1. Introduction

Artificial Intelligence (AI) represents a multifaceted construct within societal and political imaginaries. On the one hand, it is perceived as a driving force for economic growth, technological innovation, and global competitiveness with the capacity to significantly enhance productivity, reshape labour markets, and influence long-term economic trajectories (Filippucci et al. 2024). Governments worldwide have been investing heavily in AI development, endorsed by major international organisations, envisioning progress through its wide-scale deployment. On the other hand, AI has increasingly become associated with societal fears and dystopian projections of the future dominated by uncontrollable and dehumanised systems that diminish human agency and threaten individual well-being (e.g., Cave et al. 2018). These competing narratives not only coexist but frequently move in antagonistic directions, often without effective mechanisms for reconciliation or mitigation. In times of such technological and normative uncertainty, many actors seek guidance through rules, norms, standards, and principles intended to govern change and ensure a smooth transition into an uncertain future. Among these normative instruments, the law plays a distinctive role due to its binding character and formal authority in structuring governance and societal change. However, AI, as an evolving and disruptive technological phenomenon, presents a regulatory challenge.

Historically, during the phases of digitisation and digitalisation, regulatory responses often relied on self-regulation and co-regulation to manage rapid technological changes, engaging the tech industry as an assumed key stakeholder with intimate knowledge of the technologies involved (Marsden 2011). These models, in several instances, successfully fostered legal clarity and supported innovation (Brownsword 2008). In such cases, shared interests between governments and technology actors facilitated cooperative regulatory frameworks. However, in the absence of aligned policy agendas and considering the increasing dominance of oligopolistic and powerful tech companies, this cooperation frequently gave way to unilateral, corporate-driven 'solo-regulation' (Milosavljević & Micova 2016). Efforts to regulate AI have mirrored many aspects of the earlier digital technology regulatory pathways (i.e. risk-based approach). Yet, the uncertainties surrounding AI's societal implications have triggered a somewhat distinct regulatory trajectory. Rather than pursuing wide-scale collaborative regulatory models, where minimal consensus on basic normative standards might have provided a foundation, 'ethics' was introduced as a concrete policy option. This move was accompanied by the promise that "striving for ethics and ethical decision-making […] will make technologies better" (Wagner 2018, 84). However, ethics was also critiqued as a soft and potentially evasive substitute for regulation, offering a veneer of responsibility in contexts where resistance to binding legal norms prevailed: "the rise of the ethical technology debate ran in parallel to the increasing resistance to any regulation at all" (ibid., 85).

Ethics and law are closely connected yet distinct domains: ethics offers philosophical guidance on what is right or wrong, while law sets enforceable societal rules. Ethical theories, such as deontology, utilitarianism, virtue ethics, and care ethics, among others, inform moral judgment, whereas laws are shaped through political and social processes and may reflect, but not fully encompass, ethical values. While law establishes minimum standards and legal responsibility within specific jurisdictions, ethics can go beyond legal compliance, can be applied universally and more flexibly. This distinction becomes particularly important when addressing emerging and unknown challenges – such as AI – where ethical reasoning may fill gaps left by slower-moving legal systems. Hence, across jurisdictions, a wide array of ethical values, such as fairness, transparency, accountability, and human-centricity, are reflected in AI policy guidelines and strategies (Hagendorff 2020) underlying legal frameworks.

Within this context, the presence of ethics in the body of AI policy and legal instruments has been a feature since their conception, often due to these instruments' interlinkage with the concept of 'ethics' as a (desirable) guideline for the development of AI as a technology. In this respect, authors such as Hagendorf (2020), Corrêa et. al (2023), or G'Sell (2024) have mapped the presence of ethics within this corpus: either as a collection of principles providing general guidance on AI development (Hagendorff 2020, 100–101), as a growing body of global policy guidelines that use ethical values as a guiding precept (Corrêa et al. 2023, 3), or as an implicit feature in some existing governance options for generative AI, albeit with a focus on 'weaker' forms of governance, such as self-regulation (G'sell 2024, 20). This last topic has raised the most criticism among several authors, who often perceive this ethical basis as an excuse towards more deregulatory efforts, and thus, as an avenue that can be exploited by private individuals seeking to opt out of more strict forms of AI regulation (Hagendorff 2020, 99–100; G'sell 2024, 21).

