Ethical AI in healthcare: insights from European policy discourses

Abstract

Background

The rapid integration of Artificial Intelligence (AI) into healthcare presents unprecedented opportunities for innovation in diagnostics, treatment and patient care. However, this paradigm shift is accompanied by a concomitant series of critical regulatory and ethical challenges. Relying on the pillars of the New Institutional Theory, this study undertakes a comprehensive examination of how ethical AI in healthcare is discursively constructed in European Commission’s (EC) policy communications and how institutional narratives shape emerging governance frameworks.

Methods

A multi-method approach was employed, integrating both deductive qualitative content analysis and computational topic modelling using Latent Dirichlet Allocation (LDA) applied to 95 official European Commission documents published between 2018 and 2025. The sequential combination of qualitative and computational methods enable both theory-informed interpretation and systematic identification of latent thematic patterns.

Results

Our findings reveal four dominant themes: (1) healthcare AI applications and data governance, including privacy and interoperability issues; (2) regulatory risk frameworks, emphasizing liability and high-risk classification; (3) Europe’s stance as a leader in trustworthy and ethical AI; and (4) the socioeconomic ramifications of AI adoption, including risks of bias, exclusion, and unequal access. While the Commission frequently emphasizes “trustworthy AI” as a guiding principle, ethical challenges such as algorithmic bias and opacity are primarily framed at a high level of abstraction, with limited specification of operational implementation mechanisms.

Conclusions

This study adds to the ongoing discussions on ethical AI by demonstrating how institutional communication functions as a form of institutional work that constructs legitimacy, normative alignment, and ethical responsibility in healthcare AI governance. Rather than evaluating policy adequacy, the analysis highlights how ethical considerations are discursively positioned within broader regulatory narratives.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12910-026-01403-9.

Introduction

The European Commission (EC) defines Artificial Intelligence (AI) as “systems that display intelligent behaviour by analysing their environment and taking actions – with some degree of autonomy – to achieve specificgoals” (2018).1 The accelerated evolution of AI technologies has outstripped the development of corresponding regulatory and normative frameworks, particularly in domains that are pivotal to the preservation of democratic values and human rights [9, 10, 42, 61]. While governments and international institutions endeavour to address this deficit, an increasing number of nations have introduced national AI strategies and policies to facilitate the responsible utilisation of these technologies [59].

At present, over 175 countries, companies, and organisations have published documents that set out ethical guidelines for AI, which reflects a heightened focus on regulating and safeguarding these systems. However, the efforts to regulate AI have been somewhat fragmented, with multiple, often competing, initiatives. Despite this fragmentation, indications of convergence are emerging, particularly as the United States, Europe, and other nations coalesce around shared principles [67].

Furthermore, while global organisations such as the Organisation for Economic Co-operation and Development (OECD) and the United Nations Educational, Scientific and Cultural Organisation (UNESCO) have established overarching AI principles with the objective of shaping ethical standards and practices, the EU stands out due to its comprehensive, legally binding regulatory approach.

Following this approach, the EC has placed an increasing emphasis on the strategic potential of AI in addressing critical societal and human challenges, while also shaping the EU’s vision, priorities, and strategy for AI integration [12, 13, 15–18]. A significant milestone in this effort was the publication of the Ethics Guidelines for Trustworthy Artificial Intelligence in 2019,2 which delineated seven fundamental principles for responsible AI development. These criteria – rooted in fundamental rights and ethical principles – prioritize fairness, accountability, and transparency in AI governance. The guidelines align with broader global debates on AI ethics [22, 31], which emphasize the importance of balancing innovation with societal safeguards.

Overall, the EU’s AI strategy is indicative of the intersection of societal needs, ethical concerns, and regulatory challenges [63]. The EU has distinguished itself as a global leader in the governance of AI through the establishment of the Artificial Intelligence Act (AIA), a seminal piece of legislation that aims to establish a harmonised regulatory framework for the development and deployment of AI. This initiative represents a significant departure from voluntary ethical guidelines or fragmented national regulations, positioning the EU at the vanguard of AI governance on a global scale.3 Moreover, the EU’s regulatory strategies have been shown to exert a notable influence on global AI policy, a phenomenon known as the “Brussels Effect” [2]. This effect refers to the tendency for European standards to shape international markets and governance models.

To date, scholars have critiqued such high-level ethical frameworks for their lack of enforceability and sector-specific guidance [26]. For instance, while the EC promotes principles of equitable AI, the operationalization of these principles in highly regulated and high-stakes domains such as business and healthcare remain a challenge, with scholars calling for clearer regulatory mechanisms to translate these guidelines into practice [44, 45].

In the context of healthcare – a sector characterised by ethical sensitivity – the accelerated development of AI has the potential to engender both transformative prospects and substantial regulatory and ethical dilemmas [47, 51].

In contrast to other sectors, errors in algorithmic decision-making in healthcare can have immediate consequences, as these decisions can directly impact patient outcomes [38, 64]. Moreover, the necessity of holistic patient data, the intricacies of co-morbidities, and the imperative for precise diagnostics create a landscape where companies might be incentivized to navigate regulatory grey areas [23, 64]. The potential for regulatory misinterpretation or circumvention is heightened in healthcare compared to other sectors [29, 33]. These dynamics make healthcare a critical domain for studying AI governance communication. Furthermore, as AI-driven solutions increasingly influence diagnostics, treatment planning, and personalized medicine, ensuring compliance with ethical and regulatory standards is paramount [14, 19, 39, 47, 53].

The intricacies inherent in this domain, coupled with the paucity of studies that delve into the specific influence of institutional discourses on shaping AI policies in healthcare [1, 6, 61] have guided the authors’ decision to undertake a focused analysis of the EU’s communication on AI, with a specific focus on the healthcare sector.

