1. Introduction
“Really, I no longer want to share our research data, nor our community resources. […] I have the feeling, everything we make open, there come all the AI companies and scrape it, train their models [on it], and reap the benefits – while we get nothing in return.” (27.3)
This computational sociolinguist’s concern captures a growing tension in open science. Once a committed advocate, she now resists sharing in light of generative AI’s expanding reach. What began for her as a broad mission within her community to democratize knowledge production has, in her words, morphed into an “opaque one-way exchange,” where she feels disconnected from how her work is used and fears it may contribute to “some technology of social control or discrimination” (27.4). This disillusionment reflects a broader shift among several of my interview partners, who have long embraced open science, but who now question whether their openness is being appropriated to serve interests and purposes they oppose.
These are not isolated voices. A geographer reflected: “We used to believe that transparency and data sharing were key to advancing science and knowledge production for good, but now it feels like we are giving away our most valuable asset without enough recognition, and not enough protection.” (6.6) A political scientist noted: “Sharing data used to be about collaboration, but now I worry it opens too many doors for exploitation, or even misuse. […] It has gotten out of control.” (16.8)
This paper emerges from 27 in-depth interviews with computational social scientists, data scientists and data managers in Europe, conducted between 2022 and 2024 as part of a larger project on the politics of openness in research practice (Mayer Reference Mayer2020). While generative AI was not the initial focus, it quickly surfaced in the interviews as a key concern. With the mainstreaming of large-scale AI models, interviewees increasingly voiced unease that the data and infrastructures they shared in good faith were being scraped, decontextualized and monetized by corporate actors, often in ways misaligned with their ethical commitments. As I will show, the issue, for many, is not merely loss of control or recognition, but the fear that openness – once understood as a means to advance the public good – has become a mechanism of enclosure, feeding proprietary AI systems and reinforcing asymmetries of power. Openness in this context is neither neutral nor universally good; it is a contested political space, shaped by institutional mandates, infrastructural dependencies and conflicting moral claims.
This study interrogates these developments through a situated and critical lens, asking: How does generative AI reshape the normative ideals and practices of openness in computational social science? And how do researchers experience and respond to pressures to make their data open? Drawing on feminist epistemology, critical data studies, and science and technology studies, the analysis foregrounds the reflexive agency of practitioners – their capacity to reflect and recalibrate data sharing practices in response to shifting technical, ethical and political conditions. The findings show openness as a situated, relational practice, enacted at the intersection of vertical dimensions (rules, governance, infrastructure) and horizontal ones (community norms, peer relations, everyday ethics). By centering the perspectives of those at the frontline of open science in the AI era, this paper contributes to a more nuanced understanding of how openness is negotiated in practice and to the urgent question: what kind of openness is possible in the face of extractive generative AI?
The paper proceeds as follows. The first section situates the study within debates on openness, extractivism and situated ethics. The second outlines the methodological approach and empirical material. The third presents findings from interviews and document analysis, highlighting how data practitioners negotiate openness in the age of AI. The paper concludes by reflecting on the implications of these practices for responsible governance and commons stewardship.
2. Openness and ethical tensions around generative AI in computational social sciences
This paper draws on a study of actors in the computational social sciences (CSS) – an interdisciplinary field shaped by data-driven methods, digital infrastructures (González-Bailón Reference González-Bailón2013; Lazer et al. Reference Lazer, Pentland, Adamic, Aral, Barabási, Brewer, Christakis, Contractor, Fowler, Gutmann, Jebara, King, Macy, Roy and Van Alstyne2009). CSS applies computational methods – such as simulations, machine learning and large-scale data analysis – to investigate social phenomena. It integrates approaches from sociology, political science, economics, psychology, anthropology, communication studies, human geography and linguistics, often in collaboration with computer science, statistics, data science and physics, and has growing ties to all kinds of data-driven industries, such as social media. CSS spans methodological areas like social network analysis, agent-based modeling, computational text analysis, digital demography and behavioral economics. Meaningful access to all kinds of social (digital) data and its stewardship is a core condition for CSS, shaping not only what can be studied but also who can participate in the field (Olteanu et al. Reference Olteanu, Castillo, Diaz and Kıcıman2019). While CSS promises new insights through large-scale data analysis, critics have cautioned against its overreliance on methods from physics or computer science at the expense of social theory (Radford and Joseph Reference Radford and Joseph2020). Others have pointed to reproducibility challenges specific to CSS, including application programming interface (API) dependencies, data protection, data access restrictions and limited sharing incentives – issues that complicate matters of transparency and more generally the adoption of open science (Lazer et al. Reference Lazer, Pentland, Watts, Aral, Athey, Contractor, Freelon, Gonzalez-Bailon, King, Margetts and Nelson2020; Schoch et al. Reference Schoch, Chan, Wagner and Bleier2023).
