2.1. AI Ethics as a historically vertical approach

AI Ethics has largely developed along what we call a vertical ethics orientation: efforts to articulate general normative frameworks for AI through guidelines, principles and vocabularies that can be adapted and used across domains and institutions (Jobin et al. Reference Jobin, Ienca and Vayena2019). Early work in AI Ethics sought to respond to the opacity and complexity problem of machine-learning systems by translating established ethical resources into the AI domain. Luciano Floridi and colleagues, for example, proposed an influential framework for a “good AI society” that foregrounds explainability, accountability and human-centric values (Floridi et al. Reference Floridi, Cowls, Beltrametti, Chatila, Chazerand, Dignum, Luetge, Madelin, Pagallo, Rossi, Schafer, Valcke and Vayena2018; Taddeo and Floridi Reference Taddeo and Floridi2018). This early focus has been foundational in shaping AI Ethics and can be seen as a contribution to a good-faith strand within vertical AI Ethics: attempts to establish shared reference points capable of informing public debate, regulatory efforts or institutional governance.

A wide range of vertical orientations has since developed around value- and rights-oriented conceptual toolkits, focusing on privacy, transparency, explainability, fairness, nondiscrimination and autonomy (Nyholm Reference Nyholm2022). These orientations have been essential for rendering AI-related harms visible and addressable (Morley et al. Reference Morley, Kinsey, Elhalal, Garcia, Ziosi and Floridi2023). At the same time, vertical orientations tend to operate at a high level of abstraction: values appear as stable reference points that are articulated independently of the concrete practices through which AI systems are designed, trained and deployed, while their meaning and practical consequences may remain underdetermined (Mittelstadt Reference Mittelstadt2019). This does not mean to say that AI Ethics is dominated by principlism in the strict bioethical sense associated with the so-called “four-principles framework” (Beauchamp and Childress Reference Beauchamp and Childress2019). Rather, our concern is broader: vertical AI Ethics often relies on portable normative abstractions whose legitimacy stems less from consensus about their content than from their institutional legitimacy. Eventually, this orientation tends to position norms and values as external reference points, or in other words: ethical reflection is situated outside of sociotechnical practices.

This reliance on abstraction becomes problematic when vertical orientations are expected to guide action in contexts marked by conflicting normative pluralism or asymmetrical power relations (Bleher and Braun Reference Bleher and Braun2023; Clouser and Gert Reference Clouser and Gert1990). Privacy and fairness, for example, are neither self-evident nor univocally interpreted. Their significance and meaning depend on social position, institutional setting or domain (Bridges Reference Bridges2017). What counts as the “best” ethical interpretation may, in practice, be determined more by technical feasibility, economic viability or political expediency than deliberative agreement (Siffels and Sharon Reference Siffels and Sharon2024). AI Ethics can thus provide a shared language of contestation, but it cannot by itself ensure that this language is translated into accountable practice.

2.2. Corporate co-optation: a stress-test for vertical ethics

Corporate environments function as a stress test for these limitations. Major technology companies publicly endorse values such as fairness, sustainability and explainability (Hagendorff Reference Hagendorff2020) while simultaneously being criticized for practices that contradict these commitments, including unauthorized user data extraction (Zuboff Reference Zuboff2019), biased systems (Buolamwini and Gebru Reference Buolamwini and Gebru2018), industry serving benchmarks (Eriksson et al. Reference Eriksson, Purificato and Noroozian2025) and exploitative data labor (Miceli and Posada Reference Miceli and Posada2022). This pattern reflects what Sarah Ahmed calls “non-performativity”: declarations of ethical commitment operate as substitutes for substantive change (Ahmed Reference Ahmed2016). The problem is not that vertical ethics is per se reducible to “ethics washing,” but that abstract value concepts are comparatively easy to appropriate while robust enforcement and accountability mechanisms remain missing.

Institutionalization efforts such as the GDPR, the EU’s Trustworthy AI agenda and the UNESCO AI Ethics recommendations have contributed to embedding ethics into governance infrastructures, yet their practical effects remain uneven and highly dependent on implementation capacities (Goodman and Flaxman Reference Goodman and Flaxman2015). In parallel, a growing “economy of virtues” around AI has turned grassroots efforts and ethical expertise into a strategic resource, circulating between academia, industry and policy (Phan et al. Reference Phan, Goldenfein, Mann and Kuch2022).

