The ethics of AI and its social implications have attracted serious attention, and many policy frameworks and guidelines have been developed by various organizations (Zeng et al., 2018; Franzke, 2022). Additionally, each framework contains different AI ethics principles, which reflect unique perspectives on current and future AI strategies. The AI ethical guidelines serve to indicate to legislators that internal self-governance in science and industry is sufficient and that specific laws are not needed to mitigate potential technological risks and eliminate abuse scenarios (Hagendorff, 2020). It must not be neglected that ethical principles for developing and using AI systems are proposed and contextualized in countless AI ethics guidelines, which prescribe ethical direction to AI systems’ developers, users, policymakers, and other stakeholders who seek to maximize the potential benefits, minimizing at the same time the potential harms of systems’ operations (Attard-Frost et al., 2023). Therefore, we consider it important to see the convergence of these principles in order to establish a set of minimum requirements regarding the trustworthiness of AI systems. We approach the aforementioned issue by comparing various systematic or semi-systematic reviews that have been conducted in the past on the ethical principles that should govern AI systems.

We conducted a meta-analysis of selected systematic literature reviews (SLRs) to synthesize existing insights into the ethical principles underlying Trustworthy AI. One of the first SLRs was that of Floridi et al. (2018), which is actually a semi-SLR. They assessed six documents with a chronological range from 2017 to 2018, which yielded 47 principles and derived five high-level principles (Beneficence, Non-Maleficence, Autonomy, Justice, Explicability). The principle of Explicability was added by authors as a need to incorporate both intelligibility and accountability. Zeng et al. (2018) collected 27 proposals of AI principles from societal entities like Academia, Non-profits and Non-Governmental Organizations, Governments, and Industry. Thinking about the issue from a semantic perspective, they posed a set of manually chosen core terms keywords, setting the core terms of Accountability, Privacy, Fairness, Humanity, Collaboration, Share, Transparency, Security, Safety, AGI/ASI (Artificial General/Super Intelligence). Jobin et al. (2019) conducted a systematic scoping review of the existing corpus of documents containing soft-law or non-legal norms issued by organizations. In particular, they identified 84 eligible, non-duplicate documents containing ethical principles for AI, revealing a global convergence emerging around five ethical principles (Transparency, Justice & Fairness, Non-maleficence, Responsibility, and Privacy).

In the current decade, Fjeld et al. (2020) analyzed the content of 36 prominent AI principles documents to identify trends and essential components in the discussion on the future of AI technologies. They established 47 principles governing the AI, which can be categorized into eight themes: Accountability, Privacy, Fairness & Nondiscrimination, Safety & Security, Transparency & Explainability, Human Control of Technology, Professional Responsibility, Promotion of Human Values. They also point out that the aforementioned themes serve as fundamental requirements, expressing the importance of these conceptual and ethical principles. Hagendorff (2020) conducted a semi-systematic literature review that compares 22 guidelines, finding that principles of Accountability, Fairness, and Privacy appear altogether in about 80% of all guidelines and identifying 22 ethical principles in total. He also notes that these three core aspects constitute of minimum requirements for building and using AI systems ethically. Franzke (2022) conducted research on a total of 70 AI ethics guidelines. The analysis concluded that the most dominant principles are Transparency, Privacy, and Accountability. She notes that AI ethics guidelines largely ignore the crucial question of how ethical principles can be transposed onto the usage of technology. Furthermore, Khan et al. (2022) presented a systematic literature review revealing a global convergence around 22 ethical principles. Through their approach, Transparency, Privacy, Accountability, and Fairness are identified as the most common AI ethics principles. They highlight the existence of significant practical challenges involved in implementing the guidelines in real-world conditions, mentioning the lack of tools or frameworks that bridge the gap between principles and practice. Attard-Frost et al. (2023) concentrate on four a priori principles, which they have designated F.A.S.T., examining 47 AI ethics guidelines. These principles are Fairness, Accountability, Sustainability, and Transparency. Corrêa et al. (2023) took into account 200 AI guidelines, identifying 17 principles prevalent in the respective policies and guidelines. From those principles, the top five, based on citation index, were found to be similar to the ones identified by Jobin et al. (2019) and Fjeld et al. (2020). Laine et al. (2024) conducted a systematic literature review to understand ethical principles and stakeholders in ethics-based AI auditing. From the sample of 110 studies, they conducted backward citation chaining. Finally, they assessed 93 articles on ethics-based AI auditing. The results were the grouping of 54 terms related to ethics into eight principles: Justice & Fairness, Transparency, Non-Maleficence, Responsibility, Privacy, Trust, Beneficence, Freedom & Autonomy. The authors enriched the results of the systematic literature review by adding three additional principles, namely Sustainability, Dignity, and Solidarity, in order to enhance comprehensiveness, even though none of the studies explicitly mentioned them. Table 1 presents the results of the systematic literature reviews, while Figure 1 presents the outcomes of the meta-analysis.