To address this gap, our study examines the European Commission’s formal communications to identify key themes, strategic alignments, and regulatory challenges characterising AI governance and global regulatory trends.

This study aims to address the following research questions:

• RQ1. What are the principal themes and subjects addressed in the European Commission’s communications on AI in healthcare?

• RQ2. How do the European Commission’s AI strategies correspond with broader European Identity (EI) politics?

• RQ3. What policy recommendations are proposed in the European Commission’s communications on the integration of AI in healthcare?

By analysing formal communication documents, this study broadens the scope of discourse analysis in organizational research, emphasizing the role of public institutions in AI regulation.

From a practical perspective, the findings of this paper aim to provide actionable insights for regulators and policymakers on how discourse shapes public perception and policy formation around AI in healthcare.

The paper is organised as follows: Sect. 2 outlines the theoretical framework. Section 3 explains the research methodology. Section 4 presents the key findings. Section 5 discusses the implications of these findings for both policy and practice. Finally, the paper concludes with a summary of the contributions, limitations and directions for future research.

Literature review

The EU has placed significant emphasis on the concept of digital sovereignty as a cornerstone of its digital and AI policies [46]. This concept is indicative of the EU’s aspiration to exercise control over key dimensions of the digital realm, including physical infrastructure, standards and rules, and data and content. As articulated by EC President Ursula von der Leyen in 2020, digital sovereignty, in conjunction with “open strategic autonomy”, constitutes a vital component of Europe’s broader global ambitions. This approach emphasises the EU’s commitment to establish “responsible”, “human-centric”, and “trustworthy” AI as fundamental tenets of its regulatory framework.

In this innovative contextm various stakeholders, including governments, the private sector, and civil society, are engaged in developing governance models that are shaped by shared principles, norms, rules, decision-making processes, and economic factors.

The EU’s proposed Artificial Intelligence Act (AIA) is at the core of this framework, employing a “risk-based” approach to categorise AI systems. For instance, systems posing “unacceptable risks”, such as biometric surveillance technologies, are prohibited outright. Conversely, systems designated as “high-risk”, including predictive algorithms in hiring processes and AI-powered medical diagnostic tools, are subject to rigorous pre-market conformity assessments. This ensures that such systems are compliant with safety, ethical, and transparency standards prior to their release into the market. The European Declaration on Digital Rights and Principles4 reinforces this view by stating that technology should unite people and contribute to a fair, inclusive society and economy. The EC has similarly articulated a vision for a digital society by 2030,5 emphasizing that no one should be left behind in the digital transformation, with particular attention given to those who may be vulnerable or least likely to benefit.

The EU’s pursuit of digital sovereignty, characterised by its endeavours to exercise control over digital infrastructure, standards, and data governance, finds close alignment with the dynamics delineated by New Institutional Theory. This theoretical background underscores the manner in which institutional frameworks are established and deconstructed over time, frequently reflecting the actions and strategies of influential actors, known as Institutional Entrepreneurs (IEs) [4, 11, 13]. These actors engage in what is termed “Institutional Work”: actively shaping, modifying, or dismantling institutional norms and practices to align them with their objectives [27, 37].

In the context of the EU’s digital policies, particularly its AIA, IEs comprising policymakers, industry leaders, and advocacy groups play a crucial role. By leveraging legitimacy and strategic influence, these actors drive the development of regulatory norms, including the “risk-based” approach central to AI governance. This framework, anchored in the concept of digital sovereignty, gradually takes shape over time, evolving into a set of institutional prescriptions that actors seek to align with, often in pursuit of legitimacy.

The interplay between digital sovereignty and the institutional framework becomes evident when considering the EU’s emphasis on “human-centric AI” and “trustworthy AI” [28, 68].

The insights offered by neo-institutionalist theory are particularly valuable in this regard, as it focuses on key dimensions such as:

• Institutional Entrepreneurship: this dimension examines the role of IEs in influencing and reshaping institutional contexts, including their use of institutional strategies to navigate and shape the regulatory landscape [11, 27].

• Institutional Work: highlighting the broader processes of constructing, maintaining, and transforming institutions through actions by both powerful and less powerful actors, emphasising collaborative dynamics in institutional evolution [5, 36].

A more detailed examination of the theoretical framework of this paper reveals its foundation in the role of communication as a pivotal mechanism in the formation of institutional norms and practices. From the perspective of Institutional Work, communication mechanisms serve as strategic levers, essential for constructing and disseminating discourses that guide decision-making processes leading to constraints and norms (regulatory framework) within a concrete system of action, resulting from a mix of games (conflicts/negotiations), values, and rules that temporarily structure a local order [24].

It is through communication that institutional actors are able to establish their legitimacy, influence the decision-making process, and create the symbolic and cultural meanings that underpin regulatory frameworks [69]. Institutional change is often the result of a dialectic process involving competing discourses that shape and reshape norms over time. In the context of AI governance, the creation of stratified texts – such as policy documents, regulatory proposals, and strategic communications – that legitimize and promote the adoption of responsible AI practices is imperative [40].

In the realm of AI adoption across entrepreneurial sectors, IEs leverage various resources, such as:

• Reputation Capital. Success in implementing and innovating AI technologies enhances the credibility of IEs, fostering a positive reputation crucial for influencing stakeholders [3, 57, 65, 70–72].

• Apical Roles in AI Communication Processes. Holding influential roles in communication processes allows entrepreneurs to shape how AI technologies are portrayed to the market, clients, and regulatory bodies. Steering the narrative around the ethical use, benefits, and risks of AI enables Entrepreneurs to influence market trends and regulatory discussions [73].