Vertical orientations have proven indispensable for establishing AI as a matter of ethical and critical concern. Yet they have also been limited when faced with the complex power structures embedded in AI deployment. We therefore argue that it is precisely the vertical orientation that has contributed to making AI Ethics not just powerful but, in many respects, counterproductive to basic aims like preventing harm or advancing the common good. This is because the vertical character leaves AI Ethics particularly susceptible to co-optation by industry players for adversarial purposes. By relying on a narrow set of external abstract principles, AI Ethics limits its ability to respond dynamically to the situational entanglements through which values are being enacted in various AI–human constellations. This is how the hype around AI Ethics, the tech industry’s strategic influence and the absence of robust enforcement have brought to light systematic limitations of the vertical orientation, motivating the emergence of horizontal, practice-oriented approaches.

2.3. Horizontal orientations and their limitations

Horizontal approaches trace their origins to broader critiques of the political dimensions inherent in technological artifacts and seek to embed ethical reflection within the practices of design and deployment. Technologies are not ahistorical, instead, they are situated within culturally and politically contingent regimes of knowledge (Akrich Reference Akrich1992). This also applies to AI systems (Tacheva and Ramasubramanian Reference Tacheva and Ramasubramanian2023). From this perspective, AI does not merely mediate action but enacts specific epistemic assumptions, many of which are contested (Gerlek and Weydner-Volkmann Reference Gerlek and Weydner-Volkmann2025). Treating AI as shaped by and shaping culture (Seaver Reference Seaver2017), horizontal scholarship highlights how biases embedded in computer science become materialized in technical systems, while simultaneously amplifying existing epistemological regimes of dominance. One manifestation is the deeply rooted Anglocentrism that inheres in both the architectures of large language models and their training data (Bella et al. Reference Bella, Helm, Koch and Giunchiglia2024).

Frameworks such as Value Sensitive Design have, early on, advanced the idea of embedding ethical commitments throughout the development process (Friedman Reference Friedman1997; Knobel and Bowker Reference Knobel and Bowker2011; van de Poel Reference van de Poel, Michelfelder, McCarthy and Goldberg2013). These proactive approaches aim to align innovation with societal values during early and midstream development (Fisher et al. Reference Fisher, O’Rourke, Evans, Kennedy, Gorman and Seager2015), and they now thrive on the fertile ground of contemporary AI hype and doom discourse. Approaches such as Explainable AI, fairness measures, dataset debiasing, value alignment and jailbreaking for AI safety represent attempts to bridge principle and practice. As such, they form part of a horizontal ethics that seeks to operate from within technological development rather than from above (in a vertical orientation). Still, this raises key questions: who is included in this “within”? who gets to decide? and whose voices remain excluded?

Despite being more implementation-oriented than vertical orientations, many of these frameworks continue to rely on the prior articulation of societal values that are subsequently translated into design requirements. In this sense, they do not abandon vertical norm articulation but relocate it into processes of technological development. As a result, much of this work remains shaped by a design discourse that defines ethical problems in technically solvable terms, despite acknowledging that technologies can reshape the moral situation in which they operate (Akrich Reference Akrich1992; Morozov Reference Morozov2014; Verbeek Reference Verbeek2006). This is not to say that technological solutionism is bad per se, but it tends to mistake symptoms for causes (Helm and Hagendorff Reference Helm and Hagendorff2021; Siffels and Sharon Reference Siffels and Sharon2024).

The so-called Moral Machine experiment exemplifies another approach that, while appearing horizontal, still engages in reductive framing. Designed to explore people’s intuitive judgments on moral dilemmas involving autonomous vehicles (Awad et al. Reference Awad, Dsouza, Kim, Schulz, Henrich, Shariff, Bonnefon and Rahwan2018), it used a global online survey to gather input on whom autonomous cars should prioritize in fatal scenarios. Participants chose between categories such as pedestrians versus passengers, young versus old or humans versus animals. Though the study revealed cultural variation and underscored the value of diverse input, its methodology remained top-down: expert-led, based on constrained moral logics, and disconnected from everyday experiences that are shaped by power asymmetries and practical concerns and constraints.

