic was labeled by the researchers based on top keywords and representative documents, with consistency between the two serving as the primary criterion for topic naming. IV. RESULTS A. BIBLIOMETRIC ANALYSIS RESULTS

Paper

Authors

M.-K. Kim, J. Hyun, J. Lee

Publication date

2026

Read the original paper

Transcript

A transcript has not been published for this episode yet.

Licensing

Source license

Creative Commons Attribution 4.0 International

Recording license

CC-BY-4.0

The authors and publisher do not sponsor or endorse this recording.