To explore the ethical principles underlying Trustworthy AI, we conducted a scoping review of the existing literature, with the aim of identifying systematic and semi-systematic reviews that addressed these principles in a broad, cross-disciplinary context. Our primary objective was to synthesize ethical principles that transcend specific application domains, providing a comprehensive foundation for further research.

We began by conducting a comprehensive review of the available literature on AI ethics, focusing on studies that examined ethical principles at a general level. However, we observed that many results focused on domain-specific applications, such as education and healthcare, rather than addressing ethical principles at a general level. To ensure the scope remained aligned with our objectives, we prioritized studies with a cross-disciplinary focus, excluding those heavily tied to specific domains. Building upon this foundation, we employed a targeted search strategy to complement our initial findings.

During this process, we noted a significant alignment between our findings and those presented in the meta-review by Ziouvelou et al. (2024), likely due not only to the relatively short time frame between the two studies but also to the overlap in key research questions addressed in both bodies of work. Their work served as a valuable reference point, prompting a re-examination and extension of their dataset. This involved applying updated inclusion and exclusion criteria to ensure the relevance and currency of our analysis. Specifically, we excluded one empirical study from their original dataset, as it did not meet the criteria for systematic or semi-systematic reviews. In addition, we incorporated a new review published in 2024, which expanded the temporal coverage of the analysis. These steps ensured our review incorporated the most up-to-date and relevant literature.

The scoping review provided a body of papers, which we used to conduct a meta-analysis that offered a framework for identifying and synthesizing ethical principles across diverse contexts. The insights derived from this review established the conceptual foundations upon which our algorithmic approach was developed, ensuring that it aligns with the requirements of Trustworthy AI. In essence, the methodological approach adopted enables the integration of algorithmic methods, PageRank, and TrustRank within this theoretical framework.

Building upon the ethical foundation derived from the scoping review and meta-analysis, we next outline the methodological design of the algorithmic framework that forms the core contribution of this paper. The rationale for this step is to complement the subjectivity inherent in existing self-assessment techniques with algorithmic criteria that allow a more systematic and reproducible evaluation of AI systems. In this methodological design, we operationalize the seven requirements of Trustworthy AI, as defined in the ALTAI framework, into the structure of a directed graph. Each requirement is decomposed into its constituent aspects, which are then associated with specific components of the AI system under study. In this way, the abstract ethical dimensions become concrete nodes within a network, allowing their interdependencies and reliance relationships to be represented explicitly.

To analyze this structure, we employ two established Link Analysis algorithms, namely PageRank and TrustRank. PageRank enables the assessment of importance by considering the collective references among nodes, while TrustRank extends this process by propagating trust scores from a set of predefined, trusted nodes. Applying these algorithms to the graph offers an exploratory means to investigate how trustworthiness propagates within an AI system and which components emerge as critical or vulnerable.

Finally, two complementary perspectives are considered: a top-down approach, which begins from high-level requirements and propagates trust to system components, and a bottom-up approach, which begins from components and aggregates trust scores upwards toward overarching requirements. Together, these perspectives provide a methodological basis for examining both the granular and systemic dimensions of AI trustworthiness. This algorithmic design does not aim to replace qualitative assessments or human oversight but to provide a structured, less subjective method for quantifying relationships within the system. Section 4 elaborates on the implementation of this framework, illustrating its application through hypothetical case scenarios.

