4.1. Safety

‘Safety’ stands out as the most prominent and frequently mentioned principle in the entire corpus. To conduct an in-depth analysis of this theme, a set of nine related search terms was compiled, guided by the keywords in both our initial analysis and Jobin et al.’s framework: safe, safety, secure, security, protect, protection, prevent, prevention and integrity. An examination of the collocates for these terms confirmed that their semantic use throughout the corpus consistently falls within the conceptual domain of the safety principle. The emphasis on this theme is quantitatively significant. The combined frequency of these nine terms reaches 2,410 occurrences. Furthermore, their broad distribution across the documents suggests a widespread consensus on the centrality of safety among different types of actors. Specifically, the core term ‘security’ appears in thirty-eight of the forty-seven documents (81 per cent), while ‘safety’ is present in thirty-six (77 per cent).

Within the cluster of terms related to the safety principle, the pairs safety/safe and security/secure are consistently linked, necessitating a clear distinction. A contextual examination of the corpus reveals that safety typically refers to the internal integrity of an AI system and the avoidance of accidental harm, whereas security pertains to its external resilience against attacks. This distinction is explicitly articulated in UNESCO’s Recommendation on the Ethics of Artificial Intelligence, which states that, ‘Unwanted harms (safety risks), as well as vulnerabilities to attack (security risks) should be avoided…’. This conceptual difference is also reflected in their distribution and collocates. Safety appears most frequently in technical, standard-setting documents and its top collocate is function, reinforcing its connection to internal operations. In contrast, security is distributed broadly across all document types, and its most significant collocate is national, pointing to a different set of primary concerns. The prominence of national security suggests that the discourse on AI safety is deeply embedded in a framework of legally protected interests. This securitisation of AI can be understood as part of a broader logic of governance for managing uncertain futures, where principles of precaution, pre-emption and preparedness become central. The focus on ‘safety’ and ‘risk’ in the texts reflects a political desire to anticipate and neutralise potential threats before they materialise, a key feature of contemporary governance imaginaries (Lazaro and Rizzi Reference Lazaro and Rizzi2023). This framework can be conceptualised as a three-tiered model of ‘National Security – Social Security – Individual Security’, an evolution from the traditional Roman law dichotomy of public and private law. In this model, national security is the broadest level, concerning state sovereignty. Social security is the intermediate level, relating to public order and the stability of societal domains. Individual security is the narrowest, encompassing not only traditional personal and property safety but also emerging technological and industrial safety. The development and use of AI implicate all three levels of this security framework, from public interests to private ones. Because the authors in our corpus – ranging from inter-governmental organisations and states to NGOs and private firms – each selectively emphasise different facets of security, the resulting discourse is inherently multi-faceted, regionalised and context-dependent. Consequently, any effort to build a meaningful global consensus must be founded upon a clear acknowledgement of these diverse interests to fully define the principle’s intension and extension. In this context, the ‘safety’ discourse within AI soft law documents can be understood as what the legal sociologist Philip Selznick termed ‘incipient law’ – patterns and claims ‘latent in the evolving social and economic order’ that are highly probable to receive legal recognition in due course. The widespread consensus on safety, emerging from the practices and declarations of both state and non-state actors, forms a normative foundation from which future binding treaties or customary international law on AI safety may evolve. The process of securitisation, therefore, is not just a political manoeuver; it can be seen as a jurisgenerative process where social anxieties about risk are potentially translated into the building blocks of a new legal order.

‘Protection’ is another key term within the safety principle, revealing the primary objects of this concern. The collocation results for protection (Table 5) are dominated by data, privacy, personal and consumer, outlining a clear focus on safeguarding the data and privacy of individuals. This focus is consistently reinforced in the source texts, such as the G20’s call to ‘promote the protection of privacy and personal data’ and UNESCO’s mandate that data processing must align with ‘data protection and data privacy’ laws and principles. Beyond what is protected, the discourse reveals a crucial insight into how protection is framed. Drawing on Foucault’s (Reference Foucault2013) concept of discourse as a socially constructed knowledge and Van Leeuwen’s (Reference Van Leeuwen2005) notion of ‘substitution’, we can observe a key discursive strategy at play: nominalisation. Van Leeuwen argues that replacing verbs (doing) with nouns (being) objectifies an issue, framing it as a settled state. This is precisely what occurs here: the nominalised forms protection and prevention appear with far greater frequency than their verb counterparts protect and prevent. This suggests the discourse treats data and privacy protection not as an ongoing action to be advocated for, but as an already established consensus – a static feature of the governance landscape. This discursive emphasis on protection, however, exists in tension with the practical reality of AI development, where the collection and use of data serve as the technology’s essential fuel. While data, algorithms and computing power are all core elements of AI, massive datasets are what enabled the recent ‘emergent intelligence’ in models like ChatGPT, whose core algorithms have existed for years. A final analysis of the word data itself within our corpus confirms the discursive priority: protection co-occurs with data far more frequently than terms related to data utilisation, such as collection, training, sharing or processing. This suggests a central finding: while data use is paramount for technological advancement, data protection appears to constitute the core theme of the regulatory discourse.