As AI Ethics becomes increasingly institutionalized, integrated into what Phan et al. describe as “economies of virtue” (Phan et al. Reference Phan, Goldenfein, Mann and Kuch2022) and shaped by governments and corporations, approaches centered on abstract guidelines, risk mitigation, thought-experiments and design-level values prove insufficient. While horizontal scholarship moves ethical reflection closer to the practices than vertical approaches, systematic efforts to integrate research that investigates how ethical life unfolds in context remain scarce. What is needed, therefore, is not the abandonment of existing approaches but their empirical grounding: an orientation that examines how values are articulated, contested and enacted within the social, cultural, political and historical conditions from which technologies such as AI emerge and in which they operate and evolve.

3.1 Foundations: from thought experiments to empirical grounding

Emerging in the late 1990s, philosophy’s Empirical Turn describes a dynamic shift away from the dystopian critiques associated with thinkers such as Heidegger and Anders and toward scholars such as Don Ihde and Hans Achterhuis, who foregrounded technology’s entanglement with everyday practices and institutions (Achterhuis Reference Achterhuis1992). This shift parallels a wider Practice Turn in the humanities and social sciences, which treats routines and material engagements as key sites where knowledge and norms are produced (Cetina et al. Reference Cetina, Schatzki and von Savigny2005) and where practice is treated as central for understanding how bodies, technologies, meanings and norms are co-constituted (Bourdieu Reference Bourdieu1977; Butler Reference Butler1993). STS, then, extends these moves by converging in three core insights: first, phenomena are simultaneously epistemological and ontological, constituted through actor-networks (Callon Reference Callon1984); second, agency is an effect of relational practice (Callon and Law Reference Callon and Law1997); and third, symbolic and technical mediations coproduce phenomena in situated ways. Reality is inseparable from observation and intervention (Barad Reference Barad2007).

These insights also relocate AI Ethics: from the application of principles to bounded systems toward the sociotechnical processes through which AI becomes a way of knowing and governing. Louise Amoore’s emphasis on opacity and partial account-giving shifts ethical inquiry away from transparency demands and toward the practices through which algorithmic arrangements justify, distribute and deflect responsibility (Reference Amoore2020). On this basis, empirical AI Ethics can be defined as the study of ethical life in AI assemblages. Lucy Suchman’s critique of AI’s “uncontroversial thingness” adds a temporal dimension: ethical debate often begins after categories have stabilized and responsibility has already been allocated across sociotechnical chains (Reference Suchman2023). This relegates AI Ethics to a power-conserving risk-management enterprise.

What is needed are approaches that move inquiry upstream. Ethnographic work on machine learning as a mode of experimentation on thought and conduct and critical analyses of how evaluation regimes function as infrastructures of power are exemplary here (Luitse et al. Reference Luitse, Blanke and Poell2025, Mackenzie Reference Mackenzie2017). Investigations in the routines and realities of data work add another pointer to where an upstream empirical ethics must go, and what it must follow. It directs attention to the mundane but consequential sites where “AI” is assembled as a working system: the instruction sets, interfaces, taxonomies, quality checks and payment structures through which data are produced, cleaned, labeled and rendered compatible with model development (Miceli et al. Reference Miceli, Posada and Yang2021). These approaches make visible how ethical stakes are built into the field at the point of specification: in decisions about what counts as an admissible target, how “error” is rendered legible and acceptable, and how competitions, leaderboards and platform access distribute incentives, credibility and authority.

Such qualitative inquiry also clarifies why harms that are often discussed as “bias” or “misclassification” can be better understood as failures of sociotechnical recognition. Credit scoring provides a clear case. Systems trained on formal financial traces frequently misread informal practices in the Global South as indicators of risk or corruption, even when they reflect creditworthiness as it is locally produced through relations of solidarity and care (Heeks and Swain Reference Heeks and Swain2018). An empirical ethics approach does not simply register that an algorithm performs poorly on a given population. It empirically examines how model categories, data proxies and validation practices define what counts as legitimate economic behavior, and how these definitions voyage across contexts while presenting themselves as generalizable.