The concepts of trustworthiness and trust are multifaceted, and they can be understood in various ways, depending on one’s perspective. In the context of AI, it is the idea of a framework ensuring a system is trustworthy, based on evidence of its stated requirements. It ensures that the expectations of users and stakeholders are met in a verifiable way (Kaur et al., 2022; International Organization for Standardization, 2020). The conceptual basis of trustworthiness lies in the concept of trust. Kaur et al. (2022) note that sociologists see trust as relational, psychologists as cognitive, and economists as calculative. Given the subtlety of this distinction and the fact that it can sometimes be overlooked in more technical or functionally driven approaches, it is essential to briefly examine the various dimensions of trust and trustworthiness.

Returning to the issue of definitions, economists tend to reduce trust to calculative expectations or institutional guarantees (Williamson, 1993), framing it as a form of rational risk-taking. Psychologists, meanwhile, focus on internal cognitive models of the trustor and perceived attributes of the trustee (Rousseau et al., 1998). Sociologists may treat trust as a function of social embedding or institutional structures (Rousseau et al., 1998). In the management sciences, trust is defined behaviorally as “the willingness of a party to be vulnerable” in relationships involving risk and uncertainty (Mayer et al., 1995). In the context of informatics and artificial intelligence, trust is often understood as the trustor’s willingness to rely on a system’s ability to perform specific actions or provide services within a defined context (Grandison and Sloman, 2000). This reliance is shaped by beliefs about the system’s competence, integrity, benevolence, and predictability (Siau and Wang, 2018), and reflects a readiness to accept vulnerability in technologically mediated interactions. Despite their disciplinary differences, these definitions converge on the idea that trust entails accepting vulnerability, supported by perceptions of competence, integrity, and reliability.

While these definitions are contextually appropriate, they often rely, implicitly or explicitly, on a philosophical core. Both phenomena are associated with a certain degree of expectation, vulnerability, and moral hope. These concepts are examined in greater detail in philosophical discourse. The Oxford Dictionary of Philosophy defines trust as “the attitude of expecting good performance from another party” (Blackburn, 2016), tying it closely to values like loyalty, truthfulness, and promise-keeping. Philosophically, trust can be understood as the confidence one entity has in another that the latter will behave as expected. It is necessary to point out that trust is an attitude that we have towards entities in which there is the hope that they will be trustworthy, where trustworthiness is a property and not an attitude (McLeod et al., 2023). Trust involves a form of reliance, but it goes beyond simple dependence, including some extra factor (Hawley, 2014; Durán and Pozzi, 2025; Simion and Kelp, 2023). This extra factor generally concerns why the trustor (i.e., the one who trusts) would rely on the trustee (i.e., the one who is trusted) to be willing to carry out the actions for which they have been trusted.

The philosophical definition is especially valuable because it highlights the normative dimension of trust. Trust is not simply a pragmatic strategy; it is a moral commitment. To trust someone or something is not only to rely on them but also to believe that they ought to act in a trustworthy manner. This normative link is what makes trust fragile, and its violation morally significant. While definitions of trust vary across disciplines, the philosophical approach offers a rich, unified, and normative framework that is either embedded within or provides the scaffolding for those other understandings. Far from being merely abstract, the philosophical conception of trust is indispensable for coherent interdisciplinary work.

Crucially, this moral and normative dimension also underpins trust in artificial intelligence and informatics. The notion of trust in AI encompasses more than functional performance or reliability; it implicitly assumes that systems will behave in ways aligned with human expectations and ethical standards. The philosophical account of trust thus provides the conceptual foundation to understand trust in AI not only as rational reliance but also as a form of moral engagement. In this context, reliance refers to the trustor’s dependence on a system’s ability to consistently fulfill its expected functions within a specific setting. Importantly, such reliance can occur not only between a human trustor and a technological system, but also among components within the system itself, where one component or ethical requirement depends on another to operate in a trustworthy manner and align with broader system goals. However, unlike traditional software, many AI systems, particularly those based on machine learning, derive their algorithms directly from data rather than explicit human instructions. This reliance on data-driven learning underscores the need for safeguards to ensure that these models are fit for purpose, ethically sound, and free from unintended biases, as their performance is fundamentally shaped by the quality and representativeness of the training data. Recognizing this, we propose an algorithmic approach to quantify these reliance relationships, aiming to interpret this extra factor within the qualitative context of established ethical frameworks, such as the ALTAI, discussed in the literature (see Figure 2).