Table 5. Top collocations of ‘protection’

The pronounced emphasis on ‘safety’ is one of this study’s most striking findings. In contrast to earlier analyses of global AI governance datasets, the safety principle is now discussed more frequently, suggesting a new global consensus on its importance. This evolution can be understood through Balkin’s (Reference Balkin1990) concept of ‘ideological drift’. Balkin argues that as political and legal concepts are expressed through signs, their ideological leanings can shift over time as those signs are iterated in new contexts. The rising prominence of ‘safety’ thus demonstrates this diachronic function of signs, where the discourse around AI principles has evolved in just a few years.

Furthermore, the critical nature of social semiotics can reveal previously unrecognised conflicts and prompt a reassessment of existing intuitions. Paul’s (Reference Paul2023) semiotic study of the EU AI Act provides a powerful recent example. Paul deconstructs the EU’s ‘risk-based’ approach, arguing – based on legislative analysis and expert interviews – that the risk classifications were not objective but were rather a ‘political decision and value statement’. This framing was a strategic discursive choice that allowed the EU’s regulatory framework to ‘stand out’ in the global competition over AI norms. Similarly, the discursive construction of the ‘safety’ principle in our corpus merits critical reflection. A sociosemiotic perspective prompts further questions: how, precisely, was the regulation of AI ‘securitized?’ Whose viewpoints does this safety-centric discourse represent and amplify? And through what semiotic processes was this principle generated and propagated?

While fully answering these questions requires more contextual material, the current international political climate offers potential explanations. One is the broader trend of ‘pan-securitization’, where numerous international issues are framed through a security lens. For a disruptive technology like AI, a ‘safety narrative’ offers international actors more control than a ‘development narrative’. Second, the still-maturing international AI regime is heavily influenced by the domestic laws of first-mover jurisdictions, and these legal systems have themselves been observed to be undergoing a clear securitisation process (Mügge Reference Mügge2023; Zeng Reference Zeng2021).

4.2. Human-centric

Alongside safety, a human-centric approach emerges as one of the most significant principle themes in the corpus. This is immediately evident from the prominence of the keyword human, which, with a frequency of 1,686, is the most common of all principle-related terms. Its statistical keyness is second only to keywords that refer to AI itself. A qualitative review of the corpus shows that human combines with other words to form a rich semantic network encompassing concepts like rights, dignity and well-being. Therefore, this analysis uses the term ‘human-centric’ – a label common in both international law and contemporary AI ethics scholarship (Amariles and Baquero Reference Amariles and Baquero2023; Rožanec et al. Reference Amariles and Baquero2023) – as an umbrella concept to encapsulate the diverse principles revolving around the keyword human. To deconstruct this concept, a collocate analysis of human was performed, with the results presented in Table 6. Additionally, the Cluster function was used to identify common two-word phrases (n-grams) involving human, as shown in Table 7.

Table 6. Top collocations of ‘human’

Table 7. Top clusters of ‘human’

The analysis of human’s collocates in Table 6 and word clusters in Table 7 reveals that the ‘human-centric’ principle encompasses two primary dimensions. The first dimension concerns the relationship between humans and AI systems. A contextual examination of key collocates like loop, oversight and autonomy confirms that their semantic use consistently points to human interaction with and control over AI. This theme is further reinforced by the prominence of two-word clusters such as human agency and human intervention. The second dimension is an emphasis on human values, highlighted by collocates like dignity, freedoms and well-being. Within this dimension, the most significant finding is the deep integration of the international human rights framework. The high co-occurrence score of the collocate rights, combined with the statistical prominence of the cluster human rights, demonstrates that international human rights law has become a foundational component of the current AI regulatory discourse.

