2.1 Principlism and AI ethics’ scoping problem

The widespread implementation of AI in biomedical and health-related research is recognized as creating a range of ethical challenges, including algorithmic fairness and decision-making autonomy. A core aspect of AI ethics work is identifying the scope and applicability of ethical principles. However, ethical frameworks developed primarily in the clinical research context are often a poor fit for the development and implementation of AI-based tools. The blurred line of development and implementation in AI research practices obscures the subject matter of ethical guidelines and frameworks (Doerr and Meeder Reference Doerr and Meeder2022; Ferretti et al. Reference Ferretti, Ienca, Hurst and Vayena2020). First, these frameworks are predicated on a clear distinction between “research,” which aims to create new, generalizable knowledge typically relevant to populations, and “clinical practice,” which seeks to benefit specific individuals. These goals determine the prioritization of ethical principles and who has ethical responsibility. For AI applications in biomedicine, this distinction is difficult to maintain, rendering the application of ethical frameworks inappropriate or poorly matched at best, especially because research ethics analysis is largely limited to weighing risks and benefits of a particular intervention to individuals enrolled in specific studies (Bereza Reference Bereza2017). Furthermore, neither standard research ethics nor clinical ethics frameworks account for broad societal harms such as those that have already arisen from AI (Ferretti et al. Reference Ferretti, Ienca, Hurst and Vayena2020). In addition, the dependence on identifiability in the regulatory definition of human subjects in the US means that AI research is largely not considered human subjects research and thus exempt from that aspect of regulatory oversight. For example, current regulatory schemes designed to evaluate drugs, devices and even biological systems struggle to assess black-boxed and dynamic AI algorithms, especially machine learning-based models, whose performance is designed to change in response to new data but are inscrutable because they do not provide a rationale for their output (Mittelstadt and Floridi Reference Mittelstadt and Floridi2016).

Most importantly, the requirement for widely distributed networks of actors to create and analyze the vast datasets necessary for AI (Leonelli Reference Leonelli2016), or data ecosystems (NIH Office of Data Science Strategy 2022), means that responsibility is also widely distributed or difficult to assign. Although Leonelli (Reference Leonelli2016) suggests that data scientists attach ethical responsibilities according to the division of labor in the data ecosystem, standard bioethics frameworks do not account for distributed responsibilities (Herington and Cho Reference Herington and Cho2025). Furthermore, these frameworks also fail to account for the complexity of interest holders and decisions in biomedical research that have ethical implications (Shim et al. Reference Shim, McMahon, Saco, Bentz, Foti and Lee2025). For healthcare AI in particular, a host of different ethical issues are raised over the course of the translational process that involve a range of interest holders, including data donors, data collectors, data analysts, AI model developers, clinicians or patient users, AI system purchasers and implementers. The temporal and spatial diversity of the data ecosystem presents many different possible scopes for which relevant responsibilities and ethical issues can be identified.

The wider societal impact of AI, which is not generally anticipated during the development stage, poses further challenges around who is accountable for AI ethics deliberation and at what stage ethical interventions might be required. Ethical inquiries need to extend beyond immediate harms or instances of unprofessional conduct or “human subjects protections” and may require social science expertise to consider broader societal consequences that are often unforeseen during the developmental stages. Ethical reviews by the Institutional Review Boards (IRBs), with their focus on respect for individuals, are explicitly limited in their capacity to consider potential harms to groups or broader societal impacts (Reardon et al. Reference Reardon, Lee, Goering, Fullerton, Cho, Panofsky and Hammonds2023) – precisely the types of risks most relevant to AI-related research. De-identified, electronic data and automated algorithms are usually not considered the subject of IRB oversight (Metcalf and Crawford Reference Metcalf and Crawford2016). In case of underrepresented data, biased data proxies or discriminatory categories, it is not clear who is responsible to anticipate, scale and mitigate the downstream harm, which makes interventions in the development stage difficult (Lazar and Nelson Reference Lazar and Nelson2023; Metcalf et al. Reference Metcalf, Moss, Watkins, Singh and Elish2021). Responsibility for data bias is especially difficult to locate for healthcare AI that uses data collected from electronic health records, which are notorious for capturing preexisting societal biases that become reified in algorithms (Gerke et al. Reference Gerke, Minssen and Cohen2020). The current informed consent process is inadequate in protecting patients’ autonomy and privacy in cases where data are delinked, aggregated and harmonized (Emam et al. Reference Emam, Khaled, Arbuckle and Malin2011; Mittelstadt and Floridi Reference Mittelstadt and Floridi2016), which becomes an imminent challenge in the era of platformized health data. Nevertheless, for data-centric research such as AI research, the temptation to focus on ethical issues in obtaining data allows these concerns to overshadow the implications of data use or downstream effects.