2.1 Openness
Open Science emerged in the early 21st century as a response to growing concerns about the inaccessibility, opacity and inequality in publicly funded research (Mayer Reference Mayer2015). It is anchored in a normative project that frames scientific knowledge as a global public good – meant to be accessible, transparent, inclusive and responsive to diverse publics. This vision is most clearly expressed in the UNESCO Recommendation on Open Science (UNESCO 2021), which calls for “a new social contract for science” based on openness, responsibility and collaboration across societal and geographic boundaries. Drawing from traditions of the knowledge commons – including free software and open source culture (Kelty Reference Kelty2008; Weber Reference Weber2005) – Open Science has inspired a growing repertoire of practices, such as open access publishing (Suber Reference Suber2012), the FAIR (Findable, Accessible, Interoperable, Reusable) principles for data management and the CARE (Collective Benefit, Authority to Control, Responsibility, Ethics) principles for Indigenous data governance (Carroll et al. Reference Carroll, Garba, Figueroa-Rodríguez, Holbrook, Lovett, Materechera, Parsons, Raseroka, Rodriguez-Lonebear, Rowe and Sara2020; Wilkinson et al. Reference Wilkinson, Dumontier, Aalbersberg, Appleton, Axton, Baak, Blomberg, Boiten, da Silva Santos, Bourne and Bouwman2016), and citizen science, which refers to research approaches that actively involve members of the public in scientific processes such as data collection, analysis or interpretation, often with aims of democratizing knowledge production and broadening participation (Vohland et al. Reference Vohland, Land-Zandstra, Ceccaroni, Lemmens, Perelló, Ponti, Samson and Wagenknecht2021). While these practices often aim to democratize knowledge, foster reproducibility and improve public trust, their uptake varies across fields. In the natural and technical sciences, Open Science has frequently taken on a functionalist tone, aligning with efficiency and innovation agendas. In the social sciences, by contrast, uptake has been slower (Christensen et al. Reference Christensen, Wang, Levy Paluck, Swanson, Birke, Miguel and Littman2019; Laakso Reference Laakso, Engels and Kulczycki2021) and is shaped by at least three overlapping concerns: first, a reaction to replication crises in psychology and economics, which spurred calls for transparency and data availability (Korbmacher et al. Reference Korbmacher, Azevedo, Pennington, Hartmann, Pownall, Schmidt, Elsherif, Breznau, Robertson, Kalandadze and Yu2023; Ortmann Reference Ortmann2015); second, the open data movement, especially around public sector accountability (Arzberger et al. Reference Arzberger, Schroeder, Beaulieu, Bowker, Casey, Laaksonen, Moorman, Uhlir and Wouters2004); and third, a growing commitment to ethical reciprocity – especially the responsibility to “give back” to the communities under study (Cocq Reference Cocq2023). Over time, what began as a grassroots push for openness has become institutionalized through top-down mandates, with major funders and science policy bodies now requiring open access and data sharing (Mendez et al. Reference Mendez, Lawrence, MacCallum, Moar and Leonelli2020). The European Union’s investment in the European Open Science Cloud (EOSC) exemplifies this shift toward infrastructural consolidation (Burgelman Reference Burgelman2021). Open science was considered a key pillar of “Responsible Research and Innovation” (Owen et al. Reference Owen, von Schomberg and Macnaghten2021). Yet as Open Science scales, its normative ambitions increasingly encounter frictions around governance, incentives and the unequal capacities of researchers and institutions to participate.
Moreover, data sharing in science remains uneven and highly shaped by disciplinary epistemic cultures (Borgman et al. Reference Borgman, Scharnhorst and Golshan2019; Knorr-Cetina Reference Knorr-Cetina1999; Sieber Reference Sieber1991). Reichmann shows that dominant discourses on data sharing – across Open Science, research data management and Science and Technology Studies (STS) – frame it as the release of data for unspecified, global reuse, assuming that availability will naturally lead to reuse. While this literature is rich in empirical accounts of researchers’ motivations and barriers, it largely lacks theoretical grounding, takes reuse for granted and often overlooks distinctions between background (research support) and foreground (innovation) uses of data (Reichmann Reference Reichmann2022). Despite growing encouragement for open data and code, decisions about sharing often lie with individual researchers and are constrained by entrenched norms of authorship and credit (Abrams et al. Reference Abrams, Leone, Cambrosio and Faraj2025; Nelson Reference Nelson2016). Since academic recognition is still largely tied to final publications rather than intermediate outputs, researchers have little incentive to invest in the time- and labor-intensive work of curating and documenting data and code for reuse (Bezuidenhout et al. Reference Bezuidenhout, Leonelli, Kelly and Rappert2017; Pasquetto et al. Reference Pasquetto, Randles and Borgman2017). Moreover, ethical and technical challenges – such as ensuring privacy, interoperability and sustained accessibility – further complicate openness, especially in smaller teams or resource-constrained settings (Borgman and Bourne Reference Borgman and Bourne2022). Based on all these challenges, scholars have increasingly critiqued the assumption that openness is inherently virtuous.
Leonelli emphasizes that openness must be understood as a situated and relational practice; when implemented without attention to context, it can exacerbate epistemic injustice (Leonelli Reference Leonelli2023). Epistemic injustice denotes structural conditions under which certain actors’ knowledge claims are devalued or rendered invisible within dominant epistemic systems (Fricker Reference Fricker2007). While Leonelli has highlighted how open science can reproduce epistemic injustice by privileging well-resourced actors, she also emphasizes a deeper tension between openness and epistemic diversity, which is understood as the coexistence of heterogeneous methods, values and ways of knowing: standardized openness can marginalize heterogeneous methods, contexts and forms of expertise (Leonelli Reference Leonelli2022; Reference Leonelli2023). From this perspective, openness is not inherently emancipatory but must be evaluated also in relation to how data travel, are recontextualized and gain meaning across sites of reuse (Leonelli and Tempini Reference Leonelli and Tempini2020). Other authors warn of privacy risks, dual-use harms and the exploitation of marginalized knowledge (Bezuidenhout et al. Reference Bezuidenhout, Leonelli, Kelly and Rappert2017; Carroll et al. Reference Carroll, Garba, Figueroa-Rodríguez, Holbrook, Lovett, Materechera, Parsons, Raseroka, Rodriguez-Lonebear, Rowe and Sara2020; Zook et al. Reference Zook, Barocas, Boyd, Crawford, Keller, Gangadharan, Goodman, Hollander, Koenig, Metcalf and Narayanan2017). Spitale et al. (Reference Spitale, Germani and Biller-Andorno2024) and Lindemann and Häberlein (Reference Lindemann and Häberlein2023) argue that openness should be seen not as an end in itself, but as a means toward other ethical goals – which may at times conflict (Lindemann and Häberlein Reference Lindemann and Häberlein2023; Spitale et al. Reference Spitale, Germani and Biller-Andorno2024). These critiques call for an ethics of openness that is attentive to power, context and institutional embeddedness.