The theory on Institutional Work emphasizes that institutional activities are often shaped and legitimized through discourse. Communication, whether verbal or written, serves as a fundamental mechanism to influence regulatory frameworks and institutional landscapes. In this regard, the organizational discourse theory [74] is widely used in institutionalist literature to analyse Institutional Work processes [27, 36]. Discourse theory encapsulates a broad range of methods and approaches to organizational analysis, focusing on “organizational discourse” as a structured set of texts embedded within written and verbal practices.

As anticipated before, New Institutional Theory emphasizes the role of institutional structures, norms, and discursive practices in shaping governance frameworks and decision-making [12, 56]. Institutions, in this sense, serve as both enablers and constraints for policy development, influencing regulatory discourses and standard-setting processes [24, 48].

In the context of this study, Institutional Theory provides a lens through which we can understand how the EC constructs its regulatory and ethical discourse around AI governance in healthcare [66]. Institutional work aligns with the coding taxonomy (see Appendix A) by categorizing discourses that legitimize AI policies and governance mechanisms [58].

Each category can be mapped onto key dimensions of Institutional Theory:

1. Descriptive Framework on AI & European Identity (Normative Pressure): the EU’s emphasis on human-centric and trustworthy AI reflects normative isomorphism [12].

2. Regulatory Landscape & Risk Categorization (Regulatory Institutionalization): the categorization of AI systems into high-risk, limited-risk, and minimal-risk aligns with the regulatory logic of institutional frameworks [43].

3. Ethical and Transparency Issues (Mimetic Processes): the concern with algorithmic bias, privacy, and ethical considerations in AI governance mirrors institutional processes where policymakers model their frameworks based on perceived best practices and ethical norms [50, 55].

4. Data Governance and Digital Sovereignty (Institutional Entrepreneurship): the focus on data protection, privacy, and interoperability demonstrates the EU’s role as an institutional entrepreneur [5].

5. Healthcare Applications and Member States’ Digital Level (Institutional Diffusion): the taxonomy’s inclusion of AI’s role in health data governance and digital health adoption across member states reflects institutional diffusion – where policies spread across national and regional levels through regulatory alignment and policy learning [60, 75].

Methodology

The methodology employed in this research involved a systematic approach to gather and analyse communications and speeches on AI in healthcare from the European Commission’s Press Corner website (https://ec.europa.eu/commission/presscorner/). The analysis of the collected data entailed a content analysis, focusing on the communication surrounding AI in healthcare as articulated by members of the European Commission in the documents published on the official Press Corner website, from 15 March 2018 to 13 March 2025. This time frame was selected because it corresponds to the period in which the European Union significantly intensified its regulatory activity on AI, including the proposal and subsequent development of the AI Act.

Initially, a comprehensive search was conducted using specific keywords related to AI and healthcare, resulting in a pool of 202 search results. Subsequently, the authors downloaded the identified documents and meticulously screened them according to explicit eligibility criteria. Documents were included if they (1) were issued by the European Commission, (2) explicitly referred to artificial intelligence in healthcare or health-related policy contexts, and (3) addressed regulatory, ethical, or governance-related aspects of AI. Documents were excluded if references to AI or healthcare were incidental or unrelated to policy framing or governance. Following this screening process, 95 documents were selected for analysis.

To identify these documents, a detailed examination was performed, targeting specific keywords such as “health”, “medicine”, “pharma”, “artificial intelligence”, and “AI” within the text. The selected documents were then catalogued in an Excel spreadsheet. To investigate how the European Commission reconciles its AI strategy with European Identity politics in its framing practices, the authors examined four types of EC documents released during the selected period: speeches, press releases, Q&A documents and statements. These document types were chosen because they represent formal institutional communication through which the Commission articulates regulatory priorities, policy rationales, and strategic narratives, collectively shaping the discourse on AI and digital innovation in the EU [20].

In this study, content analysis was applied as a sequential mixed quali-quantitative approach, combining computational and interpretative techniques to identify and quantify recurring policy themes [34, 49, 61]. Analytical constructs or rules of inference were used to link the text to research questions [76]. The combination of methods was designed to integrate theory-driven interpretation with data-driven pattern detection, thereby strengthening analytical transparency and robustness.

The analysis, spanning July 2024 to March 2025, followed a two-stage analytical strategy integrating deductive and inductive approaches.

The study adopts a sequential mixed-methods design. Firstly, a deductive qualitative content analysis was conducted to identify theory-informed discursive categories derived from New Institutional Theory. Secondly, Latent Dirichlet Allocation (LDA) topic modelling was employed as an inductive technique to reveal latent thematic structures that were independent of the predefined coding scheme. This sequence enables the empirical validation and refinement of theory-driven interpretations through data-driven pattern detection. The integration of these methods enhances analytical robustness by combining interpretive depth with computational objectivity.

Operationalization of New Institutional Theory

In the first stage, a deductive coding framework was developed explicitly grounded in New Institutional Theory, using predefined analytical categories such as regulative mechanisms, normative frameworks, legitimacy narratives, and institutional identity. These categories were derived from core institutional concepts including legitimacy, institutional logics and isomorphic pressures, and were used to guide analytical decisions throughout the coding and interpretation process. Deductive category assignment [77] was applied for the entire data corpus in NVivo®, where the coding units were defined as sentences or logical sentence parts containing one or more of the target keywords (e.g., AI, health, data, ethics). This approach was particularly suited to addressing the descriptive dimension of the research design, which seeks to identify the elements present in the data rather than explain them (RQ1: themes and subjects).