Close-up empirical studies stand in sharp contrast to sanitized thought experiments, such as the trolley-problem framing deployed in the Moral Machine experiment, which abstracts away the power-differentiated realities through which AI systems are specified, trained, validated, distributed and adopted. By making early and implicit judgments visible and by attending to the creative and unexpected ways in which technology is shaped, empirical ethics research does not merely add context to assessment. It supplies a foundational orientation as a practice-based, power-attentive inquiry into how ethical registers are materially assembled through situated practice.

3.2 Relations: from principles to orders of worth

The empirical approaches described above may raise questions about the status of normativity. Are principles to be understood as entirely relative to practice? The answer is no: principles do not become obsolete, but their status shifts. Judgment no longer derives from a standpoint of external reasoning; it is articulated and contested in relation to communities of practice.

This analytic shift positions us well to describe what, following Boltanski and Thevenot (Reference Boltanski and Thevenot2006), can be called “orders of worth”: diverse justificatory logics that vary across communities of practice, domains and even individuals. Rather than following either a linear vertical logic or a linear horizontal logic, their framework can be used as a source of inspiration to study how actors articulate, defend and enact values in practice and how such valuations become stabilized, challenged or overturned.

Attending to orders of worth helps make sense of recurrent conflicts between evaluative preferences: an industrial order may elevate efficiency and cost–benefit rationales, while a civic order may privilege equity and collective welfare. Taking a step back to identify the relevant and potentially conflicting orders of worth in a given field generates a context-sensitive understanding that reliance on principles (vertical), or on prestructured participatory designs (horizontal), often bypasses.

Close-up observation of ethical life may also bring into view the dominance of particular orders of worth imposed upon a community that does not fully share it. This can produce a profound desire for transformation, which becomes visible, for instance, when studying the frustrations of nurses working under increasingly technocratic regimes where efficiency is granted overriding force, yet in practice prevents them from enacting what they understand as “good care” (Pols Reference Pols2015).

Furthermore, studying diverging orders of worth can bring into view the damaging consequences of vertical approaches that are too often power-preserving, as shown, for example, in the context of criminal justice. Canonical by now is the case of decision-support algorithms trained on historical crime data to estimate the likelihood of reoffending, thereby shaping parole decisions. In such high-stakes settings, it is vital that systems do not reproduce discriminatory patterns. Yet studies have shown that widely used fairness metrics often fail to capture intersectional inequalities (Barocas et al. Reference Barocas, Hardt and Narayanan2023). Rather than mitigating injustice, they can inadvertently stabilize existing hierarchies. This becomes visible in a case where they have proven insufficient to prevent disproportionately harsh treatment of defendants with darker skin tones (Lum and Isaac Reference Lum and Isaac2016). Notably, qualitative reporting from interviews also points to forms of irritation and unrest not only among those who felt unfairly targeted but also among those ostensibly advantaged by the system (Angwin and Larson Reference Angwin and Larson2015).

An empirically grounded AI Ethics that begins with the systematic observation of practices, perspectives and consequences may challenge the dominance of certain orders of worth, confronting them with others that exist, albeit at the margins of power. For instance, systems previously evaluated through utilitarian or procedural notions of justice as fairness might be reassessed in light of historically situated conceptions such as epistemic justice (Helm et al. Reference Helm, Bella, Koch and Giunchiglia2024). A reorientation can also involve extending domain-specific values into new territories, such as the application of care, traditionally confined to reproductive labor and health care, to fields like law enforcement (Asaro Reference Asaro2019). Care ethics, with its relational understanding of ethical life, resonates well with the project of empirical AI Ethics. As Joan Tronto famously stated:

Care ethics centers on experience and the moral urgency of attending to the needs of others, especially those rendered vulnerable or marginalized. By placing repair at its core, care ethics also foregrounds the historical embeddedness of ethical life, recognizing that before questions of procedural justice can be meaningfully posed, there is often a demand to redress the accumulated weight of past and ongoing harm.