Beyond technical implementation, open science has always been framed as a political intervention – challenging privatization of knowledge, competitive individualism and gatekeeping in academic meritocracy. Activist communities have developed cooperative infrastructures, reclaimed journals from commercial publishers, built open citation corpora, called for reforms of research evaluation and experimented with alternative metrics of scholarly value (Han Reference Han2025; Hicks et al. Reference Hicks, Wouters, Waltman, de Rijcke and Rafols2015; Krämer and Rooryck Reference Krämer and Rooryck2016; Priem et al. Reference Priem, Piwowar and Orr2022). In response to the increasing commodification of knowledge and data, the concept of the “digital commons” has gained traction as an alternative framework for governing shared digital resources. Rooted in Elinor Ostrom’s work on common-pool resource management and extended by scholars of digital governance, digital commons approaches emphasize collective stewardship, bottom-up governance and equitable access (Birkinbine Reference Birkinbine2018; Hess and Ostrom Reference Hess and Ostrom2007). In contrast to proprietary enclosures and top-down institutional models, these initiatives aim to keep resources open while embedding them in normative frameworks of care, reciprocity and accountability. All these efforts reframe openness as a contested and political space rather than a neutral good. Yet, as several scholars have noted, open science’s progressive ambitions are often co-opted. Early taxonomies like Fecher and Friesike’s distinguish five thought schools in Open Science that come with overlapping but also different value sets: between the “public school” and “democratic school” of open science, noting that openness in practice can prioritize economic growth over inclusion or justice (Fecher and Friesike Reference Fecher, Friesike, Bartling and Friesike2014). Mirowski shows how funders and institutions instrumentalize openness to accelerate innovation and marketization (Mirowski Reference Mirowski2018). Birch and Muniesa critique the assetization of data, while Tkacz demonstrates how the language of openness aligns seamlessly with platform capitalism (Birch and Muniesa Reference Birch and Muniesa2020; Tkacz Reference Tkacz2014). Jon Tennant (Reference Tennant2018) provocatively asked: what does openness stand for, beyond just being “open”? He argued that openness requires a clearer articulation of its underlying values, particularly equity, justice, collective governance, and the orientation of scholarship as a public good.
These debates underscore the need to examine not only open science practices but the normativities they reproduce or challenge: the culturally embedded values, expectations and moral judgments that shape scientific life. Hence, normativities in Open Science capture the values – such as transparency, accessibility, collaboration and epistemic accountability – that guide scientific conduct. Scholars like Merton have historically conceptualized openness as a scientific norm (e.g. communalism), while more contemporary works emphasize how these values or “moral economies” are being renegotiated in digital and institutional contexts (Daston Reference Daston1995; Hosseini et al. Reference Hosseini, Hidalgo, Horbach, Güttinger and Penders2024; Koenig et al. Reference Koenig, Vilain, Tempio, Getzinger, Jahrbacher and Häller2024).
2.2 The extractivist turn with generative AI
The tension between making knowledge openly accessible and retaining responsibility for its downstream use is a long-standing ethical dilemma in science – now sharpened by AI, where open research outputs are readily absorbed into opaque, commercial systems with far-reaching social consequences. Science and technology have never been value-free – a core insight of STS and feminist epistemology – yet current political AI action plans often revive the illusion that technical systems can be neutral, masking the values embedded in their design and governance (Harding Reference Harding1991). Building on this critique, this paper adopts a practice-based perspective to examine how researchers themselves engage with these tensions. Drawing on Haraway’s notion of situated knowledges and Mol’s approach of ontological multiplicity, it explores how actors in CSS reflect on and respond to the ethical and political dilemmas posed by open science in the age of generative AI – not through abstract principles but through their concrete, context-bound experiences and socio-material conditions (Haraway Reference Haraway1988; Mol Reference Mol1999). Taken together, these perspectives foreground knowledge production as situated, relational and shaped by power, guiding the analysis in this study to interpret accounts of practice as positioned engagements with the normativities and governance of data sharing.
Further critical scholarship has increasingly exposed the structural tensions underlying contemporary data practices, shifting the focus from isolated cases to broader patterns of power and inequality. Arora shows how big data initiatives in the Global South often replicate extractive dynamics and global asymmetries, raising the question of who ultimately benefits from data-intensive development (Arora Reference Arora2016). Similar concerns surface in Helm et al., who argue that research infrastructures – despite being presented as inclusive – can entrench epistemic hierarchies and contribute to forms of data colonialism (Helm et al. Reference Helm, De Goetzen, Cernuzzi, Diwakar, Hume, Ruiz-Correa, Gommesen and Gatica-Perez2023). Being sensitive to these forms of extractivism allows to describe how data, labor and knowledge are increasingly appropriated for value generation without corresponding forms of consent, reciprocity or accountability. Drawing on feminist and decolonial scholarship, extractivism refers thus not only to the removal of resources but to a relational logic that systematically externalizes costs, erases situated contributions and concentrates power elsewhere (Alcoff Reference Alcoff, Sullivan and Tuana2007; Couldry and Mejias Reference Couldry and Mejias2019). Applied to generative AI, this lens foregrounds how openness can be mobilized to legitimize large-scale appropriation while obscuring asymmetries in who benefits and who bears the risks.
This also resonates with Birhane and colleagues’ analyses of large-scale machine-learning datasets, which, even when portrayed as neutral and open, embed cultural and systemic biases (Birhane et al. Reference Birhane, Kalluri, Card, Agnew, Dotan and Bao2022). Widder et al. extend this critique to the platform level, showing how Big Tech deploys the language of openness to invite community contributions while retaining control over core infrastructures, thereby consolidating power (Widder et al. Reference Widder, West and Whittaker2023). The development of generative AI, as Pasquinelli and Joler demonstrate, makes these extractive logics particularly visible, illuminating how narratives of openness can coexist with deepening concentration of control (Pasquinelli and Joler Reference Pasquinelli and Joler2021).
Scholars have documented how industry actors dominate AI research-shaping agendas, controlling infrastructures and harvesting public datasets without consent or reciprocity (Ahmed et al. Reference Ahmed, Wahed and Thompson2023; Editorial 2022; Whittaker Reference Whittaker2021). Open source, once a model of collaborative governance, is increasingly subject to “open-washing” where openness is selectively performed while data and decision-making remain closed (Liesenfeld and Dingemanse Reference Liesenfeld and Dingemanse2024). From an extractivism-sensitive perspective, open washing functions as a legitimizing discourse that frames appropriation as openness, while leaving underlying relations of power and dispossession intact. This is part of a broader political economy of extraction rather than an isolated governance failure. As Keller and Tarkowski note, the “paradox of openness” can either challenge or consolidate power, depending on who governs access and benefits (Keller and Tarkowski Reference Keller and Tarkowski2021). Crucially, this tension is not limited to proprietary actors: open source and academic model development can reproduce similar extractive data practices when consent, attribution and accountability are unresolved. Thus, it is the governance of openness – not openness as such – that determines whether data protection, Intellectual Property (IP) rights and fairness are upheld, underscoring why responsible stewardship must apply across proprietary and non-proprietary settings alike. Finally, the environmental footprint of large-scale AI raises further concerns. The energy and resource demands of training and running generative models call for forms of openness that are accountable not only to human users but also to planetary limits (Chilet et al. Reference Chilet, Tironi, Nicenboim, Lindley, Rosén, Salovaara, Botero and Søndergaard2024).