In the professional fields of engineering and computer science, despite the fact that responsible AI and AI ethics have become a cottage industry, many of the ethical endeavors have been either limited to the issuance of relatively non-specific codes of ethics (Munn Reference Munn2023), with limited effect, or restricted to making technical fixes to address narrow ethical issues, such as debiasing or fine-tuning algorithms and models. More importantly, critics point out that the current practices of ethical AI are “bounded” as they fall short of addressing historical injustices and marginalization, which have been deeply embedded in the systematic design and data collection, and will disproportionately affect certain communities without meaningful inclusion (Creary Reference Creary2021; El-Azab and Nong Reference El-Azab and Nong2023; Ferryman Reference Ferryman2023; Ferryman et al. Reference Ferryman, Cesare, Creary and Nsoesie2024; McCradden et al. Reference McCradden, Joshi, Mazwi and Anderson2020). Only relying on a reductionist, technologist and Western-centric perspective might overlook the lived experiences and societal complexities of those affected by emerging AI systems (Amoore Reference Amoore2023).

2.2 Understand the multiplicity and heterogeneity of the scoping problem

Understanding how ethics are connected to the complex, real-world development and implementation practices in a data ecosystem requires deeper attention to how “ethics” are defined, interpreted, operationalized and institutionalized across different domains. Values and ethics are not fixed, preexisting entities. In organizational and institutional contexts where multiple stakeholders intersect, the scope of ethics is not singular and clear, but rather dynamic, heterogeneous, shaped and stabilized during interactions and negotiations around technological and material artefacts. The field of Science and Technology Studies (STS) provides useful conceptual tools to unpack and understand the multiplicity, heterogeneity and dynamics in scoping ethics. Influenced by Actor-Network Theories and Praxeography (see Latour Reference Latour1996; Mol Reference Mol2008; Reference Mol2003; Mol and Berg Reference Mol and Berg1994; Schatzki et al. Reference Schatzki, Knorr-cetina and von Savigny2001), STS scholars have investigated the multiplicity of practices during biomedical knowledge-making, which lead to various framing strategies among various agencies, and such knowledge reflects not only material existence but also relational networks and resources that researchers enlist, engage and enroll in (Crabu Reference Crabu2018; Hollin Reference Hollin2017; Wilson‐Kovacs and Hauskeller Reference Wilson‐Kovacs and Hauskeller2012). For example, Crabu (Reference Crabu2016) examined the laboratory arrangements performed by scientists and clinical oncologists to render biological materials and knowledge immediately “usable” in clinics. Similarly, AI and data-driven systems produce a multitude of societal actions regarding what AI predictions mean in clinics and how the numbers are situated in management systems, which require negotiation, repair and reconciliation (Elish Reference Elish2018; Elish and Watkins Reference Elish and Watkins2020; Hoeyer Reference Hoeyer2019). AI ethics, in this context, are ontologically disparate regarding goals, measurements and outcomes (Wehrens et al. Reference Wehrens, Stevens, Kostenzer, Weggelaar and de Bont2023) and need to be developed in line with the emergent power of AI and algorithms that mobilize a network of agents and decision-making processes (Ananny Reference Ananny2016). A relational perspective is particularly key in understanding the constitution of ethics, as Pols (Reference Pols2015) has shown, in which the patients’ perception of ethics and care are shaped through technological interactions and clinical relationships. The ethics of AI in biomedicine and healthcare require a dynamic, heterogeneous and grounded point of view that examines sociotechnical relationships and phenomena to contextualize discussions of ethical principles in practice.