Axial coding techniques were used to group related codes into more general categories [78]. An initial codebook was drafted, classifying keyword mentions into core categories – AI (Artificial Intelligence), healthcare, and medicine – which served as the primary search and classification anchors. These represented the foundational keywords for document identification. However, additional related terms (e.g., health data, ethics, risk, digital health, trust) were subsequently incorporated during the coding process to capture emergent concepts. Following this approach, four broad discourses encapsulating how the Commission frames AI were identified: 1) Ethical and Legal Regulation, 2) Innovation and Competitiveness, 3) Data Governance and Trust, and 4) European Leadership and Identity. These discourses reflect distinct institutional narratives through which AI governance is framed and legitimized in EU policy communication.

Two authors independently coded and analysed the full selected documents, comparing findings and resolving areas of confusion. Independent double coding was adopted to minimise potential sources of bias, including individual coder subjectivity, inconsistency in code application, and ambiguity in category interpretation [35, 79]. Revisions to category definitions were discussed and incorporated into the final codebook, ensuring analytical consistency [49].

While the initial search strategy used a combination of AI- and healthcare-related keywords to identify relevant EC communications, the subsequent analyses were conducted on the full text of each of the 95 selected documents. Accordingly, frequency counts reflect the empirical distribution of terms across entire documents rather than search-driven sampling effects. This means that frequency counts (e.g., for ‘health’, ‘data’, or ‘intelligence’) reflect the actual distribution of terms across full documents, rather than a search bias or excerpt-level sampling. Thus, the thematic prominence of healthcare-related vocabulary emerged empirically from the corpus as a whole, not from the inclusion criteria alone. Illustrative examples of codes aggregated under each discourse and representative quotations from the coded corpus are provided in Appendix B.

In the second stage, Latent Dirichlet Allocation (LDA) topic modelling was carried out for the entire corpus employing R (v. 4.3.2) as an inductive analytical complement to the deductive qualitative analysis. This technique, widely recognized in policy analysis and digital text mining [7, 30], enables the detection of latent structures in the dataset, allowing for an empirically grounded identification of discourse trends. The computational analysis aimed to uncover latent thematic patterns independent of the pre-established codebook. The dataset was first loaded and standardized through a pre-processing phase that included text cleaning, lemmatization, and the extraction of frequent bigrams and trigrams [32]. This approach ensured to preserve key multi-word expressions (e.g., “artificial intelligence”, “data protection”) for further analysis [32].

The processed text has then been transformed into a Document-Term Matrix (DTM), which represents the frequency of each term in each document [8]. The number of topics was determined by model fit criteria. The final LDA model has been trained on the DTM, uncovering latent thematic structures in the text. LDA’s assumption that documents are mixtures of topics and each topic is a mixture of words makes it particularly suitable for identifying dominant policy narratives in institutional communication [25].

The results from the LDA were interpreted using a qualitative approach and compared with the codes extracted in NVivo®. Convergences and divergences. In this sequential process, the LDA acted as an inductive validation step, confirming and expanding the results of the prior deductive analysis. Thus, NVivo® enabled theory-guided interpretation, while LDA (computed in R 4.3.2) enhanced analytical transparency by revealing latent thematic patterns.

By doing so, this process allowed both the computational objectivity and interpretive depth: while NVivo® coding was grounded in theoretical constructs, topic modelling through LDA picked out themes that might otherwise have remained undetected.

Finally, interpretive synthesis of manually coded data was used to address RQ2 and RQ3. For RQ2, themes identified through qualitative coding were conceptually mapped onto dimensions of European Identity politics, interpreted through an institutional lens emphasizing legitimacy, values, and collective identity formation. For RQ3, the policy recommendations expressed in the communications were analysed in relation to coded categories that corresponded to regulatory mechanisms. This guaranteed methodological coherence across research questions while maintaining consistency in the application of institutional theory.

To ensure analytical traceability, institutional theory informed the analysis at three interconnected levels: (1) the construction of the deductive coding framework (e.g., legitimacy, regulative and normative mechanisms, and institutional identity); (2) the aggregation of first-order codes into higher-order discourses reflecting institutional narratives; (3) the interpretive mapping of empirically identified themes onto institutional processes, such as coercive pressures and legitimacy construction. In this manner, New Institutional Theory provided guidance not only to conceptual interpretation but also to concrete analytical decisions throughout the research design.

Findings

An examination of the European Commission’s communications on AI in healthcare reveals several key themes that align with the core objectives of the study. The analysis integrates results from deductive qualitative coding in NVivo® and inductive topic modelling (LDA), ensuring both interpretive and data-driven robustness. In line with the sequential mixed-methods design, qualitative coding provides contextual and theory-informed interpretation, while topic modelling serves as an inductive validation and enrichment tool. Specifically, this section explores the principal themes and subjects articulated in the European Commission’s communications on AI in healthcare (RQ1); how the European Commission’s AI strategies correspond with broader European Identity (EI) politics (RQ2); and the policy recommendations proposed in these communications concerning the integration of AI in healthcare (RQ3).

Dominant lexical and thematic patterns

This subsection directly addresses RQ1. The results presented in this section derive from the NVivo®-based qualitative coding and lexical frequency analysis of the 95 European Commission communications included in the corpus. The most frequently cited terms serve to emphasise the centrality of healthcare in the general disclosure on AI by the members of the EC. For the purposes of this analysis, a term was considered “frequent” when it appeared consistently across multiple documents and document types, rather than being concentrated in a small subset of texts.

The term “health” appears as the dominant expression (2,282 occurrences), followed by “data” (2,032) and “intelligence” (1,751). This reflects a pronounced emphasis on the deployment of AI-driven systems within the context of healthcare, with a particular focus on the role of data in these innovations.