2.3 Situated ethics and reflexive agency
Following the extractivist dynamics outlined above, this section conceptualizes situated ethics and reflexive agency as forms of response that researchers develop when navigating openness under conditions of asymmetric power and data appropriation. Much of the AI ethics literature revolves around the imperative to ensure transparency, accountability and fairness in the development and deployment of AI systems. These values resonate strongly with the normativities of open science, which regard openness as innovation, reproducibility and public accountability in knowledge production. Scholars have called for explainable models, traceable data provenance and participatory forms of governance – not only as technical solutions but as ethical obligations (Binns Reference Binns2018; Gebru et al. Reference Gebru, Morgenstern, Vecchione, Vaughan, Wallach, Iii and Crawford2021; Selbst Reference Selbst2021).
While AI ethics and open science have evolved in separate domains, their shared concern with contesting opacity and redistributing epistemic power has led to similar approaches in conduct and governance – such as the development of guiding principles, ethical frameworks, self-assessment tools and community-based norm systems. Yet these principles often fall short when reduced to technical compliance. As critical scholars warn, ethical commitments can become depoliticized when they are translated into abstract principles detached from structural critique (Green Reference Green2021), much like openness itself.
In line with the call for this special issue, scholars, thus, advocate for a relational or situated ethics – one that foregrounds lived experience, context and power dynamics (Birhane Reference Birhane2021; Mager et al. Reference Mager, Eitenberger, Winter, Prainsack, Wendehorst and Arora2025). This perspective challenges abstract, one-size-fits-all checklists and instead conceptualizes ethics as a dynamic, participatory process embedded in sociotechnical practice. Building on Mol’s notion of ontological politics, that is, how realities are enacted through practices, ethics is understood as relational and enacted through concrete interactions and practices of care (Mol Reference Mol1999). Following this tradition, the present paper examines how researchers in CSS reflect on their own entanglements with openness and AI in data sharing. Rather than invoking abstract ideals, their reflections are anchored – as I will show – in specific dilemmata around integrity, collaboration, authorship, visibility, control and institutional incentives. When viewed within the broader context of knowledge governance, these micro-level accounts illuminate how ethical tensions are negotiated and made actionable in practice. The focus on reflexive agency offers a lens for understanding how actors are making sense of the ethical and political tensions of openness in the context of AI. I use reflexive agency to denote the situated capacity of researchers to negotiate, resist or strategically comply with openness demands in contexts shaped by extractive data economies, rather than assuming individual autonomy. Drawing from Giddens’ idea of reflexivity in late modernity and Beck’s reflexive modernization, it captures how individuals critically respond to the risks and contradictions of contemporary systems (Beck Reference Beck1992; Giddens Reference Giddens1984). It shifts attention from abstract norms to lived negotiations – where actors weigh epistemic values, infrastructural dependencies and political economies. As Couldry argues, such reflexivity is key to resisting extractive data logics, while Baack shows how openness can foster collaborative civic agency when critically attuned to power (Baack Reference Baack2015; Couldry Reference Couldry2014). In the empirical sections that follow, I show how such reflexive agency is enacted through selective sharing and contextualization practices in response to AI-related reuse pressures.
3. Methods
Methodologically, this approach follows Leonelli’s argument that understanding data requires taking the perspective of those who produce, curate, use and share it, and attending to their situated judgements and constraints (Leonelli Reference Leonelli2016). This study draws on 27 interviews from a larger corpus comprising 66 qualitative expert interviews and one recorded group discussion conducted between 2022 and 2024. For this article, I purposively selected interviews in which (a) participants were directly involved in CSS or adjacent data practices and (b) generative AI, AI-readiness or related concerns around extraction and enclosure featured as a substantive part of the conversation. Interviews that focused exclusively on other aspects of open science (e.g. citizen science, open peer review) without explicit engagement with AI were excluded from the present analysis but are included in the wider project. As generative AI gained prominence in late 2022, the topic was introduced iteratively into the interviews after early participants themselves raised it and is analyzed in relation to this temporal emergence rather than assumed to be uniformly present across the corpus. Interviewees were primarily researchers from CSS – including computational linguistics, psychology, sociology, political science and human geography – as well as research data stewards, data managers and data scientists engaged with social data, notably within social media research contexts. Participants spanned career stages from early-career researchers to senior professors and held positions in universities, research institutes, NGOs and media, mainly in the public or non-profit sectors of the European research landscape. While many were not European nationals, all resided in Europe at the time, so most perspectives reflect a European standpoint. According to their pronouns, 18 persons identified as male, 7 as female and 2 as diverse. This sample was not systematically constructed nor intended to be representative of the field; rather, it emerged organically from participants’ spontaneous reflections, providing exploratory insights into how generative AI, and AI technologies more broadly, may reshape the foundations and ethical dimensions of scientific practice. While the degree of engagement with AI varied among these 27 participants, all were selected for their substantial track records as advocates of open science, particularly in data sharing. The accounts presented here are understood as partial and thus positioned perspectives emerging from specific institutional, disciplinary and infrastructural locations.
Interviews (direct quotes represented by their IDs and section numbers) were conducted mostly in English online using a semi-structured narrative format (Brinkmann and Kvale Reference Brinkmann and Kvale2018). The format combined a short list of core prompts with ample room for participants to develop their own stories and examples. Interviews conducted before late 2022 primarily addressed open science and data governance in general terms; from early 2023 onward, the guide was adapted to include explicit prompts on generative AI. In the analysis, I treat references to AI in light of their temporal context, using them to trace a gradual problematization of openness rather than to make strict before/after comparisons. The interview guide opened with questions about participants’ trajectories in and around CSS and their involvement in open science and then invited them to recount concrete experiences with data sharing, governance and ethical dilemmas. Follow-up questions probed, where relevant, the perceived impact of AI and generative models on these practices. All participants provided informed consent for anonymized quotations. Due to the sensitivity of workplace-related content and the small, highly specialized communities involved, full transcripts cannot be shared. A detailed codebook and a separate methodological paper with aggregated findings are published open access online (REF anonymized).
