The article maps the contemporary AI ethics landscape by shifting the organizing question from what moral concerns are discussed to how those concerns are made sense of. It argues that many overviews implicitly treat AI ethics as a field mainly about principles and guidelines, but that framing hides the real diversity of normative “sense-making” already present in the literature. To make this diversity navigable, the authors propose a tripartite structure and explicitly present it as a practical tool for orientation in a fast-expanding, hard-to-survey corpus.
The core aim is structural: to provide “some structure” to an “astounding breadth” of work on moral concerns raised by AI by identifying three approaches that function like lenses for ethical interpretation. The article’s purpose is therefore not to offer yet another list of concerns, nor another inventory of application areas, nor a periodization of “waves,” but to classify recurring moral concerns by the explanatory style and normative grammar used to interpret them. The same concern, the authors stress, can look fundamentally different depending on the lens used, and this difference changes what counts as an adequate response.
Methodologically, the paper is a narrative synthesis designed to detect patterns in how scholars formulate moral concerns. The literature base is built primarily via snowball sampling, starting from an initial search in databases such as Web of Science, Google Scholar, and PhilPapers for existing AI ethics overviews, then expanding with targeted searches using keywords that track moral concern framing as well as specific normative traditions. Inclusion criteria require that works explicitly or implicitly address moral concerns, use a moral framework, or critique AI systems, including critiques of dominant principle-first approaches. The review focuses on English-language peer-reviewed articles, conference papers, and reputable policy documents and explicitly notes that this introduces a Euro-US emphasis and possible bias in what lenses emerge as dominant.
The first lens, the principle-based approach, is described as the predominant framing in AI ethics. It interprets moral concerns through ethical principles assumed to be universal, stable, and portable across contexts, with a strong precedence of theory over practice: principles are articulated conceptually and then meant to be implemented in real systems. Within this approach, the paper distinguishes subcategories where principles are formulated as rules and duties, or as values that can be embedded into design. The promise here is scalability: principles can become guidelines, translate into governance, and travel across domains without requiring deep local immersion each time.
Yet the article’s key move is to show that this principle-based dominance is also a bottleneck. The paper summarizes two main limitations that follow from abstraction. First is a principles–practice gap: principles are detached from local context and far removed from sites of implementation, risking “checklist ethics,” where ethics becomes the evaluation of predefined items rather than a rigorous engagement with what AI actually does in situated use. Second is a systemic blind spot: by staying at the level of general principles, the approach often under-attends to the cultural, political, and economic contexts in which AI is developed and deployed, including structural inequalities and power relations. The critique is not that principles are useless, but that they can become thin, frictionless, and easy to “comply with” performatively, while missing the lived and structural conditions that generate harms in the first place.
The second lens, lived realities, re-centers ethics on human–AI relations as they unfold in practice. Instead of treating values as fixed “out there,” this approach treats them as interactive, dynamic, and potentially changing through real encounters with technologies. The emphasis is typically micro-level: local practices, everyday experiences, and context-sensitive meaning-making. The article links this orientation to traditions such as (post-)phenomenology, virtue ethics, and Science and Technology Studies, and it highlights that the methods often include ethnographic, empirical, and hermeneutic approaches. The point is to make ethics sensitive to how AI actually mediates perception, agency, care, work, and social interaction rather than assuming an abstract user or a generic deployment environment.
A particularly important contribution of the lived-realities lens, as framed here, is the concept of value change: interactions with AI can reshape what people count as fair, responsible, trustworthy, or even humanly flourishing. The article treats this not as rhetorical flourish but as an analytic stance with consequences: if values can shift in practice, then ethical design and governance cannot rely only on pre-specified value lists; they need feedback loops that notice transformation in norms and expectations as systems are adopted. This complicates many compliance-driven models of “ethical AI” that treat ethics as requirements engineering.
At the same time, the article does not romanticize micro-level richness. It highlights limitations that follow from local specificity. First, findings can be hard to generalize across contexts because the empirical site is often narrow by design. Second, the close focus on individual or local experience can obscure the macro-structures that shape why AI is built the way it is and which options are even available to users. The lived-realities lens is strong at showing how people interact with AI, but it can be weaker at translating those insights into broad prescriptions or addressing the political economy that constrains design choices upstream.
The third lens, power structures, interprets moral concerns through domination, exploitation, discrimination, and social justice, and demands attention to the sociopolitical and economic contexts of AI development and deployment. It operates primarily at the macro level and is associated with traditions such as political philosophy and critical theory. In this approach, “ethical problems” are not merely side-effects to be mitigated but expressions of structural arrangements: who owns infrastructures, who sets incentives, who is surveilled, who is categorized, who is rendered governable, and who bears externalities. This lens is also presented as a corrective to what the paper identifies as under-addressed systemic questions in principle-first framings.