However, there remains a limited understanding of how, and to what extent, ethical values, principles and ethics-based policies have been embedded in binding legal acts that are directly or indirectly relevant to AI. This gap constitutes the starting point for our review article, which analyses current legislation in the European Union (EU) – an actor that aspires to function as a global ethics-based standard setter and to exercise normative regulatory power in relation to emerging technologies such as AI. To this end, the article first outlines the ethics-informed principles relevant to AI law and regulation, drawing on a survey of policy documents that have shaped EU legislation. Next, we identify the legal acts subject to closer examination, whereby our corpus comprises seven EU legal acts (483 pages in total) selected to provide a comprehensive legal context for AI governance. We introduce these legal texts and explain their relevance after first tracing the origins and political dynamics behind AI regulation in the EU, particularly the tensions between safeguarding European values and ensuring global competitiveness. Through interpretative legal analysis, we assess the extent to which these principles are reflected in the legal texts, highlighting those that are prioritised and those that are marginalised. In particular, we examine whether ethics-based principles are incorporated into the enforceable (binding) or interpretative (non-binding) provisions of the legislation. We then focus on the most prevalent principles and contrast them with their modes of legal incorporation through doctrinal legal analysis (Byrne and Olsen 2024). Finally, we reflect on the EU's ambition to realise an ethics-informed regulatory approach to AI.

2. Defining Artificial Intelligence for Legal Inquiry

Artificial Intelligence systems have existed since the rise of computing and automation. However, definitions of what constitutes "Artificial Intelligence" have varied throughout the last decades, especially as systems evolve, becoming more complex, varied, and autonomous. The earliest mentions of AI stem from the birth of computing and automation in the 1950s. Here, the Dartmouth Proposal, written by McCarthy et al. (1955), defines AI as the possibility for a machine to simulate any aspect of learning or any other feature of intelligence as precisely as a human can (McCarthy et al. 2018). This definition was later expanded upon to include "the science and engineering of making intelligent machines, especially intelligent computer programs […] that can solve problems and achieve goals in the world as well as humans" (McCarthy 2007, 2).

Other definitions for AI have underscored what capabilities a system would need to be considered intelligent. For example, Russel and Norvig (2022) provide a definition that is based on the Turing Test, that is, a system that can achieve natural language processing, knowledge representation, automated reasoning, and machine learning (Russell and Norvig 2022). However, since the advent of the most recent AI boom in the 2010s and 2020s, later definitions for AI focused on the technology's capabilities and societal implications have emerged from policy actors. For example, the EU High-Level Expert Group on Artificial Intelligence defines AI as a system "that displays intelligent behaviour by analysing its environment and taking actions – with some degree of autonomy – to achieve specific goals" (High-Level Expert Group on Artificial Intelligence 2019, 24). Most notably, the Organization for Economic Cooperation and Development (OECD) defined AI as "a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments" that can also act autonomously and adapt to their environment after their deployment (OECD 2024, 6–7).

The definition of an AI system included in the EU AI Act¹ is borrowed practically unchanged from the OECD's definition (EU AI Act, Art. 3), a widely accepted description eligible for legal interpretation. Therefore, for the purpose of this paper, we also define AI as a machine-based system that is designed to operate with varying levels of autonomy, that receives inputs and generates outputs with specific goals behind that can influence physical or virtual environments, and that may exhibit adaptiveness to its environment after its deployment.

3. Regulating AI in the European Union (EU)

Technological regulation in the EU has evolved rapidly, accelerating since 2022 with the adoption of a risk-based approach aimed at mitigating digital harms, exemplified by the Digital Services Act (DSA)² and the Digital Markets Act (DMA)³. Together with the General Data Protection Regulation (GDPR)⁴, these instruments have reinforced the EU's normative power (Manners 2002), consolidating its reputation as a global standard-setter through norms (Wessel 2021) and underscoring its ambition for leadership in technology regulation as a form of geopolitical influence (Broeders, Cristiano, and Kaminska 2023). Within this broader context – particularly given the rapid rise of AI as a frontier technology – the notion of a global 'race' for AI regulation has gained prominence in policy debates. Scholars have observed the emergence of a competitive dynamic in global AI governance. Smuha (2021) and Koniakou (2023), for example, describe an intensifying race for regulatory hegemony, reflecting both the pace of technological development and the strategic importance attributed to AI. Within this landscape, ethics-informed, 'trustworthy', and human-centric AI regulation has become strongly associated with the European policy sphere (European Commission 2025). This approach has attracted attention for its emphasis on ethical principles rather than alternative regulatory paradigms (Auernhammer 2020). At the same time, it has been subject to sustained critique, and some have questioned the adequacy of ethics-based regulation in the context of global competitiveness and innovation (Schneider 2024), while others highlight concerns about vagueness and limited enforceability, which may undermine regulatory certainty (Csernatoni 2025; Paul 2024). Such critiques have also been echoed by actors within EU institutions themselves (European Commission, European Political Strategy Centre 2025).