German and French interviews were translated to English with the help of the software DeepL. Rather than explicitly focusing on AI, the interview guide posed open-ended questions designed to elicit critical reflections on the ethical dilemmas, practical tensions and institutional challenges encountered in the implementation of open research data practices. Through this reflective dialogue, many interviewees spontaneously introduced the topic of AI, articulating connections to broader concerns about data governance, infrastructural capacities and the ethical risks associated with data exploitation. Interviews were audio-recorded and transcribed using the open-source transcription tool Atrain, built upon OpenAI’s Whisper model (Haberl et al. Reference Haberl, Fleiß, Kowald and Thalmann2024). This reliance on proprietary AI infrastructure highlights the entanglements that the paper aims to render explicit, including the dependence of research on supposedly “open” practices within corporate AI ecosystems. Transcripts were subsequently reviewed manually for accuracy. Alongside interviewer notes, these transcripts were imported into MAXQDA software for qualitative content analysis.
The coding process adhered to Mayring’s (Reference Mayring2000) methodological approach, emphasizing systematic, iterative and rule-guided categorization of textual material to uncover latent themes, structural patterns and recurrent dilemmas. Categories emerged both inductively from empirical data and deductively from prior literature on open science, data ethics and AI governance as outlined in the section before. Particular attention was given to capturing not only the explicit narratives but also the contextual nuances and situated ethical reasoning through which interviewees articulated and negotiated tensions between competing values, such as openness, accountability, care and justice. In this sense, the interviews themselves constituted methodological interventions enabling participants to critically examine their moral reasoning in situ, reflecting the lived complexity of ethical life and the ontological politics of practice (Keane Reference Keane2015; Mol Reference Mol1999).
Complementing the interviews, I collected extensive field notes through participant observation at conferences and workshops on open science, research infrastructures and data governance between 2022 and 2024. Field notes were written during and immediately after events and subsequently imported into MAXQDA, where they were coded using the same coding scheme as the interviews. They served primarily to contextualize interview accounts, refine categories and identify convergences or tensions across settings rather than to provide standalone evidence. In addition, institutional policies and guidelines related to open science, research ethics, data sharing and occasionally AI were collected from universities, national research councils, repositories and professional organizations or data infrastructures. Documents were analyzed through qualitative content analysis, with deductive codes derived from the literature on open science, data governance and AI, and inductive codes capturing how openness, data reuse and AI were framed in practice (e.g. obligations, risks and rights). Often, interviewees participated actively in identifying these documents during or immediately after interviews, highlighting frequent institutional gaps. Notably, generative AI and machine learning were either marginally referenced or entirely absent from most policies. This gap itself became a significant empirical finding, reinforcing the uncertainty and lack of clear institutional guidance expressed by interviewees concerning emerging AI-related ethical and infrastructural challenges. Together, these data sources form the empirical basis for examining the reflexive agency of researchers, data stewards and research managers in open science in the era of intensifying AI integration. The following section, thus, presents the empirical findings, offering a situated account of how researchers and data practitioners experience, negotiate and respond to the shifting demands of openness in AI-affected research environments.
4. Engaging with openness in the age of AI: situated accounts from research practice
“We can still share. We can still collaborate. We can still be transparent. But in the age of AI and its insatiable appetite for data, we can’t afford to be naïve anymore.” (22.2) This statement, voiced by a data scientist involved in evaluation research, captures a sentiment echoed across many of the interviews: a growing awareness that while the principles of open science remain vital, they must now be rethought in light of new forms of data exploitation, institutional ambiguity and shifting ethical terrain. Drawing from expert interviews, this section foregrounds researchers’ own experiences, dilemmata and reflections. It highlights how practices of data sharing, valuation and governance are being reshaped by the growing influence of generative AI and machine learning systems – often in ways that clash with the normative ideals of open science. Rather than offering a singular narrative, the chapter is structured around recurring tensions and situated strategies, illustrating the contingent and often contested character of openness as it is enacted in everyday research practice.
4.1 “Openness was already hard work before AI” (13.5)
At the outset, participants often described their early engagement with open research data as a process of learning and implementing a set of well-defined, largely technical norms that dominate current data-sharing practice. These include first and foremost meeting funder and institutional mandates for openness, like depositing datasets in recognized repositories, adopting standardized metadata schemata, ensuring interoperability through common formats and ontologies and applying permissive open licenses such as CC BY. While these universal requirements are designed to enable maximum discoverability, accessibility and reuse across domains, they rest on an implicit assumption that openness is both an unqualified good and equally achievable in all contexts. For the computational social scientists in this study, fulfilling these norms was often experienced as administratively demanding and procedurally rigid, with limited room to account for the sensitivities of social data. The ethical tensions surfaced here concerned not only participant privacy and consent but also the researcher’s responsibility to anticipate potential misuse, safeguard against exploitation and preserve the contextual integrity of the data. Meeting abstract technical standards could, in practice, pull against these situated ethical commitments, forcing trade-offs in compliance.
In the context of CSS, data sharing commonly refers to making datasets, accompanying documentation and sometimes also code publicly accessible – either via disciplinary repositories, institutional platforms or informal peer networks. Shared datasets range from large-scale digital trace data (e.g. social media archives), survey data, open government data and experimental logs to annotated text corpora and network graphs. Increasingly, researchers are also expected to provide metadata, variable dictionaries and processing scripts to support transparency and reuse. This shift has brought about new infrastructural requirements, including the creation of data steward positions, development of data repositories, version control systems and standardized workflows – what some interviewees described as emerging “pipelines of accountability” (22; 18) – and has contributed to the broader adoption of the FAIR data principles as a guiding framework (Mons Reference Mons2018).