The terms “risk” (418 occurrences), “trustworthy” (362), and “rules” (296) were used in a variety of communications, most often in relation to legal frameworks, ethical standards, and technical compliance.

According to a contextual reading of these references:

• “trustworthy” was frequently associated with the EU’s Trustworthy AI framework and calls for human-centric innovation (Tech and Geopolitics: Building European Resilience in the Digital Age – Speech by Commissioner Thierry Breton, 5 September 2023; President von der Leyen’s Speech at the High-Level Opening Session of the 2021 Digital Assembly: “Leading the Digital Decade”, 1 June 2021);

• “rules” was frequently used to indicate legislative safeguards or standard-setting initiatives (Commission launches AI innovation package to support Artificial Intelligence startups and SMEs, Press Release, 24 January 2024; Europe Fit for the Digital Age: Commission proposes new rules and actions for excellence and trust in Artificial Intelligence, Press Release, 21 April 2021); and

• “risk” was typically used in relation to data misuse, algorithmic bias, or patient safety (U.S.-EU Summit Joint Statement, 20 October 2023; Questions & Answers: AI Liability Directive, 28 September 2022; EU-US Trade and Technology Council Inaugural Joint Statement, 29 September 2021).

Taken together, these patterns suggest the recurrent co-occurrence of ethical terminology and governance-related vocabulary across the corpus. From an institutional perspective, this coupling reflects a strategy of legitimacy construction, whereby regulatory authority is bolstered with normative language. The term “public” was referenced frequently (247 times), in relation to public interest, public trust and public health. This recurring vocabulary aligns with the EU’s institutional logic of positioning AI governance as oriented toward collective welfare rather than purely market optimization.

Integration of computational and qualitative results

The Latent Dirichlet Allocation (LDA) topic modelling analysis, computed in R, systematically identifies clusters of semantically related terms, revealing dominant themes within the European Commission’s communications on AI in healthcare.

Figure 1 visualises the aggregation of 38 qualitative codes into four higher-order discourses. These discursive clusters reflect recurring thematic patterns identified across both computational and qualitative analyses (see Appendix B for examples of underlying codes).

Fig. 1.

Thematic extraction process through computational LDA. Source: our elaboration

To further validate the emerging themes, the results of the LDA topic modelling were cross-referenced with findings from the deductive qualitative coding conducted in NVivo®. While significant overlaps exist, the two methods provide unique insights that enhance the overall understanding of AI governance discourse.

The NVivo® analysis provided contextual depth, capturing explicit policy narratives, regulatory justifications, and ethical framings. The qualitative coding captures recurrent policy narratives related to trust, safety and European identity. Recurrent narratives of trust, safety, and European Identity illustrate how the Commission legitimizes technological innovation through moral and institutional principles.

Qualitative coding highlights explicit policy narratives and framing strategies, such as the EC’s repeated assertion that AI policies align with European values, democratic governance, and ethical responsibility. For example, manually coded themes emphasize narratives of trust, safety, and European Identity, that are instead not explicitly discernible from the LDA topic clusters.

The topic modelling method detects word associations that may not be immediately apparent in manual coding. For instance, Topic 6 in LDA links “business”, “service”, “biotechnology”, and “citizen”, suggesting a strong connection between AI adoption, the private sector, and public-facing services. Similarly, Topic 7 introduces “robotic”, “billion”, “private”, and “investment”, pointing to an economic dimension of AI.

The findings suggest that the EC’s discourse on AI revolves around:

1. Healthcare AI applications and data governance (confirmed by both NVivo and LDA Topic 1),

2. Regulatory risk frameworks and liability concerns (confirmed by NVivo and LDA Topic 9),

3. Europe’s strategic leadership in AI governance (confirmed by NVivo and LDA Topic 4),

4. AI’s intersection with business and investment (captured more explicitly by LDA Topics 6 and 7).

European Health Data Space (EHDS): obstacles vs benefits

This subsection contributes to addressing RQ2 and RQ3 by providing the results of the computational text analysis – i.e. word frequency and topic modelling – concerning the dominant lexical patterns in the European Commission’s communications on AI in healthcare. The results provide a basis for understanding the framing tendencies examined qualitatively in the following sections.

The comparison diagram between the “Challenges & Obstacles” and the “Expected Benefits” (Fig. 2) elucidates a stark dichotomy in the European Commission’s discourse regarding the European Health Data Space (EHDS) and broader digital healthcare initiatives. The analysis provides a concise visual comparison of the obstacles and benefits most frequently identified in the corpus, thus illustrating the EC’s balanced narrative. This narrative recognises systemic constraints yet frames digital health innovation as an ethical and social good. The prevalence of terms such as “health”, “data”, and “trust” underscores the Commission’s positioning of AI within a socio-technical paradigm that prioritises public value. This distinguishes EU rhetoric from market-optimisation models observed in other contexts [52, 62].

Fig. 2.

Challenges & Obstacles versus Expected Benefits of the European Health Data Space (EHDS). Source: our elaboration based on coded data from European Commission communications (2018–2025)

On the left side of the diagram, the coding references under “Challenges & Obstacles” indicate a range of structural and infrastructural issues that impede the full realisation of the potential of the EHDS. The recurrent themes of “complex obstacles” and the necessity for “further digitalisation at the national level” indicate that, despite a desire for cross-border health data exchange, many Member States still lack interoperable infrastructure. Moreover, the issue of restricted access to primary data for industrial use, coupled with the specific challenges associated with the collection, storage and utilisation of health data, reflect the ongoing difficulties encountered in the operationalisation of these systems.