Regarding the human–AI relationship, the corpus reveals a strong emphasis on human control over technology. This control manifests as: human oversight of automated decisions; human independence within the decision-making loop (ensuring autonomy and agency); and the capacity for direct human intervention. The concept of ‘oversight’ primarily refers to the human review of an AI system’s decisions, typically conducted ex post to correct outcomes. This requires that humans retain effective autonomy, as noted in the Principles for Responsible AI Innovation: ‘human control and oversight are the ability and opportunity for humans to adequately supervise, engage and interfere with an AI system… law enforcement agencies are advised to verify that… humans have the last word’. Documents like the Blueprint for an AI Bill of Rights further specify that such oversight should be enhanced in sensitive domains like the criminal justice system, employment and health care. However, despite this emphasis, the requirements for human oversight are often described in general terms, lacking detailed scenarios or clear hierarchical rules. This is reflected in practice, where legal professionals acknowledge the necessity of verifying AI-generated outputs to mitigate risks like ‘hallucinations’, but formal training on how to do so remains scarce. The reliance on human oversight, therefore, becomes a matter of individual professional responsibility rather than a systematically implemented rule (Terzidou Reference Terzidou2025). Different documents propose varying levels of review, typically based on risk. For example, the EU’s Ethics Guidelines for Trustworthy AI notes that oversight can be required ‘in varying degrees… depending on the… application area and potential risk’, and operationalises this through a questionnaire-style self-assessment list. While flexible, this approach delegates the interpretation of oversight intensity and mechanisms to the implementing entity. This may lead to a diversification of standards in practice, undermining the goal of achieving international consensus. Such vagueness is characteristic of the current international AI regime, which is dominated by non-binding soft law instruments (Elias and Lim Reference Elias and Lim1997). This implementation challenge is rooted in the nature of international legal communication. As a cultural semiotic system, international law must be interpreted by a diverse global community (the ‘receivers’) who may not share the same context as the drafters (the ‘senders’). The comparability and effectiveness of any legal regime depend on a shared interpretation of meaning (Anne et al. Reference Anne, Matulewska and Cheng2020). The inherent ambiguity of soft law, while inclusive, exacerbates this problem, as it can lead to divergent practices that reduce comparability. Therefore, from the perspective of ensuring clear semiotic communication and robust dialogue, hard law mechanisms – such as the Council of Europe’s Framework Convention on Artificial Intelligence – represent a crucial path forward for achieving a more thoroughly negotiated and stable consensus.

In the human–AI relationship, the concepts of human independence, oversight and control are tightly interwoven. The keyword loop, for instance, is central to a tripartite classification found in roughly a quarter of the documents: human-in-the-loop, human-over-the-loop and human-out-of-the-loop, which represent escalating levels of automation and risk.Footnote ² A core tenet of human independence is that even in human-out-of-the-loop scenarios, a person retains the ability to exit the automated process. In the corpus, this capacity is represented by human autonomy and human agency. The Ethics Guidelines for Trustworthy AI define human agency as the ability for users to ‘make informed autonomous decisions regarding AI systems’, while UNESCO’s Recommendation on the Ethics of Artificial Intelligence links human autonomy and human agency directly to the protection of human dignity. However, the principles of control and intervention reveal a significant tension at the heart of AI governance. While regulatory texts call for human control, technical standards often define autonomous systems in contradictory terms. For example, the IEEE Recommended Practice for Assessing the Impact of Autonomous and Intelligent Systems on Human Well-Being defines an intelligent system as one that can ‘carry out some task with or without limited human intervention capable of decision making by independent inference’. This characterisation of AI as inherently free from intervention directly conflicts with the human-centric demand for control. This contradiction likely stems from what appears to be one of the deepest conflicts within AI regulation: the tension between development and safety. This development-versus-safety conflict brings the second dimension of the human-centric principle – human values – into sharp focus. On the other side of the risks posed by AI is its immense potential to benefit society; collocates of human show that ‘development’ and ‘well-being’ are also treated as key legal interests. The governance challenge, therefore, is a fundamental value conflict between the legal interest of safety and that of development. As long as different legislative contexts prioritise these competing values differently, the formation of an effective and stable global consensus will remain a significant challenge.