An embedded model, in which scientists work alongside ethicists and other social scientists to identify and resolve ethical challenges, has been put forward as a viable option. Embeddedness ideally allows ethical deliberation through iterative processes between ethicists and scientists, and responding to emerging issues (McLennan et al. Reference McLennan, Fiske, Tigard, Müller, Haddadin and Buyx2022). The interaction and relationship, therefore, become a point of inquiry to understand how ethics are conceived and practiced within multidisciplinary spaces.

2.3 Embeddedness: interdisciplinary promise and limitation

The US National Institutes of Health (NIH) ELSI research program was originally developed in tandem with the Human Genome Project by the National Center for Human Genome Research (now the National Human Genome Research Institute, or NHGRI) (Thomson et al. Reference Thomson, Boyer and Meslin1997). Initially implemented as supporting independent streams of genetic and ELSI research, NHGRI more recently has sought to integrate social sciences and humanities into large genome research consortia, to examine societal and ethical implications, regulation and governance, public accountability and scientific responsibility for genetic research. It has supported a community of scholars who address socially sensitive issues related to research of the human genome, and later neuroscience research (Balmer et al. Reference Balmer, Calvert, Marris, Molyneux-Hodgson, Frow, Kearnes, Bulpin, Schyfter, MacKenzie and Martin2015, Reference Balmer, Calvert, Marris, Molyneux-Hodgson, Frow, Kearnes, Bulpin, Schyfter, Mackenzie and Martin2016; Karabin et al. Reference Karabin, Adsit-morris, Lee, Fullerton, Cho and Reardon2024). The idea of embeddedness indicates a high level of interdisciplinary collaboration and cross-institutional, transdisciplinary communications and ELSI mandate and requirements have been brought into state-funded, large-scale programs such as the BRAIN initiative and, more recently, the Bridge2AI project funded by the National Institute of Health. In these projects, this interdisciplinary collaboration was not just aspirational, but a requirement of funding and a way to support science as to create policy tools for widen science’s public good (Adsit-Morris et al. Reference Adsit-Morris, Collins, Goering, Karabin, Lee and Reardon2023; Juengst Reference Juengst1996). Although the ELSI model is particular to the US context and applied mostly in the publicly funded project, similar models could be seen elsewhere and were referred to as ELSA in Europe.

Day-to-day work through interdisciplinary collaboration is essential for addressing emerging ethical issues. However, interdisciplinarity alone does not guarantee effective incorporation of ethics in ELSI initiatives. Critics have argued that the ELSI program was not designed to transform science in ways that would address and mitigate potential harms and inequities. Rather, its primary function was to protect and support the expansion of “big science.” Those initiatives often incorporate justice and ethical concerns in limited and marginal ways. The goal of ELSI, they contend, has been to develop policy tools that would facilitate the growth of genomic research, rather than to ensure that the public equitably benefits from the substantial investments in genomics (Adsit-Morris et al. Reference Adsit-Morris, Collins, Goering, Karabin, Lee and Reardon2023; Juengst Reference Juengst1996; Reardon et al. Reference Reardon, Lee, Goering, Fullerton, Cho, Panofsky and Hammonds2023). The prioritization of science and the biomedical research enterprise results in tensions and marginalization of ELSI scholarship and justice and equity goals (Conley et al. Reference Conley, Prince, Davis, Cadigan and Lazaro-munoz2020; Martschenko et al. Reference Martschenko, Grannuci and Cho2024). For example, the diversity goals of precision medicine are usually framed around recruiting participants, not equity and inclusion of partners and communities (Jeske et al. Reference Jeske, Vasquez, Fullerton, Saperstein, Bentz, Foti, Shim and Lee2022; Vasquez et al. Reference Vasquez, Foti, McMahon, Jeske, Bentz, Fullerton, Shim and Lee2023). In practice, scholars with expertise in relevant domains are typically brought on board too late in the process to provide meaningful input into decisions about the categories that frame research and the questions and aims that guide it (Reardon et al. Reference Reardon, Lee, Goering, Fullerton, Cho, Panofsky and Hammonds2023). However, including ELSI investigators while preserving their autonomy, objectivity and intellectual independence also presents ongoing challenges (McEwen et al. Reference McEwen, Boyer, Sun, Rothenberg, Lockhart and Guyer2014).