The above definitions contribute to a theoretical understanding of the concept, but they lack clear and actionable guidance for the practical implementation of trustworthiness, particularly in the context of AI. In 1999, Lawrence Page, describing the PageRank algorithm, which is a method for objectively and mechanically rating websites, noted that PageRank could help a user decide if a site is trustworthy or not (Page, 1999). These words give us the impetus to take a deeper look at the present algorithm and its extension, called TrustRank, proposing both algorithms as a solution to bridging the gap between theoretical and practical approaches to Trustworthy AI.

PageRank’s main focus is to measure the importance of a webpage through the analysis of its inbound links. The core idea is that a page is important if many other important pages link to it (Brin and Page, 1998; Niu et al., 2018). PageRank plays a central role in search engines as it reflects not only the popularity of a webpage but also serves as a rough approximation of its reliability and quality. This is based on the principle that pages linked to by many others are generally regarded as reliable and valuable sources of information, as the linking pages themselves likely possess some level of authority and credibility. However, while this approach captures a form of collective trust, it may not always fully account for nuanced dimensions of trustworthiness, such as the accuracy, intent, or ethical alignment of the content. PageRank’s thesis is that a webpage is important if it is pointed to by other important pages (Langville and Meyer, 2006; Sharma et al., 2020). Furthermore, PageRank does not define or consider trust explicitly. It assumes that links themselves indicate a form of confidence from one page to another, with no distinction between trustworthy or untrustworthy links, but taking into account the popularity importance. The algorithm computes the probability that a random web surfer lands on a particular page by following links. Pages with more inbound links from important or highly linked pages are given a higher PageRank score (Brin and Page, 1998). Also, it requires a complete analysis of the link structure of the web or graph. This involves indexing the web to understand how pages are linked to each other. Of course, the algorithm is not without its drawbacks, the most important being that it fails to differentiate between trustworthy and spammy pages. Spam can rank highly if it receives enough inbound links, since the algorithm’s approach does not incorporate any knowledge about the quality of a site (Gyöngyi et al., 2004), nor does it explicitly penalize badness.

The challenge of defining trustworthiness in practical contexts brings the TrustRank algorithm to the forefront as a relevant methodological approach. Gyöngyi et al. (2004) developed the TrustRank, shifting the focus towards trustiness and potentially to trustworthiness. The intuition behind TrustRank is that a page with high PageRank, but without a relationship with any of the trusted pages, is suspicious (Giménez-García et al., 2016) and, by extension, less trustworthy. The algorithmic approach begins by assuming that trusted sites, chosen by humans, are unlikely to link to untrusted or spammy sites. So, the selection of seed nodes by humans structures the definition of trust. TrustRank begins with a set of manually selected seed pages that are known to be trustworthy. It then propagates trust scores from these seed pages to other linked pages, setting that the pages closer to the seed pages receive higher trust scores. The limitations of this approach are related to the subjectivity introduced into the model by the human factor. Just as with email spam, determining if a page or group of pages is spam is subjective. However, we remain skeptical that a fully autonomous mechanism for evaluating the trustworthiness of another system can exist independently of human oversight, given the inherently anthropocentric nature of trust and the epistemic challenges involved.

After describing the algorithms, it is equally essential to consider the algorithmic process of these methods. In Algorithm 1, we describe the stages of PageRank, with inputs being the set of pages (P), the damping factor (α), and the convergence threshold (ϵ). The damping factor prevents the algorithm from getting stuck in infinite loops. The intuition behind this factor is that a higher value means more weight to the link structure, while a lower value gives more weight to random jumps. Also, the factor is typically set to 0.85 (Brin and Page, 1998), meaning there is an 85% chance the user will continue clicking on links and a 15% chance they will stop and start over from a random page, ensuring that the PageRank values converge. The convergence threshold determines the level of accuracy required, with a smaller threshold meaning more precise calculations, but longer computational time.