What, then, would it mean to evaluate an AI system in criminal justice not through the logic of distributive fairness, commonly quantified via the proportional distribution of false positives and negatives, but through the logic of care? Such a shift would reorient our focus: from abstract metrics to historically accumulated vulnerabilities. It would prioritize the repair of social worlds, in which trust in the criminal justice system has, for many, been eroded by experiences of mistreatment, often based on markers of social difference. This erosion of trust is not incidental; it feeds into cycles of recidivism and violence, perpetuating the very forms of marginalization that AI systems purport to neutrally assess. Contrary to the orders of worth underlying distributive fairness, from a care ethics perspective, addressing algorithmic bias is not a matter of tuning models or refining datasets to level the playing field. It becomes a demand to transform the social and historical conditions that produce inequality in the first place. Here, the primary order is to engage in the situated work of repair.

3.3 Scales: from micro–macro to assemblages of concern

Where empirical investigations into varied orders of worth reveal a demand for change, it may be fruitful to seek guidance from contexts that are most different. A plurality of conceptions of ethical life flourishes beyond the confines of dominant Western principles of fairness, accountability and transparency – the go-to principles of vertical ethics. Amerindian notions of Sumak Kawsay (buen vivir), for instance, foreground the entanglement of human and more-than-human actors by emphasizing conditions of living well in reciprocal relation with others and with the land (Acosta Reference Acosta2013). This orientation shares with the orders of care a rejection of anthropocentric and dualistic models of agency. Instead, both assume an infinite regress of relationality, where no actor, human or otherwise, exists in isolation. When applied to the domain of AI, such perspectives invite an ethical gaze that foregrounds the impact of technological systems on more-than-human well-being. Within this worldview, any AI innovation that causes environmental harm cannot plausibly be considered ethically desirable, rendering efficiency-based cost–benefit analyses as logically and morally untenable, even when short-term individual gains accrue (Taffel et al. Reference Taffel, Bedford, Mann, Phan, Goldenfein, Mann and Kuch2022).

At the same time, these perspectives bring a tension into view: between situated inquiries into ethical life and the scale at which generalist AI models operate. Generalist models come with massive infrastructural apparatuses that can be socially and materially destructive. This matters for empirical ethics because such systems are typically justified through efficiency-oriented orders of worth (speed, throughput, optimization). The tension is that what counts as “efficient” or “successful” at scale is often achieved by shifting costs onto local communities of practice, but the sources of harms remain invisible and, hence, incontestable when inquiry focuses only on local context.

While it is essential to cultivate contextual sensitivity, we must also grapple with the undeniable need for coordinated responses to the global implications of AI. The deployment of energy-hungry models across sectors necessitates the establishment of regulatory processes aimed explicitly at safeguarding the flourishing of plural forms of life (Hecht Reference Hecht2018). How can we find modes to address the relationality of various scales, while not reducing sociotechnical complexity to quantifiable threats that translate into mere risk-mitigation measures or overly abstract frameworks that do not correspond to the situated reality of ethical life?

In a world of entangled infrastructures, planetary exhaustion and governing AI models, regulation demands more than policy; it requires onto-political awareness. The idea of publics as coming into being, theorized by Bruno Latour, and its methodological operationalization by Noortje Marres are instructive in this context (Reference Marres, Castelle, Gobbo, Poletti and Tripp2024). Publics are not preexisting entities but emerge through controversy, breakdown and shared matters of concern (Latour and Weibel Reference Latour and Weibel2005), ultimately building assemblages of concern. Technologies such as AI, however, do not merely mediate communication or disseminate information; they participate in the assembly of publics by shaping what is rendered visible, debatable or ignored. If we assume that publics play a role in shaping regulation and development, and that AI is simultaneously co-constituting the very publics from which consensus is meant to emerge, we risk a recursive paradox: the goat becomes its own gardener. How to escape this dilemma?

What is needed are cross-scale interventions, which make us recognize how ethical life can find its expressions through assemblages: situated and emerging formations that link locality to wider networks of responsibility and meaning. A concrete methodological approach to studying such multi-scalar assemblages lies in what Huub Dijstelbloem calls infrastructural investigations, which focus on the practices by which journalists, activists, artists and organizations contest diffuse forms of power (Reference Dijstelbloem2021). Using techniques such as mapping, counter-databases, interactive visualizations and testimonials, these actors reconfigure relations of visibility and accountability. They counter forms of organized irresponsibility, evident in domains such as border politics, journalism and supply chains, where harm is systemic yet dispersed, and difficult to attribute, a situation exacerbated by the opacity of AI black box technologies.