Against this background, understanding how and why ethical principles became central to the EU's approach to AI regulation is crucial. Ulnicane (2022) frames this development as both a continuation of the EU's foundational values and part of a longer debate on the legal and moral status of robots and autonomous systems – technologies historically shaped by ethical considerations. Similarly, Müller and Kettemann (2024) link ethics-based AI policy to established European legal traditions, notably respect for human rights and the Charter of Fundamental Rights of the European Union. In parallel, the EU's articulation of digital sovereignty – closely tied to human-centredness, trustworthiness, and ethical AI – has become a key element of its geopolitical positioning (Broeders, Cristiano, and Kaminska 2023) and an explicit political ambition (Delponte 2018; Christou, Meyer, and Fanni 2025), particularly in a global environment where ethical considerations are often marginalised in AI development (European Parliament 2020; Daly et al. 2021; Franke 2021).

Taken together, these dynamics suggest that an ethics-focused AI regulation constitutes a European geopolitical ambition that both reaffirms the EU's legal and policy traditions and positions it as a normative counterweight to less ethically grounded AI development globally (Sajduk and Dziwisz 2024; Franke 2021). At the same time, it may also serve as a means for the EU to assert global relevance in AI governance, primarily as a regulator and norm diffuser rather than as a technological frontrunner. This ambition must therefore be understood within the broader legislative ecosystem shaping AI in Europe, with the AI Act as a central pillar alongside the GDPR, DSA, DMA, and related frameworks governing data and intellectual property.

4. Laying Out an Ethics-Based Analytical Frame for Legal Analysis

Ethics, as a branch of philosophy, is concerned with questions of what is right, what is wrong, and how individuals ought to behave. Deontological theories, such as those deriving from Kant's categorical imperative, assess actions based on their intrinsic nature, namely, whether the act itself is right or wrong. In contrast, teleological theories, such as utilitarianism, focus on outcomes and judge actions by their consequences, for example, by aiming to maximise overall well-being. Other significant ethical theories include virtue ethics, which traces back to Ancient Greece and Aristotle and centres on the moral character of the actor, and care ethics, which has its roots in nursing ethics and prioritises relationships, empathy, and context-specific moral considerations.

More specifically, political philosophy is concerned with the relationship between ethics and law. Burg (2011) describes ethics and law as twin disciplines, noting the difficulty of clearly delineating the boundaries between legal norms and moral values, given the existence of numerous exceptions (Shklar 1964, as cited in Burg 2011). While the law usually reflects society's minimum expectations and codifies and enforces ethically desirable behaviour, ethics extends beyond these minimum standards (Trevino and Nelson 2021). Lennerfors (2019) argues that laws are socially and politically constructed and that the relationship between law and ethics is not inherent; laws may emerge from political negotiations that prioritise certain ethical values while neglecting others, and they may reflect unequal distributions of power within society. Nevertheless, it is also argued that unjust laws can and should be challenged through ethical critique and public discourse. Furthermore, actions may be considered ethical but not legal, and vice versa (Kotsios, Lennerfors, and Laaksoharju 2025; Chance et al. 2025).

A further dimension concerns legal compliance. Adhering to the law merely as a minimum standard, without considering broader ethical implications, can lead to a form of moral complacency. Taebi (2021) argues that ethical guidance can help clarify what constitutes the "right" action, particularly in situations where the law remains silent or insufficient. In the context of AI, it is important to note that law often lags behind technological developments, whereas ethics can be more responsive and adaptive. This has been demonstrated by Poel and Royakkers (2023), for example, in their analysis of the evolving concept of responsibility in relation to AI. However, the risk of "ethics washing" has become increasingly salient. Wagner (2018) warns that while policymakers broadly agree on the need for ethical oversight of AI, there is far less consensus on what this should entail. The invocation of vague ethical principles without concrete enforcement mechanisms risks delaying substantive regulation and the prevention of societal harms, and ethics "toothless" (Rességuier and Rodrigues 2020) and "useless" (Munn 2023).

The EU's ethics-focused approach to AI regulation draws on a longstanding tradition of embedding fundamental rights, human dignity, and democratic values within its legal and policy architecture. While this vision is normatively compelling, it raises important questions regarding how abstract ethical principles are selected, operationalised, and translated into binding legal norms. To critically examine EU legal instruments relevant to AI, it is therefore necessary to systematically explore how ethical reasoning is embedded in the law, to what legal effect (binding or non-binding), and how key ethical concepts are interpreted in legal terms. Accordingly, our analysis first outlines the ethics-informed principles relevant to AI law, drawing on a survey of policy documents that have shaped EU legislation.