“Honestly, I would love to have a data steward on my team-someone who actually understands the legal stuff, the metadata standards, the repository requirements. I just want to focus on the research, not spend days figuring out licenses or file formats …,” says a social network analyst (3.5). The push for openness, many noted, was accelerated by funder mandates that turned what had once been an ideal into an obligation – often without corresponding institutional support. As one education researcher remarked, “Suddenly we had to comply with open data rules, but no one had built the structures to actually help us do it.” (9.24) This gap in institutions without robust open science policies and infrastructures – only two interview partners mentioned that they have support by dedicated faculty data stewards –, left individual researchers handling the complexities largely on their own. Effective data sharing requires not only technical skills but also a solid understanding of the legal conditions surrounding data use, modification and redistribution. Even data derived from publicly accessible sources – such as government portals or news archives – can become legally restricted for reuse, due to copyright and liability laws, and even more so once enriched, restructured or combined with other sources. Researchers frequently noted that they were not allowed to redistribute modified public data due to licensing terms or contractual obligations.
Additionally, the General Data Protection Regulation (GDPR) imposes strict conditions on the reuse of personal data, particularly the requirement to limit processing to the purpose originally consented to. This creates uncertainty for researchers working with sensitive or mixed data sources, who must balance legal compliance, ethical responsibility and the expectations of openness in their fields.
While GDPR is often understood by interview partners in particular, but also social scientists more generally, as a privacy barrier to openness, its regulatory landscape is more nuanced: multiple legal bases for research data processing exist alongside exemptions for scientific use, but their application varies substantially across national interpretations and institutional compliance cultures. This legal heterogeneity reinforces the finding that researchers navigate not a single rule but a patchwork of responsibilities and risks, shaping how they negotiate openness and protection in practice.
Thus, data sharing in CSS is far from uniform; it spans a spectrum of practices shaped by regulations, disciplinary norms, technical constraints and institutional or cultural expectations.
“I strongly believe in the importance of data sharing – it’s essential for transparency and cumulative science. But to be honest, it was already a legal and ethical minefield before AI came onto the scene. Now it’s just even messier …,” a sociologist mentions (15.18). Despite the considerable challenges, most interview partners did not question the fundamental value of openness for science itself. They emphasized its importance for transparency, cumulative knowledge production and collective scrutiny – especially within the academic community. However, enthusiasm for openness was more tempered when it came to its public reception. Outside of specific user groups such as journalists or policy advisors, researchers were unsure whether the broader public truly benefits from access to raw scientific data. More pressing for many was the sense that their contributions to open science – especially in the form of curating, documenting and sharing datasets – remained largely invisible in academic evaluations and career negotiations. Several expressed frustration at how little recognition such labor receives, both institutionally and in funding structures. There was also a shared sense that research policy often fails to appreciate the long-term value of replication and transparent data practices, focusing instead on novelty and innovation metrics. The COVID-19 pandemic reinforced these concerns: when preprints also in the domain of public health were publicly circulated and interpreted outside the usual peer review system, several interview partners observed confusion and mistrust among non-expert audiences. As one interviewee from the public health sector put it, “People started reading raw drafts without understanding what a preprint even is-suddenly, openness became a double-edged sword.” (25.13)
At the same time, many interview partners spoke with genuine enthusiasm about the new research cultures that have emerged around data sharing and open science. They described a growing ecosystem of support and recognition – from dedicated data tracks at conferences and the rise of data papers, to communities built around replication, transparency and methodological rigor. These developments provided not only new venues for professional exchange but also a renewed sense of belonging. Several researchers reported that they had “found their people” (22.21) – colleagues who, like them, cared about the integrity and quality of science. Practices such as preregistration and data management planning, once associated primarily with clinical research, were initially perceived as “bureaucratic burdens” (1.17). Yet over time, many even came to appreciate how these tools improved their work, providing structure and accountability. As one social media scholar reflected, “It may not count much on my CV, but it made my science better […] improved my integrity, and in the end that counts for me.” (20.20)
Beyond individual benefits, several interviewees highlighted a growing spirit of global collaboration emerging around data sharing-one that felt more horizontal and inclusive than the traditional, top-down modes of scientific internationalism. Unlike the era of large-scale, diplomatically brokered research infrastructures or “statistical cathedrals” (4.12), this new culture enabled researchers to connect across borders through shared practices, mutual support and common commitments to openness and integrity. It fostered a sense of distributed solidarity, driven not by institutional mandates but by the practical and ethical demands of doing science well in a “data-intensive world” (8.2).
A striking commonality across the interviews was the frequent invocation of the phrase “as open as possible, as closed as necessary.” Popularized by the European Commission under then-Commissioner for Research, Science and Innovation Carlos Moedas (Moedas and Commission Reference Moedas2016), this slogan was adopted by 10 interviewees as a guiding principle for their data-sharing practices, but also this principle has shifted with personal experiences: “Once I had been on team ‘as open as possible,’ now I tend to say […] ‘only open as necessary’,” (12.18) a data scientist tells me. Originally intended to promote openness while recognizing legitimate limits such as privacy, security or intellectual property, it has evolved into a normative shorthand for balancing transparency with responsibility. Its appeal lies in its flexibility: researchers can adapt it to specific contexts, satisfying funder expectations while retaining discretion over sensitive aspects of their work. Yet the interviews also expose the uncertainties embedded in this mantra.
Determining what counts as “necessary” closure has become more fraught in light of generative AI, which introduces risks of decontextualization, misappropriation and opaque downstream uses, as my interview partners note. In this sense, the phrase operates not only as a boundary marker between public and protected data but also as a focal point where ethical reasoning, institutional ambiguity and infrastructural constraints converge.
Taken together, reflections point to the need not only for better infrastructures and incentives but also for renewed attention to the social and epistemic conditions under which openness is made meaningful – both within and beyond the academy. Taken together, the accounts reveal that even before AI entered the scene, the normativities of openness were far from settled for these people. Researchers grappled with competing demands: between transparency and care, between policy imperatives and epistemic integrity, between individual responsibility and systemic support. What emerges – as I will show – is not resistance to openness as such, but an ongoing negotiation over under what conditions openness can be practiced responsibly.
4.2 “From sharing to scraping” (12.19): the AI turn in data use
While the practitioners interviewed were already handling the complexities of open science, the arrival of large-scale AI systems – particularly generative models – has introduced a new set of challenges. Several interviewees described a growing sense that data, once shared in the spirit of academic openness, are increasingly being appropriated for commercial AI development without consent, attribution or reciprocal benefit. This shift was especially visible in cases where carefully annotated data corpora, multilingual datasets or domain-specific taxonomies created for scholarly use were repurposed to train proprietary models. The frustration was not only about recognition or lost credit but about a broader sense of misalignment – between the values that motivated data sharing and the uses to which those data were ultimately put. As one computational linguist explained, “We shared our data to help other researchers, not to have it turned into a black box product sold back to the public.” (6.21) This sense of extraction – not openness – has come to dominate many researchers’ recent experiences with data reuse in the age of AI.