Whether instantiated through crude AI quality benchmarks (Raji et al. Reference Raji, Bender, Paullada, Denton and Hanna2021), algorithmic news recommenders (Helberger Reference Helberger2021) or habit-forming interfaces (Ash et al. Reference Ash, Anderson, Gordon and Langley2018), exposing sociotechnical vehicles as forms of power operating within broad ecosystems of influence challenges linear notions of causality and, by extension, vertical ethics and top-down governance. Making these dynamics legible and contestable as assemblages of concern can complement vertical ethics with a relational ethics of cross-scalar co-responsibility. A direction for the formation of such assemblages of concern as an ethics of co-responsibility comes from the notion of pluriversality, which is often referenced in tandem with the abovementioned buen vivir.

Inspired by the Mexican Zapatista movement, pluriversality invites us to envision “a world where many worlds fit.” Rooted in decolonial and post-development thought, pluriversality creates space for diverse ways of knowing, being and doing (Escobar Reference Escobar2018). Embedded within it is not only a sharp critique of the universalizing tendencies of vertical ethics but also of the well-meaning horizontal orientations of co-creation, which can reproduce colonial patterns by exporting Euro-American values under the banner of participation (Helm et al. Reference Helm, de Götzen, Cernuzzi, Hume, Diwakar, Ruiz Correa and Gatica-Perez2023). In contrast, pluriversality supports autonomous designs of sociotechnical assemblages: forms of thinking, knowing, making and doing that emerge from within communities of practice but respond to concerns that can only be understood by acknowledging interdependencies of ecosystems and infrastructures. With their planetary perspectives, yet concern for local autonomy, notions of pluriversality and buen vivir offer normative orientations for cross-scalar intervention. The idea is to reframe agency not as possession, but as circulation and to practice ethics not as distant assessment but as situated making and doing (Ratto Reference Ratto2011). In the following, we highlight strategies for the enactment of this idea.

3.4 Transformation: from risk mitigation to technologies of hope

To emancipate itself from its role as a whitewasher, rather than being driven by a relentless pace of innovation where damage prevention dictates the agenda, AI Ethics could contribute to manifesting alternatives. Just as the historical development of STS evolved from critique into an engaged program (Sismondo Reference Sismondo, Felt, Fouché, Miller and Smith-Doerr2017), Ethics should not exhaust itself in abstract assessment. Such an orientation entails a conscious ethical choice not to dwell in despair but to engage in the generative power of concrete hope.

Amidst empirical contextualization, certain foundational commitments remain that most would recognize: if harm or destruction is observed, there is an imperative to respond. Yet before action can emerge, there must first be a sensation, a representation and an arousal that catalyzes the will to act, and from this, hope is born (Willow Reference Willow2023). Not the dulled optimism that numbs into inaction (Berlant Reference Berlant2011) but a hope that energizes. The kind that Ernst Bloch (still haunted by the cruelties of the Second World War) termed concrete hope (Reference Bloch1985). Conceived as something tied to the concrete, hope is not a projection onto distant or ill-defined futures. It emerges from engagement in feasible alternatives, which respond to local concerns but are open to scale. Concrete hope acknowledges conditions of constraint and powerlessness without equating improbability with impossibility (Han-Pile Reference Han-Pile2017).

This orientation resonates with Appadurai’s distinction between an “ethics of probability” and an “ethics of possibility” (Appadurai Reference Appadurai2013, pp. 188, 295). The former limits us to the logic of what is most likely, a logic that binds us to focus on the visions of those who are in power. The latter, in contrast, opens spaces for imagination and expands our collective capacity to aspire. Empirical AI Ethics can move beyond the well-worn terrain of probability by opening itself to pluriversal figurations of technology.