Within the power-structures approach, the article points to subareas such as surveillance, bias and discrimination, and political economy. A notable analytic distinction it draws is between treating fairness and discrimination as formalizable technical properties versus treating them as historically situated and institutionally reproduced realities. From this lens, bias is not just an error in data or modeling; it can be an artifact of historical and institutional injustices that AI systems reproduce at scale. This matters because it changes what “fixing bias” means: it shifts attention from tweaking models to interrogating the social conditions that generate biased datasets and asymmetrical impacts.
The authors also specify limitations of a power-only optic. First, power can become reductive, flattening other dynamics into a single explanatory category. Second, the lens can drift toward deterministic readings that treat people as passive victims and underplay agency, creativity, resistance, refusal, and appropriation, which lived-realities research tends to document more carefully. Third, the lens can generate incisive critique but less actionable guidance, producing what the paper quotes as “rich food for thought” yet insufficient pragmatism for regulation and concrete intervention. In short, the lens is necessary for structural clarity, but it risks becoming normatively loud and operationally quiet if it cannot connect critique to implementable change.
The article’s synthesis becomes most useful when it treats the three lenses as complementary rather than mutually exclusive. Table 1 is the compact operational summary: principles focus on universal fixed principles and rely mainly on conceptual methods; lived realities focus on relationality, human flourishing, and local practice with ethnographic and hermeneutic methods; power structures focus on domination, discrimination, exploitation, and social justice with macro-level conceptual and empirical approaches. Each lens has characteristic limitations: principles can become predefined and detached from practice; lived realities can be subjective and hard to generalize; power structures can become deterministic and critique-heavy in ways that inhibit constructive responses.
A practical way to read the paper is as a diagnostic for why AI ethics debates so often talk past each other. When one group argues for clearer principles, another insists that everyday experience is where ethics actually happens, and a third says the entire discourse is compromised unless it confronts structural injustice, they are not necessarily disagreeing about the moral stakes. They may be disagreeing about what counts as an explanation and what level of analysis is ethically primary. The paper turns that confusion into an explicit map: it makes the hidden assumptions in each framing visible and thereby allows more disciplined disagreement, or better, structured collaboration.
This mapping also has an applied payoff across sectors. Because AI is rapidly integrated into domains like healthcare, education, and public policy, the authors position their structure as a bridge for comparing AI ethics with other applied ethics areas and for surfacing tensions and complementarities in how moral concerns are framed. If an organization adopts only a principle checklist, it may miss micro-level harms that emerge in practice and macro-level harms rooted in power. If it adopts only ethnographic sensitivity, it may struggle to scale governance. If it adopts only structural critique, it may struggle to specify what to do on Monday morning besides writing sharper essays. The paper’s contribution is to make these trade-offs explicit so that responsible AI efforts can assemble the lenses rather than picking one as an identity badge.
Mini dictionary of key concepts follows, written as short definitional paragraphs aligned with how the article uses them (Groen et al., 2026).
AI ethics is the interdisciplinary field that studies, critiques, and guides the moral concerns raised by AI systems, including issues such as privacy, surveillance, discrimination, and the conditions for human flourishing in AI-augmented societies (Groen et al., 2026).
Principle-based approach is a way of framing AI’s moral concerns through universal, stable, and fixed ethical principles that can be articulated conceptually and then applied across contexts, often producing guidelines, governance frameworks, and design requirements (Groen et al., 2026).
Lived realities approach is a way of framing AI’s moral concerns by foregrounding human–AI relations in local practices and everyday experience, treating values as interactive and dynamic and often relying on ethnographic, empirical, and hermeneutic methods to understand what AI does in context (Groen et al., 2026).
Power structures approach is a way of framing AI’s moral concerns by centering domination, discrimination, exploitation, and social justice, and by analyzing the sociopolitical and economic structures that shape AI development and adoption at macro level (Groen et al., 2026).
Technological mediation is the idea that technologies do not merely impact people externally but actively shape perception, practices, and moral experience, including the possibility that interactions with AI contribute to value change rather than simply triggering predefined ethical principles (Groen et al., 2026).
Human flourishing in this article’s framing is the ethical focus on how human lives can go well in relation with AI systems, not just how harms can be avoided, and it appears as a characteristic concern within the lived-realities lens that studies relationality between humans and technology (Groen et al., 2026).
Checklist ethics is the risk that ethics becomes a procedural tick-box exercise evaluating predefined concerns while leaving principles insufficiently scrutinized and inadequately connected to local context, practice, and the sociotechnical realities of implementation (Groen et al., 2026).
References: Groen, E. M., Sharon, T., & Becker, M. (2026). An overview of AI ethics: Moral concerns through the lens of principles, lived realities and power structures. AI and Ethics, 6, Article 121. https://doi.org/10.1007/s43681-025-00955-7