In contrast, the “Expected Benefits” column presents the Commission’s forward-looking narrative. The EHDS is expected to result in significant economic savings, with projections indicating that €11 billion could be saved over a ten-year period through enhanced access to health data and innovations in healthcare delivery and research (Q&A on the European Health Data Space, 24 April 2024). The benefits include accelerated drug development through enhanced data sharing, the advancement of personalised medicine, and the development of more sophisticated telemedicine capabilities. Additionally, there is a pronounced emphasis on the empowerment of individuals through the control of their health data, the fostering of a “single market for digital health services”, and the improvement of research capabilities. The diagram summarises the prospective innovations in disease prevention, diagnosis, and treatment, along with the capacity to enhance healthcare efficiency and patient safety. It juxtaposes the aspirational outcomes of the EHDS against the real-world challenges that need to be addressed, thereby providing a clear illustration of the potential benefits and obstacles inherent in the implementation of this ambitious project. Although the projected benefits are considerable and have the potential to transform healthcare delivery in the EU, they are contingent on overcoming a number of entrenched obstacles related to infrastructure, privacy, and cross-border interoperability. The frequency and specificity of both the challenges and benefits coding indicate that the European Commission is actively engaged in addressing these issues while maintaining an optimistic outlook for the future of digital health in Europe.

This comparison serves as a visual representation of the EC’s balanced narrative, which acknowledges the significant hurdles while also emphasising the transformative potential of digital health innovations under the EHDS framework.

General statements on AI: issues and benefits

The present subsection draws upon the quantitative patterns identified in Sect. 4.1 to explore the coalescence of qualitative codes into broader thematic categories, indicative of the manner in which the Commission articulates AI-related priorities and values in its communications.

The diagram comparing the benefits and issues surrounding AI in various speeches and documents (Fig. 2) provides qualitative elaboration and evolution of the quantitative results, visually synthesizing the benefits and issues associated with AI as expressed across Commission communications. The following is a detailed commentary on the elaboration.

Issues: the left side of the diagram illustrates the pivotal issues and challenges inherent to the development of AI. These challenges are situated within the legal, ethical, and operational domains.

• Ethical Concerns and Bias: a considerable number of references indicate that AI presents a number of ethical challenges, including issues surrounding transparency, potential bias and discrimination. The potential for AI systems to replicate societal inequalities is a significant concern, as evidenced by the observation that women are less likely to receive job offers or that minorities are disproportionately targeted in predictive policing. These observations underscore the necessity for the formulation of ethical frameworks to mitigate the potential risks associated with biased outcomes of AI systems.

• Liability and Legal Uncertainty: the advent of AI has given rise to a number of crucial questions pertaining to liability, particularly in instances where AI systems have resulted in adverse outcomes. The lack of legal clarity for businesses, the disparate regulatory responses, and the opacity of AI algorithms are identified as significant obstacles to the adoption of AI. The intricate nature of AI systems renders it challenging to ascertain accountability, which can impede access to justice for those adversely affected by AI-driven harm, as evidenced by references to liability directives.

• Trust and Transparency: the lack of transparency in AI systems is a recurring theme, with speakers expressing concerns about the opacity of the decision-making processes employed by these systems. This phenomenon, which has been described as a “black box”, contributes to a broader deficit of trust among businesses and consumers. It is imperative that legal and regulatory measures are implemented to guarantee that AI systems operate in a transparent manner, thereby safeguarding fundamental rights and preventing unintended consequences.

• Social and Economic Risks: the potential impact of AI on employment is a significant concern, particularly the possibility of job losses due to automation. Small businesses often face challenges in implementing AI, while larger economic actors must navigate intense competition in the global market. These socio-economic challenges are crucial and require support and resources to facilitate a smooth transition to AI-driven systems.

Benefits: the diagram’s right side emphasises the numerous advantages of AI, particularly in domains such as healthcare, biomanufacturing, and broader societal applications. The principal themes are as follows.

• AI in healthcare: the presence of multiple coding references suggests that significant advancements in medical care are imminent, with AI set to improve diagnostics, personalised treatments and early disease detection. For instance, the capacity to analyse thousands of comparable medical images is highlighted as a notable advantage. One of the most notable advantages of AI is its capacity to assist medical professionals in making diagnoses in a more expeditious and precise manner.

• Economic and Industrial Gains: AI is regarded as a pivotal driver of innovation and industrial transformation. In their addresses, speakers have emphasised the capacity of AI to enhance operational efficiency in biomanufacturing, to inform financial decision-making, and to facilitate the development of new services in sectors such as agriculture and public administration. Furthermore, the economic impact of AI is evidenced by improvements in product and process optimisation across a range of sectors.

• Environmental and Social Benefits: AI is frequently cited as a means of addressing global challenges such as climate change, disaster forecasting, and cyberattacks. In addition to these, AI is also being deployed to manage health crises, optimise energy use, reduce traffic accidents, and enhance climate neutrality efforts. Furthermore, AI has the potential to assist marginalised communities, such as the elderly and disabled, and reduce inequalities in healthcare.

• Personal Empowerment and Data Use: the application of AI technologies has the potential to empower citizens in a number of ways. Firstly, it can facilitate greater control over one’s own health data. Secondly, it can foster cooperation across different Member States. Thirdly, it can enable new opportunities in the field of personalised medicine. Another notable theme is the focus on data altruism, whereby citizens choose to contribute their data to support healthcare innovations.

European Commission’s AI strategies alignment with broader European Identity (EI) politics

This section synthesises the previous findings by mapping dominant discursive patterns onto European identity politics. This interpretive synthesis links the textual evidence to the New Institutional Theory framework, demonstrating how AI policy discourse reflects the EU's broader normative identity.