A wide range of ethical values – such as fairness, transparency, accountability, and human-centricity – feature prominently in AI strategies and guidelines across jurisdictions, including those of the EU (Hagendorff 2020). An expanding body of scholarship addresses which values should guide AI development and which ought to be embedded in AI systems themselves. Floridi and Cowls (2019), for instance, propose five core principles for ethical AI, drawing on the bioethical principles of beneficence, non-maleficence, autonomy, and justice, while adding explicability as a fifth principle. Hagendorff (2020, 103) further observes that certain values – such as accountability, explainability, privacy, justice, and robustness – are more readily operationalised in technical systems and therefore more likely to be implemented in AI design. At the same time, Sutrop (2020, 57) raises normative concerns regarding value alignment, questioning whether shared understandings of desirable outcomes are feasible given the plurality of human goals and preferences.

This plurality of values is also reflected in major international AI ethics instruments that have informed the EU's legislative work, including the Ethics Guidelines for Trustworthy AI (European Commission 2019)⁵, the OECD AI Principles (OECD 2019)⁶, and the UNESCO Recommendation on the Ethics of Artificial Intelligence (UNESCO 2021)⁷. These documents constitute forms of applied ethics and articulate key themes such as non-maleficence, autonomy, and justice, while seeking to operationalise them within policy frameworks. Methodologically, the ethical values and principles articulated in these instruments provide analytical access for examining EU AI legislation by tracing how such principles are embedded, prioritised, or diluted within the legal texts. By grounding our analytical categories in these authoritative documents, we aim to capture both the internal logic of EU AI policymaking and its external projection of regulatory power. Accordingly, we combine the values and principles from these instruments into a single interpretative framework (see Appendix 1) for analysing EU laws relevant to AI, with particular attention to the distribution of ethical values across binding and non-binding provisions, in order to assess the normative weight accorded to ethics within the EU's AI regulatory framework (Floridi 2018).

5. Legal Interpretive Analysis of AI-Relevant Laws and Regulations of the EU

Our empirical work followed a two-step approach. First, we identified a set of key legal instruments that directly or indirectly shape the regulatory landscape for AI within the European Union. These instruments were selected because they represent the EU's most comprehensive and strategic efforts to govern different dimensions of AI development, deployment, and use. While the AI Act constitutes the central legislative instrument specifically dedicated to AI, the remaining texts reflect the broader regulatory ecosystem within which AI operates. These include foundational legislation on data protection (GDPR), platform governance (DSA), data access and sharing (DGA⁸; Data Act⁹), and intellectual property rights (DSM Directive¹⁰) (see Table 1 for an overview of the most relevant EU legal acts with direct or indirect relevance to AI).

Analysing this corpus of legislation enables us to trace how ethics-based values and principles, as identified in the three AI policy guidance documents, are translated into binding and non-binding legal norms across different domains of EU law relevant to AI governance. This analysis was guided by three central research questions: (1) which ethical values and principles are embedded in EU AI-relevant legislation; (2) which of these values are prioritised or marginalised; and (3) to what extent they are legally enforceable.

In the second step, we draw on the results concerning the most prevalent ethical values to conduct an in-depth legal doctrinal and ethics-informed analysis, focusing on the alignment – or lack thereof – between ethical principles and their legal operationalisation. Finally, we reflect on the EU's ethics-informed regulatory ambitions in light of our empirical findings.

The empirical analysis in this study is grounded in a broadly applied legal doctrinal methodology (Hutchinson and Duncan 2012) involving a systematic and interpretive engagement with legal texts in order to identify, explain, and reconstruct the normative logic underpinning legal rules and principles. As applied here, this approach enables a close reading of how legal instruments express and embed ethical values and is particularly well suited to examining the internal coherence, justification, and normative content of EU AI-relevant legislation. Our interpretative reading adopts an internal perspective on law, treating legal norms as part of a normative order shaped by shared meanings and legal reasoning, entering into a dialogue with the legal texts to reconstruct the reasoning structures (Byrne and Olsen 2024).

We applied such legal interpretation to the selected set of EU legal instruments that either directly regulate AI (such as the AI Act) or indirectly shape the governance environment in which AI systems are developed and deployed (including the GDPR, DSA, DGA, Data Act, and DSM Directive). These instruments were analysed for references to the predefined set of eleven ethical values and principles derived from the three international guidelines (Appendix 1). Particular attention was paid to the legal placement of ethical values, specifically whether they appear in binding provisions (such as Articles and Annexes) or in non-binding elements (such as Recitals). This distinction is central to assessing the legal enforceability of ethical values, as binding provisions impose enforceable obligations, while Recitals primarily guide interpretation and clarify legislative intent.