TrustRank, as an extension of PageRank, generally has a similar structure. Noteworthy is that the intuition behind the decay factor (α) indicates that a higher value of the factor gives more weight to the link structure, while a lower value gives more weight to the initial trust assigned to seed pages. Furthermore, Algorithm 2 illustrates the phases of TrustRank.

Before assessing the trustworthiness of an AI system, it is essential to understand its architecture and components. The components of an AI system may vary depending on the organization that develops it, as well as the nature of the application and its intended use. Although there are shared conceptual foundations across different AI systems, it is important to note that each organization delineates the components and structures of the software it aims to develop in distinct ways. The proposed algorithmic approach assumes that the reliance relationships between the software components and the trustworthiness requirements of the relevant framework, in this case, the ALTAI, have already been defined. We propose applying Link Analysis algorithms to assess the trustworthiness of AI system components. By assessing these components, the overall trustworthiness of the AI system, concerning specific requirements, could be determined. The proposed methodological approach enables the investigation of the propagation of trustworthiness within the system and the dependency relationships between system components in terms of trustworthiness. As the conceptual foundation for our approach, we adopted the ALTAI schema, which represents a synthesis of the meta-analysis presented in Section 2, as it closely aligns with the content and findings of that analysis. Figure 3 shows the Trustworthy AI requirements and aspects of each, as described by the AI HLEG (European Commission, Directorate-General for Communications Networks, Content and Technology, 2020; High-Level Expert Group on Artificial Intelligence, 2019). These requirements and aspects serve as part of the elements of the graph-based algorithmic approach presented in what follows.

In the previously discussed examples, the outcomes of the PageRank and TrustRank algorithms were obtained for the graphs presented in Figures 6, 10. Most of the time, a clear distinction is evident between the scores obtained by the two algorithms. However, there are instances where the scores exhibit a high degree of similarity. The cases can be included in the following general conditions:

If the scores of a component satisfy condition (a), then it can be classified as critical, central and not well-connected to trusted nodes. These components need to be monitored to verify the trustworthiness of their operations. In the case where a component satisfies condition (b), then it can be classified as less critical but trustworthy. It may not require immediate monitoring but it could serve as a standard for trust. Condition (c) is characterized by a dual status. The two scores can be similar, taking either low or high values. If the two scores of a component are both similar and low, this component can be regarded as less critical and weakly connected to trusted nodes. In such a case, it is appropriate to try to improve the trustworthiness of the component and to reflect on whether it can be omitted from the system or not. Conversely, in cases where the scores are both similar and high, the corresponding component may be regarded as critical and trustworthy. In the previous case, which is the optimal one among those described, it is sufficient to supervise the component and endeavor to maintain its trustworthiness level, as it seems to be of central importance.

It is essential to mention that the choice of the direction of the connections within the graph representing the AI system significantly affects the interpretation of the results. As previously noted, in the top-down approach, entities at lower levels (i.e., system components) are expected to exhibit higher scores. This enables a clearer assessment of the trustworthiness of components within the AI system, as it provides evidence of the propagation of trustworthiness from higher-level entities to their corresponding lower-level counterparts. Antithetically, the bottom-up approach exhibits a tendency for higher-level entities, such as aspects of AI trustworthiness requirements, to achieve higher scores. This allows us to gather evidence for the trustworthiness of the system concerning overarching aspects, such as Social Impact, linked to requirements like Societal and Environmental Wellbeing. In conclusion, the proposed algorithmic approach enables the determination of the AI system’s trustworthiness level by analyzing the interdependencies among the entities embedded within its structure. Furthermore, it offers insights into the propagation of trust throughout the system, thereby allowing for sensitivity analysis to assess how trust shifts in response to changes in the trust status of specific entities within the AI system.