The European Commission’s AI strategies demonstrate a clear alignment with broader European Identity (EI) politics, emphasising adherence to core EU values, democratic principles and stringent data protection standards. The coding references consistently demonstrate how AI initiatives are integrated within the EU’s ethical and legal frameworks, reflecting the Commission’s commitment to the preservation of European values while fostering innovation.

• The Commission reiterates on numerous occasions that AI development must adhere to the principles set forth by the EU (Daily News 24 January 2024, Press Release; Commission launches AI innovation package to support Artificial Intelligence startups and SMEs, Press Release, 24 January 2024). AI is presented as a means of bolstering European resilience and competitiveness while upholding the ethical standards that define the EU (Tech and Geopolitics: Building European Resilience in the Digital Age – Speech by Commissioner Thierry Breton, 5 September 2023). For instance, the EU-US Trade and Technology Council Inaugural Joint Statement (29 September 2021) emphasises the compatibility of AI with “shared values and fundamental freedoms”, guaranteeing that the development and deployment of AI respect human rights and democratic governance. This reflects the EU’s aspiration to spearhead global efforts in establishing a responsible AI ecosystem that echoes European values (President von der Leyen’s speech at the high-level opening session of the 2021 Digital Assembly: “Leading the Digital Decade”, 1 June 2021).

• The protection of data and the safeguarding of privacy appear to be of paramount importance. A significant aspect of the Commission’s AI strategy is the focus on data protection and privacy. A considerable number of documents attest to this, including Questions and Answers on the European Health Data Space (24 April 2024) and European Health Union: A European Health Data Space for People and Science (3 May 2022). Moreover, the commitment to the protection of personal data ensures that AI initiatives are aligned with the EU’s strict General Data Protection Regulation (GDPR), thereby embedding AI within a uniquely European approach to digital governance. This guarantees the maintenance of public trust while facilitating technological advancement (Daily News 24 January 2024; U.S.-EU Summit Joint Statement, 20 October 2023).

• The Commission’s AI strategies position the EU as a global leader in establishing ethical and regulatory standards for AI, thereby reinforcing European Identity politics. For example, President von der Leyen’s 2023 State of the Union Address (13 September 2023) states that “Europe has become the global pioneer of citizens’ rights in the digital world” illustrates how AI policies are positioned as extensions of European leadership in protecting citizen rights on a global scale. The references to the EU’s partnerships with regions such as Africa and Latin America in advancing health products (Commission takes action to boost biotechnology and biomanufacturing in the EU, Press Release, 20 March 2024) demonstrate the Commission’s aspiration to influence global AI governance through a framework of shared values and collaboration (U.S.-EU Summit Joint Statement, 20 October 2023). This approach aligns with the EU’s broader global role as a proponent of ethical and responsible technology.

• A recurrent theme is the balance between innovation, safety, and trust, which is central to the European Commission’s approach to AI. The phrase “in Europe, safety, trust, and innovation go hand in hand” (Artificial Intelligence: in Europe, Innovation and Safety Go Hand in Hand – Statement by Commissioner Thierry Breton, 18 June 2023) encapsulates the Commission’s vision of AI as a driver of economic and social progress that remains aligned with Europe’s foundational principles of trust and safety. The equilibrium between technological advancement and ethical responsibility represents a fundamental aspect of the European Identity. AI initiatives such as the European Cancer Imaging Initiative (23 January 2023) are explicitly linked to the European Strategy for Data (European Health Union: A European Health Data Space for People and Science, Press Release, 3 May 2022).

• The commitment to democratic values is a fundamental tenet of the Commission’s approach. The Commission’s strategies place an emphasis on the necessity for AI to operate within the bounds of “shared democratic values” (U.S.–EU Summit Joint Statement, 20 October 2023; EU-US Trade and Technology Council Inaugural Joint Statement, 29 September 2021). References to AI being developed in accordance with democratic principles reflect the broader alignment between AI strategies and the EU’s commitment to strengthening democracy both within Europe and in its international partnerships (President von der Leyen’s speech at the 2021 Digital Assembly, 1 June 2021).

A conceptual mapping was developed to further demonstrate how the EC’s AI strategies align with the fundamental aspects of European Identity politics. The primary thematic categories found in the qualitative coding (such as ethics, data protection, and democratic governance) are connected to the corresponding normative aspects of European identity, such as digital sovereignty, human rights, and participatory democracy, in this mapping, which is shown in Appendix B.

These discursive alignments position the EC as an actor that not only regulates AI but also performs European Identity through its policy language. This finding lends further support to extant literature on the EU’s “normative power” [41, 57], demonstrating that the Union’s approach to AI regulation simultaneously serves as an articulation of its unique political and ethical order. The findings indicate that AI policy communication functions as a mechanism for identity construction, in addition to its role in technological governance.

Discussion

From the perspective of institutional theory, the recommendation guidelines on the governance of AI issued by the European Commission (EC) portray how regulatory instruments, ethical vocabularies, and institutional values are discursively connected to shape the digital landscape of European governance (RQ3). Rather than treating regulation, ethics, and identity as separate domains, the Commission’s discourse presents them as mutually reinforcing dimensions of governance.

As outlined in the Findings section, these dynamics can be interpreted through the core mechanisms of New Institutional Theory. Healthcare is a key area that provides a critical testing ground for such institutional dynamics, given the high-risk application of technology, the ethical scrutiny, patient vulnerability, and public accountability.