An example comes from a project in which members of the Guarani-Kaiowá and Ikpeng-Xingu, Indigenous peoples of the Upper Amazon, were invited to imagine technologies they desired. Many participants responded with paintings of mythical beings: animals with human traits. At first, the facilitators assumed the prompt had been misunderstood. But the images were intentional. Among them was the Nyra Nhe’Éngatu, a radiant hybrid of bird and human, drawn from Ikpeng cosmology. This creature bridges earth and sky – guardian of the unseen, responsible for the care of communication itself. It is said to visit those in need of sending urgent messages, especially in moments of injury or danger. One 12-year-old boy painted this being to represent what he believed a smartphone should do for his people (Bonami et al. Reference Bonami, Helm, van Voorst and da Silva2026).

Unlike Western AI benchmarks, which are typically governed by anthropocentric ideals or operational efficiency, the Nyra Nhe’Éngatu functions as an onto-epistemic opening (Cadena Reference Cadena2015): an orienting force that resonates closely with ideas of care ethics while diverging fundamentally from anthropocentric frameworks of explainability or fairness. Through collaboration with local communities of practice, an empirical ethics approach has the potential to challenge dominant Silicon Valley narratives that valorize human-likeness as primary indicators of successful AI innovation (Brown et al. Reference Brown, Mann, Ryder, Larochelle, Ranzato, Hadsell, Balcan and Lin2020). This dominant model presumes the inevitability of centralized, resource-intensive and infrastructure-heavy AI systems. The children of the Guarani-Kaiowá and Ikpeng-Xingu, in contrast, offer a radically different vision for AI. It is imagined not to replicate human capacities for convenience, nor driven by an obsession for explainability, but to assist in tasks that humans are inherently unable to perform, all while remaining within deeply humane registers of worth: supporting and caring for those in need.

By emphasizing possibilities rather than probabilities, we move away from mere reactionary responses to technological advances, risks and harms. An empirical ethics that seeks to be transformative could focus on elevating the hopes, desires and aspirations of otherwise marginalized and unheard people. This helps to move away from a mere critical perspective on AI as a “naughty child” needing discipline. And indeed, across the globe, numerous initiatives illustrate how machine learning can be harnessed to empower marginalized communities, for restorative justice, to challenge oppressive systems and foster domain-specific values. In the following, we provide two examples of such attempts to enact orders of worth through the development and adjustment of technical tools.

The Live Language Initiative serves as an example of a technology of hope that strives to elevate marginalized voices. This project functions as a database, catalogue and collaborative platform, surpassing existing language databases not in volume but in its diversity. Its name, Live Language, reflects the idea that languages are alive; they are born, evolve and die, along with the people who speak them (Koch et al. Reference Koch, Bella, Helm and Giunchiglia2024). While most language databases focus on the 40 most resourced languages, thereby pivoting around English, Live Language strives to document the linguistic diversity of the remaining 6,960 languages and dialects spoken today (Skirgård et al. Reference Skirgård, Haynie, Blasi, Hammarström, Collins, Latarche, Lesage, Weber, Witzlack-Makarevich, Passmore, Chira, Maurits, Dinnage, Dunn, Reesink, Singer, Bowern, Epps and Hill2023). The vision is to collaborate with marginalized speakers around the world to create a pluriversal alternative to centralized models, fostering the development of tailored alternatives: a world where many languages fit.

Another noteworthy initiative focuses on efforts to enact domain-specific orders of worth, thereby countering the top-down tendencies associated with the algorithmization of the public sphere (Aradau and Blanke Reference Aradau and Blanke2022). This approach resonates with the earlier discussion of the need to assess the enactment of values differently across distinct domains, where diverse communities of practice may prioritize different orders of worth. Cultural citizenship, for instance, has been established as a key order within the journalism sector. Ferraro et al. (Reference Ferraro, Ferreira, Diaz and Born2024) therefore prototypically enact four concrete steps for the redesign of recommender algorithms for music curation to reflect this order: (1) moving beyond a commercial orientation, (2) recognizing the cumulative influence of recommendations, (3) integrating a metric of commonality that measures how well recommendations familiarize a given listener population with specific content categories and (4) avoiding a top-down normative approach by complementing it with experiments that assess the effectiveness of commonality in achieving the target value. This reconfiguration of algorithmic recommendation in the music domain is a prime example of enacting an ethics of possibility to create technologies of hope.