One of the most significant findings relating to the human-centric principle is the widespread integration of the international human rights law (IHRL) framework, mentioned in thirty-three of the forty-seven documents. This trend is not isolated to AI governance; it is part of a broader ‘transnational human rights’ movement where international norms, constitutional rights and humanitarian law increasingly interact and shape both state and non-state action (Klug Reference Klug2005). This is supported by the high statistical keyness of the terms human and rights and the frequent human rights law cluster. The contextual analysis shows that human rights has a dual meaning: it refers both to core values, often listed with dignity and fundamental freedoms, and to the formal legal framework itself. These two meanings are tightly linked; when documents elaborate on values, they consistently align them with IHRL. For instance, the Principles for Responsible AI Innovation states that its concepts ‘are derived from the Universal Declaration of Human Rights’, while other reports use specific ICCPR articles to define principles like non-discrimination, directly aligning AI ethics with established human rights law. The broad application of IHRL signals a growing recognition of its utility for AI governance. The IHRL framework provides a mature, systematic and universal methodology for assessing and regulating AI, built on decades of global consensus and already integrated into national legal systems (Rodrigues Reference Rodrigues2020). Its suitability stems from its focus on human dignity and its clear delineation of state and private sector responsibilities (Donahoe and Metzger Reference Donahoe and Metzger2019). The direct applicability of this framework is clear when mapping cornerstone documents like the Universal Declaration of Human Rights (UDHR) to key AI principles. For example, the UDHR’s principles of equality and non-discrimination (Article 2) directly address AI fairness; the rights to remedy and due process (Articles 8–11) provide a foundation for transparency and accountability; and the right to privacy (Article 12) anchors data protection requirements. Ultimately, the relationship between the nascent discourse of AI regulation and the established IHRL framework can be understood as a form of ‘manifest intertextuality’ (Fairclough Reference Fairclough1992). The two function as interconnected legal semiotic systems at the international level. The semiotic meaning of AI regulation is endowed and specified through its dialogue with IHRL. Concurrently, the semiotic meaning of the human rights framework is itself reinterpreted and enriched as it is applied to the new and complex problem domain of artificial intelligence, suggesting a dynamic and mutually constitutive relationship.

4.3. Fairness

As one of the three most significant principles revealed in the corpus, fairness is distinguished by having the richest set of associated keywords, indicating its broad semantic range. The centrality of this theme is underscored by the fact that its antithesis, discrimination, is mentioned in thirty-four of the forty-seven documents. The keywords encompassed by this principle include fairness, justice, equality, diversity, inclusion, consistency, accessibility, unbiased and non-discrimination. As a deontological principle, fairness is an essential requirement for ‘AI for good’. However, it is also notoriously vague, containing multiple, often competing, meanings that can lead to value conflicts. This presents a challenge for purely qualitative analysis, as noted in the Artificial Intelligence Risk Management Framework: ‘Standards of fairness can be complex and difficult to define because perceptions of fairness differ among cultures and may shift depending on application’. To navigate this complexity, this study utilises corpus methods (frequency and keyness calculations) to distil the ‘greatest common denominator’ from the discourse across the extensive document set. Based on semantic proximity, keywords under principles of fairness in Table 4 can be grouped into four primary thematic categories: (1) Non-discrimination and Non-bias (non-discrimination, unbiased); (2) Fairness and Justice (fairness, justice, equity); (3) Equality and Accessibility (equality, accessibility, with equity also being relevant); and (4) Diversity and Inclusion (diversity, inclusion). Based on a ranking of their combined frequency and keyness, ‘Non-discrimination and Non-bias’ emerges as the most significant dimension of the fairness principle. This is followed, in order, by ‘Fairness and Justice’, ‘Diversity and Inclusion’ and finally ‘Equality and Accessibility’. The following analysis will further examine these quantitative discourse results.

Non-discrimination and the prevention of bias represent the most significant dimension of the fairness principle, with discussions of bias found in thirty-four of the forty-seven documents and non-discrimination in twenty-six. The discourse is primarily concerned with algorithmic bias, which is defined as ‘systematic and repeatable errors in a computer system that create unfair outcomes’ in the Assessment List for Trustworthy AI. The root causes are identified as flawed or unrepresentative training data and poor algorithmic design. As many documents note, the harms of such bias are amplified in sensitive domains, with frameworks like the AI Safety Governance Framework stating that ‘poor-quality datasets can lead to biased or discriminatory outcomes… regarding ethnicity, religion, nationality and region’. However, a key limitation of the current discourse is that it concentrates on bias as a problem to be avoided rather than articulating a clear, operational vision for an unbiased AI system. This stems from the semantic uncertainty of the term bias itself. Most texts address discrimination in general terms, often using vague modifiers like harmful, unjust or unintended. Even the most precise definitions, such as that in the EU’s Ethics Guidelines, offer only a broad framing of bias as ‘an inclination of prejudice towards or against a person, object, or position’, without providing concrete operational criteria. This lack of specificity can be explained by the inherent distinction between the legal and moral dimensions of discrimination. The moral understanding of bias is often unstable and subject to rapid semantic shifts based on social trends. Law, adhering to a principle of contraction, is inherently cautious about codifying such fluid moral concepts. This challenge, combined with the fact that the meaning of bias is spatiotemporally dependent – varying significantly across different regions and legal cultures – results in a principle that is widely endorsed in theory yet lacks the concrete substance required for effective legal practice.