The scope of ethical and societal issues has been influenced by federal research funders. For example, in recent years, the NIH directed attention at equity issues, addressing injustices in biomedical research ranging from underrepresentation of gender, racial and ethnic minorities in clinical studies to lack of diversity in the scientific workforce. As a result, the NIH instituted a number of comprehensive, large-scale initiatives such as the UNITE Program to address structural racism in the biomedical research enterprise (Collins et al. Reference Collins, Adams, Aklin, Archer, Bernard, Boone, Burklow, Evans, Jackson, Johnson and Lorsch2021). Individual research projects had additional requirements to ensure workforce diversity, equitable research participation and inclusion of diverse perspectives in research. For some of the large consortia with requirements for ethics expertise, there were also requirements for subgroups to address workforce issues specifically. However, as shown in the empirical findings in later sections of this paper, it was also not clear who was responsible for addressing diversity, equity or inclusion mandates in these consortia.

Such institutionalized and formalized ways of integration and governance are also challenged by political uncertainties, and concepts linked with social justice can be subject to changing ideologies. In 2025, the United States, the Trump administration denounced the Diversity, Equity and Inclusion (DEI) initiatives in education and employment (some required by NIH policy developed under previous administrations), terminating grants that were focused on social justice issues such as health equity or workforce diversity, which included a number of AI ethics initiatives (National Institute of Health 2025). The administration also rescinded the previous administrations’ AI safety initiatives (The White House 2025), resulting in uncertainty and ambiguity in setting AI governance goals, especially in publicly funded projects, in terms of how the governmental agencies determine their ethical standard in response to public accountability. Elsewhere, institutionalized assessment methods, such as Responsible Research and Innovation and anticipatory governance models, indicate a need for governance and public accountability; still, those approaches are subject to criticisms such as methodology, vulnerability to politicization and misuse of public participation (Casiraghi Reference Casiraghi2023; Shanley Reference Shanley2021). There is a lack of consensus and scattered international efforts, leaving a patchwork of models and ambiguities regarding jurisdiction and impact.