4.2.1 Exploited openness
“First, they made us pay thousands of euros to make our papers open, and now those same publishers turn around and sell everything – including our work, our metadata, even peer reviews – to train AI systems. We weren’t asked. We signed over our rights, and now it’s out of our hands.
It’s extractive, it’s cynical, and it completely undermines the ideals that open science was supposed to stand for.” (2.15), a social media researcher said.
Several interview partners expressed growing skepticism toward the current open access publishing model, which, while promoting openness in principle, has in practice reinforced profit-driven dynamics and shifted financial burdens onto researchers and institutions. Furthermore, the AI boom transformed open research data into what one interviewee called “training fuel” (27.8) – a raw high-quality resource scraped from public repositories and repurposed to build commercial models. These practices were experienced as extractive, not collaborative. Researchers described discovering their open-coded datasets embedded in proprietary classification systems or seeing their domain-specific annotations reproduced without citation in the documentation of commercial products. One social scientist recounted recognizing their labels, developed for a project on online radicalization, in a start-up’s sentiment analysis tool. “It was our taxonomy, even our examples – just lifted and used commercially without a word,” (23.28) they noted. They described that their emotional response ranged from disappointment to a profound sense of violation. The labor invested in contextualizing and documenting the data was not only unacknowledged but stripped of meaning in the transition from open science to commercial AI for online marketing.
Several interviewees voiced concern that the FAIR data principles – originally designed to enhance transparency, interoperability and reusability – are increasingly being reframed as pipelines for what is called “AI-readiness.” AI-readiness typically refers to the process of preparing data in ways that optimize it for machine learning systems – through standardization, labelling, formatting and metadata enrichment to ensure seamless computational processing and integration (Breuer Reference Breuer, Sudmann, Echterhölter, Ramsauer, Retkowski, Schröter and Waibel2023). While some interview partners acknowledged that this emphasis might be beneficial for fields such as astronomy, geology or pharmaceutical research, where highly structured datasets lend themselves to machine learning applications, they argued that this logic risks flattening the epistemic and ethical complexities of social science data. As one computational sociologist and repository manager put it, “FAIR has started to mean ‘feed AI reliably’ – and that’s fine if you’re talking about galaxies, rocks or molecules. But our data is about people, communities, politics. […] Computability is nice, but FAIR is also about trust. With all those black boxed AI models, how do I know what information is preserved, and if it does not strip away the very meaning we worked so hard to preserve, all the context information […]” (13.20). This unease highlights a tension between the technocratic push for AI-readiness and the situated, context-rich nature of social science data, raising questions about whose needs FAIR is really serving. For social data, the drive to make datasets “AI-ready” exposes further a profound ethical tension between FAIR-inspired norms of unrestricted reusability and the legal–ethical principles of purpose limitation and informed consent. Many social science datasets were collected under consent agreements tied to specific research aims, with no provision for unpredictable, large-scale AI applications or commercial reuse. Preparing such data for AI is creating uses that participants neither anticipated nor approved. This clash sharpens the ethical stakes of openness: compliance with abstract interoperability standards may come at the cost of violating the situated commitments researchers have to their participants. Interviewees’ concerns about AI-readiness, loss of context and unintended reuse resonate with Leonelli’s argument that data reuse across domains entails trade-offs between interoperability and epistemic diversity, particularly when data are detached from the conditions under which they were produced (Leonelli Reference Leonelli2022; Leonelli and Williamson Reference Leonelli and Williamson2023).
Interestingly, even some of the most committed advocates of copyleft and cultural remix began to question the implications of radical openness in light of generative AI. One data scientist, who had long championed free software, described a moment of unease when copyright activists from the media and entertainment industries began protesting the unauthorized scraping of creative content by AI companies. “I’ve always defended sampling, collage, cultural re-use; those are human acts. But when a machine does it, at scale, without context or intention, it feels different. It’s not remix – it’s automation.” (5.16) He went on to describe a growing fear of what he called a “global automaton” – a system that ingests the products of human culture indiscriminately and reproduces them without situated knowledge, without memory and without responsibility. This shift, for them, was not about reversing his stance on openness but about confronting the ontological and ethical rupture introduced by machine-driven reproduction.
4.2.2 Openness at risk: privacy, misuse and loss of control
The promises of openness in science rest heavily on trust: trust that data will be reused responsibly, contextualized appropriately and handled with care (Grand et al. Reference Grand, Wilkinson, Bultitude and Winfield2012; Oreskes Reference Oreskes2021). But with the rise of generative AI and thus increasingly powerful data fusion techniques, interview partners expressed heightened concerns about what openness means in practice when it comes to privacy, interpretive integrity and long-term accountability. In particular, interviewees report that concerns about re-identification have intensified. Interviewees working with evaluation data, student records or vulnerable populations described how data once considered safely anonymized now appeared much more vulnerable to computational recombination. Several interviewees expressed concern that large-scale language models obscure the provenance and context of scientific data once incorporated into training pipelines, while, on the other hand, they have the potential to re-identify anonymized data. Large-scale language models and probabilistic merging techniques have rendered it possible to triangulate information across datasets, reconstruct social profiles or infer sensitive attributes (Birhane et al. Reference Birhane2021). “I read that datasets that were thought sufficiently anonymized could potentially be pieced back together,” (21.7) explained one concerned repository manager. Another researcher working with at-risk populations noted that they now avoided sharing even certain aggregated datasets altogether, not because they rejected openness but because the technical landscape had fundamentally changed. These developments seem to “undermine long-standing de-identification protocols” (9.28) and start reshaping the ethical calculus of what can and should be shared.