The collocation analysis shows that fairness and justice are core components of the broader fairness principle, frequently appearing alongside non-discrimination. Similar to the other concepts, it has multiple meanings. Documents like A pro-innovation approach to AI regulation acknowledge this complexity, suggesting that fairness should be interpreted in line with existing legal frameworks such as human rights law. A particularly useful approach, proposed in the EU’s Ethics Guidelines for Trustworthy AI and supported by a contextual review of the corpus, is to divide the concept into substantive fairness and procedural fairness. Substantive fairness, though often defined in general terms, appears widely. For example, the Principles for Responsible AI Innovation defines it as ensuring ‘a just and non-discriminatory treatment of individuals and groups and a contribution to a more equitable society’. Procedural fairness complements this by safeguarding individual recourse. The same documents emphasise that procedural fairness ‘requires that agencies safeguard people’s ability to contest decisions supported by AI systems and to be compensated if such decisions are harmful to them’. This distinction highlights a key dynamic: because the concept of fairness and justice in legislative documents remains generalised and slow to evolve, enforcement and judicial action often take the lead in guaranteeing fairness in the AI domain. However, a significant gap remains, as these enforcement-led approaches currently lack effective, detailed guidance on how to reconcile the competing values of fairness and innovation.

Inclusion, primarily represented by the keywords diversity and inclusion, which often appear together, is another important dimension of the fairness principle. A contextual analysis reveals that inclusion is applied in two primary scenarios. The first is inclusion in the design of AI systems, which requires that development processes be multi-faceted and participatory. As the EU’s Ethics Guidelines for Trustworthy AI states, ‘It is critical that… the teams that design, develop, test and maintain… these systems reflect the diversity of users and of society in general’, including different genders, races, nationalities and cultures. This extends beyond the design teams to also consider the diversity of the technology’s intended audience. Furthermore, organisations are called upon to support this participation, with the Partnership on AI noting that ‘committing organizational resources demonstrates a different level of dedication to pursuing inclusively… developed technology’. The second scenario is the inclusive sharing of AI’s benefits. This interpretation of inclusion is widely mentioned throughout the corpus and serves to link the principle of fairness with the broader goals of human-centricity and trustworthy AI. For example, the OECD’s Policy and Investment Recommendations for Trustworthy AI explains inclusion as ensuring that all people can benefit from artificial intelligence and that it contributes to social well-being.

Finally, the cluster of words represented by ‘equality’ has a certain semantic similarity to ‘fairness and justice’, but from the context of the signs, equality has a different use case compared to ‘fairness’. ‘Equality’ is more absolute, meaning that regardless of whether the situations of different groups are similar, they should receive the same opportunities and protection, whereas fairness is more about adapting to local conditions, focusing on guaranteeing the treatment and opportunities of similar groups. In the corpus, the most common collocates of ‘equality’ include ‘race’, ‘gender’, ‘dignity’, ‘economic’ and ‘rights’. Taking ‘race’ and ‘gender’ as examples, even if there are large differences between individuals, they are ‘absolute’ equality matters in the AI context. In addition, the principle of ‘equality’ is often mentioned together with the principle of ‘non-discrimination’, but at the semantic level, the principle of ‘equality’ goes a step further than the principle of ‘non-discrimination’. In the AI context, non-discrimination still retains the possibility of treating different situations differently based on objective reasons, whereas equality means applying the same rules – for example, the same algorithm rules – so that all groups have equal rights to information, data, knowledge, market access and the distribution of utility (Yu Reference Yu2020). In the corpus documents, equality in the AI context mainly includes three aspects: equality from a human rights perspective (race, gender), equality in the accessibility of technology and equality in sharing the results of AI technology. For example, the Recommendation on the Ethics of Artificial Intelligence released by UNESCO mentions that the goals of the recommendation include equality from a human rights perspective, ‘to protect, promote and respect human rights and fundamental freedoms, human dignity and equality, including gender equality’, as well as equality in accessing AI knowledge and in sharing the results of AI development, ‘to promote equitable access to developments and knowledge in the field of AI and the sharing of benefits’.