2.4 Vertical versus horizontal: is embeddedness the answer?

Principalist approaches to ethics, along with related governance frameworks, fall short in addressing the complex, systemic and heterogeneous nature of AI. Treating ethics as the application of abstract principles through a top-down regulatory framework and focusing primarily on relationships and responsibilities between individuals has been increasingly critiqued for its limitations in AI, particularly its lack of clear articulation of problems, mechanisms for accountability and lack of power (Chi et al. Reference Chi, Lurie and Mulligan2021; Mittelstadt Reference Mittelstadt2019; Rességuier and Rodrigues Reference Rességuier and Rodrigues2020; Shneiderman Reference Shneiderman2020). Similarly, previous ELSI models also fall short in responding to broader social issues such as public trust (Juengst Reference Juengst1996; Reardon et al. Reference Reardon, Lee, Goering, Fullerton, Cho, Panofsky and Hammonds2023) and fail to represent the interests of the marginalized communities (Lee et al. Reference Lee2008). The top-down, principle-based approaches correspond to what is termed as “vertical” (Helm and Gerlek Reference Helm and Gerlek2026 in this issue). In contrast, a horizontal approach seeks to engage with the broader system and lifecycle of AI technologies, including its structural and historical injustices and downstream consequences. Such an approach advocates for a contextualized and situated understanding of ethical principles, grounded in everyday practices (Keane Reference Keane2016; Pols Reference Pols2024). It emphasizes the sociotechnical network in which AI is embedded and aims at co-framing ethical considerations among stakeholders, ensuring that responsibilities for ethical design, data practices and implementation are fairly and effectively distributed across the AI lifecycle. Therefore, constructing a horizontal approach requires critical reflection not only on how ethics are understood and conceptualized, but also on where (and when) ethical deliberation occurs and who holds the authority to enact it, especially how the “embeddedness” is solving the aforementioned challenges in the decision-making. The calls for embedded ethical practice and the lack of institutional, conceptual and operational clarity still remain a consistent challenge. From the concept of ELSI and embeddedness to the project designs, institutional arrangements, on-the-ground practices and relationship-building, there has not been a standard or a consensus, and reports from embedded experts have made substantial critiques and reflections on their experience (e.g. Balmer et al. Reference Balmer, Calvert, Marris, Molyneux-Hodgson, Frow, Kearnes, Bulpin, Schyfter, MacKenzie and Martin2015). Empirical investigations, therefore, are needed in answering these key questions: Where are ethics teams located within research and development processes? What mechanisms ensure that ethical objectives are incorporated into scientific goals and not subordinated to them, or subject to conflicts of interest? In multidisciplinary teams, who has the expertise – and power – to define what counts as “ethical AI”? Who is held accountable for ethical outcomes throughout development and deployment? Finally, does the embeddedness of ethics promote deeper engagement with systemic and societal concerns that may lie beyond the immediate scope of a given scientific and technical project?

Our empirical study adopts a sociotechnical and grounded approach to examining the embedded ethics and the role of the ethicists. We focus on the differences and barriers in the context of large research consortia, and how organizational and institutional goals influence the framing and scoping of ethics work. We also focus on the embeddedness of value and the process value-(dis)alignment during interdisciplinary collaborations, and the facilitators and barriers of such processes. Such approaches will inform the “crux” of interdisciplinary collaboration in the era of AI and how to tackle it – building teams, sharing power and aligning values during negotiation and co-production of ethics, responding to both technological evolvement and historical injustice. Importantly, our empirical study examines these dynamics not just in framework or principles (vertical) but in practice (horizontal): how are ethics goals framed and scoped within large AI research consortia? Where does the vertical logic of top-down governance create friction with the horizontal goals of embedded ethics? What structural and relational conditions enable or undermine meaningful ethics integration?

4.1 Techno-fix tendency and disciplinary barriers

We have observed a significant discrepancy between definitions of ethics and the scope of ethics work between different disciplines, and major barriers in collaboration and communication. Without clear guidance, ethics teams struggled with data scientists over what these commitments actually meant. As one ethicist (AI-D) noted: “ethical sourced data is a concept they (the funding agency) put in there. But nobody knows what it means.” The scope and definitions of fairness or trustworthiness remained very general and detached from the research context, creating significant ambiguities about what constitutes ethical work in AI development.

Amid uncertainty and lack of clarity of what the scope of ethics work, our findings reveal that AI’s unique characteristics create distinct barriers to ethics integration that manifest through undefined requirements, competing frameworks and systematic disciplinary marginalization. Unlike more regulated biomedical research, AI development occurs within a context that privileges technical solutions over interdisciplinary collaboration. AI’s engineering-dominated culture creates a significant barrier to ethics integration through “techno-fix,” or the systematic assumption that ethical problems can be, and should be, solved through technical measures. Social and political complexity of ethics then becomes “reduced” to questions of engineering and design, which limits the scope of ethical work and fails to respond to further societal or justice concerns (Leese et al. Reference Leese, Lidén and Nikolova2019; Sætra and Selinger Reference Sætra and Selinger2024; Selbst et al. Reference Selbst, Boyd, Friedler, Venkatasubramanian and Vertesi2019). This culture also treats ethical issues as technical problems requiring computational solutions, systematically excluding non-technical disciplines from meaningful participation in defining and addressing ethical challenges.