6. Ethical Values and Principles in the EU Laws

Drawing on the eleven ethical values and principles defined as analytical categories (Appendix 1), we interpreted the selected EU legal instruments. For each principle, key terms, indicative expressions, and definitional markers were extracted from relevant policy guidelines and translated into the formal legal language of EU law. For example, fairness encompassed references to bias mitigation, equality of treatment, and equitable data governance; explicability included traceability, auditability, and duties to inform affected persons; and non-maleficence was linked to harm prevention and risk-based safeguards. Coding focused on the proximity between ethical categories and their legal expression, allowing assessment of substantive presence, normative priority, and enforceability. This approach moved beyond keyword searches towards a context-sensitive interpretive analysis. Results are summarised in Appendix 2.

7. Discussion of the Ethical and Legal Conceptualisation of the Core Principles

8. Conclusion and Looking Ahead

The EU's AI-relevant legislative framework reflects a multi-value ethical architecture, with non-maleficence, fairness and privacy forming the enforceable core, and explicability playing a significant supporting role. Sustainability, solidarity and democracy remain underdeveloped in binding terms, largely confined to aspirational statements. The reliance on non-binding language for key ethical commitments risks leaving them under-realised, particularly as generative AI and other emerging technologies expose new normative challenges. We are not naïve: our finding goes in line with the political agreements at the international level on this matter, where non-binding recommendations have taken the forefront in international AI regulation – at the cost of risking a high level of protection of several fundamental rights.

Our aim was to draw attention to values that underpin European democracies and to ask whether legislation on AI reflects these values adequately, and whether the EU's proclaimed policy of trustworthy and human-centric AI is truly embedded in law. The findings reveal shortcomings: values tied to democratic processes are treated more as an "afterthought" than a priority, while human-centricity lacks binding and enforceable guarantees of meaningful human oversight. At the same time, the vagueness of fairness across instruments and its differing interpretations highlight the instability of this principle as a regulatory foundation, also due to its multi-faceted and highly context-dependent nature.

A limitation of our paper is that it does not explore in depth the political and legislative dynamics behind these outcomes, nor the division of responsibilities between EU institutions and member states in governing AI. Due to the complexity of a study of that nature, It rather takes a pragmatic stance in the geopolitical realities into which these instruments were conceived: a global environment where few global players concentrate AI development, and an incipient yet growing global regulatory response that is rich and consistent in terms of principles and standardised practices but vague and reliant on non-binding instruments to ensure their functioning. Nevertheless, the analysis demonstrates how this global landscape results in ethics-based commitments that are then selectively translated into law and identifies the gaps in leaving central democratic and societal values outside the enforceable core. Future research should investigate not only the effectiveness of ethics-based regulation, but also how ethical values can be more robustly enacted in legal frameworks to support both technological development and the legitimacy of AI governance in Europe.

Appendix 1: AI Values and Principles

Appendix 2: Ethical values and principles in the EU AI-relevant legislative instruments

AI Use Statement

The authors used an AI-based language model (ChatGPT, version 5.2) solely for language-related assistance, including improving clarity, coherence, consistency of terminology, and proofreading of the manuscript. The AI tool did not contribute to the development of the research design, the formulation of research questions, the analysis or interpretation of data, or the development of substantive arguments. All intellectual content, analytical decisions, and conclusions presented in the article are the sole responsibility of the authors.

Funding Information

This work was partially supported by DIACOMET ("Fostering capacity building for civic resilience and participation: Dialogic communication ethics and accountability" project, funded from the European Union's Horizon Europe research and innovation grant agreement No 01094816).

Competing Interests

The authors declare that they have no competing interests.

Author Contributions

Krisztina Rozgonyi coordinated the project behind this research. Krisztina Rozgonyi and Mari-Liisa Parder conceptualized the study and designed the research methodology. Krisztina Rozgonyi, Mari-Liisa Parder, and Rodrigo Conde Jiménez contributed to the theoretical framework. Krisztina Rozgonyi carried out the formal analysis and data collection for this article. Krisztina Rozgonyi drafted sections 1, 5, and 6 of the manuscript, and contributed to sections 2, 7, and 8. Mari-Liisa Parder drafted sections 4 and contributed to sections 7 and 8. Rodrigo Conde Jiménez drafted section 3 and contributed to sections 2, 7, and 8. All authors contributed to reviewing and editing the manuscript and approved the final version for publication.