The risk-based approach advocated by the EC can be interpreted through the lens of coercive isomorphism [12], insofar as formal regulatory requirements and compliance obligations are imposed on actors operating in high-risk sectors such as healthcare. The categorisation of AI systems according to levels of risk functions as a mechanism through which institutional pressure is translated into concrete expectations of conformity, not only in legal terms but also in ethical and procedural standards.

Simultaneously, mimetic isomorphism emerges through the adoption of regulatory sandboxes and testing facilities by the EC, mirroring policy instruments used in financial and data protection regulation, suggesting policy learning and emulation rather than regulatory novelty. These sandbox mechanisms enabled controlled experimentation and knowledge exchange while maintaining public oversight, reflecting the EU’s preference for experimentation within boundaries rather than deregulation.

On the other hand, normative isomorphism is exemplified by the Commission’s efforts towards ensuring the interoperability and cross-border data sharing, particularly through initiatives like the European Health Data Space (EHDS). The European Commission (EC) has adopted a strategy of repeated references to shared standards, professional norms, and ethical principles with a view to promoting the diffusion of AI governance practices across Member States. This will promote the diffusion of the AI standards within the Member States and will solidify the collective commitment to transparency, accountability and trust.

The qualitative coding analysis confirms that indeed the EU’s framing of the governance of AI is very much embedded within the mainstream discourse of digital ethics. Consistent with previous research, “trustworthy AI” functions as both a regulative instrument and a justificatory narrative [21, 54]. Importantly, AI is not framed as an autonomous or disruptive technological sphere, but as a domain that must remain compatible with pre-existing European legal, moral, and democratic frameworks. The constant emphasis on safety, fairness, and the accountability of democracy signals a logic of justification through shared values, in that AI must remain subordinate to the institutional norms that sustain European legitimacy.

In this regard, the EU’s discourse aligns with the idea of Normative Power Europe [41], understood here not as an empirical claim about global influence, but as a discursive strategy through which governance is performed and communicated. Regulation thus becomes a performative act that projects the moral and ethical standards as a distinct global identity. While scholars have appropriately raised questions regarding the coherence and consistency of this normative self-image [57], its persistence in AI policy discourse highlights its ongoing significance in framing emerging technologies as public-value innovations rather than purely market-driven products.

More broadly, the Commission’s discourse embeds the development of AI in a sociotechnical imaginary [80] that focuses on the progress of technology in the realm of democracy, citizens’ trust, and egalitarianism – core features of the EU’s identity politics in the digital era.

The European Health Data Space is one such example of this institutionalization where AI is integrated into a legal-ethical structure with inclusivity, data solidarity and public empowerment at its core. This initiative represents the Commission’s value-driven approach, institutionalizing digital ethics as both governance logic and identity practice. This strengthens the EU identity as a normative entrepreneur who leads through ethics and morals instead of competing for liberalization.

Nevertheless, the analysis also reveals important tensions and limitations. These stem from fragmentation between Member States, from a lack of ethical-legal enforcement, and from algorithmic biases. The reliance on a risk-based logic, while ensuring procedural accountability, risks reinforcing bureaucratic rigidity that may stifle innovation. In this sense, while the EC’s AI policy discourse institutionalizes trustworthy AI as a legitimizing narrative, its effectiveness ultimately depends on whether these values translate into measurable social and regulatory outcomes.

In sum, the findings indicate that the EU’s institutional strategy is both normatively ambitious and structurally constrained. Through coercive, mimetic, and normative mechanisms, the EC seeks to align AI innovation with European democratic ideals – simultaneously pursuing governance legitimacy and identity construction. This ambidextrous role of the EU highlights how AI regulation operates not only as a technical policy domain but as a form of values-based politics, through which Europe negotiates its moral and institutional coherence in the digital era.

Conclusions and limitations

The present study demonstrates that European Commission communications on AI concurrently undertake governance and identity functions: they regulate technological innovation whilst constructing a narrative of European moral leadership. The analysis demonstrates that the Commission’s discourse on AI is structured around four interrelated discourses, aligned with the EU’s broader normative project: ethical and legal regulation, innovation and competitiveness, data governance and trust, and European leadership and identity.

While the research provides a comprehensive analysis of European Commission communications on AI, it is important to acknowledge the methodological limitations of the study. First, the corpus was limited to publicly available communications, which may not have captured internal deliberations or informal policymaking dynamics. Second, although the combination of computational and manual coding enhances validity, the hybrid approach potential methodological bias: topic modelling results depend on algorithmic parameterisation (e.g., number of topics and pre-processing steps), while qualitative coding remains subject to interpretive judgement despite intercoder checks. Third, the selection of English-language documents may underrepresent multilingual nuances in EU discourse or national-level rhetorical variations. Finally, the temporal scope of the study (2018–2025) reflects the evolution of AI policy discourse, but not its implementation outcomes. The methodological constraints thus delineate the boundaries of inference, while concomitantly suggesting directions for future comparative and longitudinal studies. Future work might extend the dataset to encompass multilingual and cross-institutional examples, employ a different set of computational methods (e.g., topic modeling via BERT), and examine cross-regional comparisons that can determine the applicability of a discursive framing of the EU. These findings underscore the need for continuous dialogue between policymakers, clinicians, and ethicists to ensure that AI governance frameworks remain both evidence-based and ethically grounded.

Authors’ contributions

S.C. wrote the main manuscript text and conducted the analyses; E.I. and C.N. supervised the research and reviewed the manuscript.

Funding

This research did not receive funding.

Data availability

The datasets generated and analysed during the current study, together with the R scripts used for preprocessing and topic modelling, are available in the Zenodo repository at https://doi.org/10.5281/zenodo.18385144. All analysed documents are publicly available from the European Commission Press Corner website.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.