One interviewee (AI-M) explained how this technical framing narrows ethical consideration:

This techno-fix tendency manifested in disputes over the scope of ethics work. One interviewee (AI-C) complained that some data scientists claimed “they’re doing ethics,” “but in reality what they’re really doing is just saying how do I debias an algorithm.” In contrast, a bioethics expert (AI-K) argued for examining “the way the data were collected, who had a say and how the data were collected, how representative the data [are], and their disparate impacts downstream,” which are considerations that the technical fix approach systematically overlooks.

Part of the techno-fix tendency stems from the problem- and technique-oriented engineering culture that ignores human factors. It creates novel ethical tensions through competing frameworks that operate simultaneously within the projects’ collaborative space, generating critical differences. The engineering culture in the AI projects treated data as nothing but numbers and AI as tools, making ethical conversations particularly contentious. For ethicists who previously worked with clinicians, the lack of focus on human factors in the “data people” was difficult to work with. The bioethicist (AI-K) noted that,

Another interviewee (AI-J) highlighted the conflict between “different ethics”:

The open-source ethos of making data freely available for technological advancement directly conflicts with the emphasis on privacy protection and patient safety in healthcare ethics. This tension is particularly acute because the same datasets serve both open-source AI development and clinical applications, creating competing ethical demands. Interviewees expressed the intention of looking for a “middle ground.” However, these competing constitutions of practices operate with fundamentally incompatible accountability mechanisms, stakeholder networks and success metrics, making consensus-building extremely difficult. These conceptual differences created significant communication challenges that reinforced disciplinary isolation. The communication barriers were exacerbated by mutual gaps in expertise. As one interviewee (AI-C) described:

Multidisciplinary experts capable of bridging these conversations were rare, and the techno-fix tendency meant that ethical expertise was systematically devalued. Computer scientists and data scientists typically lack exposure to human factors and bioethics training, and they “don’t necessarily see ethical issues” (AI-M) as relevant to their work. One interviewee (AI-P) described resistance when “ethics wanted to be part of those conversations” about tool development related to privacy, but “data collection [team] had their lane and they didn’t think that ethics should be in their lane.” This territorial defensiveness emerges because data work is perceived as purely technical rather than involving human considerations despite its ultimate human impacts. An interviewee (AI-A) noted that,

These disciplinary tensions resulted in the systematic marginalization of ethics expertise. Ethicists reported feeling “unintentionally disrespected by people who just wanted to keep doing data generation” (AI-Q). They struggled to establish credibility within AI teams and recognized that gaining respect required “ELSI folks demonstrating our value to the scientific folks” (AI-K) – a challenging and time-intensive process.

Critically, this marginalization was structural rather than interpersonal. The research programs were organized in ways that systematically sidelined ethics work, creating barriers to the horizontal integration that embedded ethics approaches require. As the interviewees noted, ethics work was “by design” undermined by program structures that prioritized technical milestones over ethics integration.

4.2 Widgeti-zation

These disciplinary tensions become operationalized through specific funding mechanisms and organizational structures that systematically marginalized ethics work. Despite rhetoric about embedded ethics in the funding documents, the actual implementation treated ethics as what we term “widget-ization” or a modular add-on rather than an integrated component of research design.

The contrast between ambiguous ethical goals and clear technical objectives became operationalized through funding structures that prioritized data collection milestones. Both funding agencies and research consortia asserted clear and pressing scientific goals to generate AI-ready data and shareable databases. Specific funding mechanisms tied data collection milestones directly to continued funding, creating immense performance pressure on principal investigators (PIs) who were accountable for these deliverables.

This “centrality of ‘science goals’” systematically sidelined ethics considerations, reinforcing the techno-fix identified in the previous section. The data collection and dissemination components, referred to by ethicists as the “science part,” reported directly to administration and funding agencies, gaining significantly more institutional traction than ethics work.

Meanwhile, funds for ethics teams were frequently significantly delayed so that teams could not even be formed, and ethics proposals were regularly rejected.

Despite project rhetoric about bringing diverse disciplines and stakeholders together, execution remained highly centralized, particularly in the distribution of grant funding. The decision-making structure was hierarchical, as one interviewee (AI-B) explained: “[The PI’s University] was the ‘prime’ and everybody else was a ‘sub’.” This hierarchy enabled PIs to unilaterally override interdisciplinary considerations, operationalizing the power imbalances implicit in the techno-fix tendency. The same interviewee described how collaborative input was dismissed:

Rather than being embedded and integrated as promised in funding calls, ethics work was treated as a modular component in “widget-ization.” This means positioning ethics teams as replaceable resources without embedded deliberation or grounded empirical research integration. It also sees aerial expertise as of less value than fulfilling certain functional objectives or “checkboxes” such as drafting legal documents. As one interviewee (AI-A) described their role: “we are available as a resource. And we’re called upon as a resource.” Widget-ization manifested through activities like hosting office hours and organizing seminars, positioning ethics teams as external consultants rather than research collaborators. The utility of ethics teams depended entirely on how they were “called upon” and how technical teams understood their value. One ethicist (AI-L) noted that during office hours, “rarely anybody came in with ethical questions, but technical ones.” Importantly, interviewees did not critique consultation models per se, recognizing their legitimate role in responsible research practices. Their frustration stemmed from the gap between promised “paradigm changing” approaches to ethically sourced data and the reality that their scholarship on how to bring paradigm change was downplayed.

The widget-ization was compounded by team restructuring processes that treated ethics expertise as transferable between projects without consideration for specialized ethics knowledge, preexisting collaborative relationships or contextual knowledge. A funding agency program officer explained that some original ethics teams “scored low” in grant reviews, leading to their replacement with “high scoring ethics teams from the unfunded proposals” who were added to other funded projects. This restructuring process scrapped the originally developed proposals and research plans. One interviewee (AI-D) expressed the frustration, summarizing what was relayed to them: “The [funding agency] said we don’t like your ethics work. So we expect you to get rid of the team that you planned for ethics; and to integrate this new team, that you never heard of, into your already developed project.” Another interviewee (AI-K) explained: “the scope of work that our team came in with based on the expertise of the folks who were involved was no longer going to happen.” New ethics teams inherited predetermined scope and goals from project leadership, with little to no impact on the research design itself.

The restructuring created “plug-and-play” ethics, which reflected a perception held by both funding agencies and project team members that ethics work could be seamlessly transferred between projects without consideration for the embedded scholarship that effective ethics integration requires.

The structural marginalization created critical timing problems that undermined ethics integration. Ethics work was consistently “brought on too late” in project timelines, contradicting embedded ethics principles that require early integration into research design. One team attempted to establish a community advisory council for input on data collection. As an interviewee (AI-B) reported,

Interviewees consistently reported insufficient time and effort allocated for teams to build relationships before beginning technical work. In reorganized projects where new ethics teams were added through the “plug-and-play” restructuring process, trust levels remained particularly low. Several interviewees described perceived hostility toward ethics teams, in which ethics work was viewed as a barrier rather than an asset.

Even in some more successfully integrated cases, ethics teams required “constant negotiation of deliverables to justify their work.” Much ethics work fell outside formal deliverable lists, particularly when addressing emerging issues such as raising and discussing data privacy concerns across the consortium before the data is released. An interviewee (AI-D) said,

Some interviewees attributed the widgeti-zation to a lack of history. Unlike more mature fields such as ELSI in genomics, interviewees reported that the field of AI has not grown to recognize the value of ethics work. Most ethics teams were recruited through existing relationships, but many of these were “inherited” from genomics research and required additional effort to adapt to AI contexts. One interviewee (AI-K) reflected on this transition:

These findings reveal that the hierarchical, “vertical” project structure and widget-ization of ethics work fundamentally limited how ethics could be scoped and operationalized. The gap between embedded ethics rhetoric and modular implementation demonstrates that current organizational approaches systematically undermine the collaborative scholarship that embedded ethics requires. The systematic nature of this marginalization through funding mechanisms, hierarchical decision-making, team restructuring and timing misalignments shows that techno-fix becomes institutionalized through specific organizational structures.