Importantly, what is reportedly at stake is not merely epistemic control – that is, the ability to ensure that data is interpreted in line with its original scientific framing – but a deeper concern about ontological and relational consequences. AI systems do not “misinterpret” information; rather, their outputs reflect the statistical associations learned from training data and the design choices embedded in their models and deployment contexts; they restructure, repurpose and remix data across contexts. Several participants described cases where their open datasets had been used in misleading or even harmful ways. One interviewee recounted how a chatbot drew from their policy landscape evaluation and declared the sociopolitical impacts a success, omitting not only the methodological limitations and contextual nuances that were central to their analysis but also “obviously [the model] not being very well versed in correctly understanding negation in texts” (16.20). Another flagged literature and instances where AI models trained on linguistic datasets from Indigenous or marginalized communities had stripped language from its cultural referents, reinforcing stereotypes rather than fostering understanding (21.11).
“The point of sharing was to make science better – not to let it be weaponized or misused in ways we can’t trace,” (15.2) a sociologist emphasized. That inability to trace downstream uses was a repeated concern. Researchers lamented the lack of mechanisms to track or govern how data circulates once released, particularly when scraped or absorbed into AI training pipelines. Several noted that they had no way of knowing whether their datasets had already been ingested by commercial language models. The lack of transparency around training data – combined with the scale and automation of AI systems – left many feeling uneasy and left alone with a growing desire to take back responsibility that does not end with the publication of a dataset – which could only “be realized in a radically fully open environment” (22.9) – including open infrastructures and open data.
In parallel, interviewees also noted a subtle but probably significant shift in the research process itself, as AI was increasingly used not just to analyze data but to synthesize or even simulate it. They explained how language models were employed to generate hypothetical interview responses, summarize documents or classify social media posts – sometimes even standing in for human subjects. While some saw these developments as useful tools, in particular in early exploratory research phases, others worried about the epistemic and ethical implications: “What happens when highly biased AI begins to substitute the social world it was meant to help study?” (14.14) asked a political data scientist. Rather than treating openness as a basis for transparency and collective scrutiny, some researchers now see it as a shortcut to automation, using Large Language Models (LLMs) trained on open datasets to generate synthetic responses that simulate social behavior. As one senior computational sociologist in urban studies warned: “It’s incredibly dangerous when social scientists start thinking they can replace human participants with chatbots just because it’s cheaper or faster. Without a solid understanding of the statistical machinery behind these models, they risk generating synthetic data that looks plausible but has neither epistemic grounding nor any robust representation of diversity in the social world. It’s not just bad science – it’s a misunderstanding of what social research is supposed to do.” (26.22)
Last but not least, the topic of tensions between global infrastructures and local responsibilities became especially visible in the experiences of interview partners having roles as data managers and repository administrators. As open science infrastructures became targets of large-scale data scraping for AI training, many found themselves overwhelmed by uncontrolled bot activity, which not only strained bandwidth and budgets but also threatened the long-term viability of open access systems. “Every euro we spend fending off bots is a euro we don’t spend supporting research in the public interest,” (7.15) one infrastructure lead noted. These incursions prompted the implementation of new layers of technical defense – IP rate limits, bot detection, access protocols – that, paradoxically, constrained openness in the name of protecting it. Here, the material limits of openness become palpable: bandwidth, server capacity and maintenance budgets are not abstract concerns but the very substrate on which open practice depends.
At the same time, many interviewees voiced discomfort with how global cloud infrastructures and AI platforms, largely operated by private corporations, were becoming essential yet unaccountable layers in the open science ecosystem. While researchers were being pushed to comply with funder mandates for openness, they were rarely supported in handling the ethical and practical consequences of platform dependence. This disconnect between institutional expectations and situated ethical concerns led to a broader sense of disempowerment and lack of control: researchers felt responsible for the risks but lacked influence over the systems in which those risks materialize. As one put it, “We’re supposed to open everything up, but no one talks about where it ends up-or who profits from it.” (25.18)
Taken together, these reflections reveal how AI has expanded the stakes of openness. It is not simply about whether datasets are reused properly, but about how powerful automated systems can recombine, decontextualize and instrumentalize data at scale, often without the knowledge or consent of those who created it. The result is a growing tension between the normative ideals of open science, which are also embedded in funder mandates, and the shifting technical and political realities of AI-driven research environments. The interviews show that decisions about openness were shaped by a mix of ethical obligations and lived professional commitments. Participants spoke of a “responsibility to protect” research participants from harm, of “not handing over years of work” and of avoiding “feeding the big models” without consent or reciprocity. Others stressed the need to “keep trust intact” within research communities and to share “only when it makes sense and is safe.” Concerns about “losing control” over datasets, “being misrepresented” through decontextualized reuse and the “commodification of public data” were weighed against long-standing ideals of transparency and collaboration. These negotiations were often framed in terms of care, accountability and fairness, yet the balance point shifted depending on institutional rules, community norms, and the technical and legal affordances of infrastructures. Rather than abandoning openness, participants actively recalibrated it – narrowing access, setting explicit conditions or building alternative, commons-oriented infrastructures – in order to uphold what they saw as the core ethical purpose of their work in an AI-saturated research environment. In the next section, we turn to how researchers are actively negotiating this shift – developing new strategies, advocating for institutional change and reimagining what responsible openness could mean under AI conditions.
5. Coping with the AI turn
If the preceding sections diagnosed the tensions, what follows traces the generative responses: the reflections, strategies and role models through which practitioners attempt to reclaim openness as a commons-oriented practice rather than a vector of extraction. When I spoke with academic peers and open science activists about my research and my observations from the interviews, the most common response was a shrug: “That’s the tragedy of the commons,” they would say, as if the story had already been told, as if nothing could be done beyond accepting the inevitability of overuse and exploitation baked into the very idea of openness. I heard this analogy so often that it began to feel like a conversational closure – an all-purpose explanation that risks foreclosing alternative imaginaries. Yet Elinor Ostrom’s work reminds us that such tragedy is not inevitable but a design problem: commons thrive when governance is participatory, rules fit local contexts, and trust and reciprocity are actively maintained (Ostrom Reference Ostrom1990). The digital commons movement – with its open infrastructures and data cooperatives – demonstrates that shared digital resources can flourish under these conditions, even within the AI era’s political economy of corporate capture, crazy energy use, concentrated compute and uneven technical capacity (Krewer and Warso Reference Krewer and Warso2024).
When I asked interview partners to share examples of how openness could be done differently in the AI context, several pointed to projects they saw as inspiring counterpoints to extractive and closed corporate models. These were not abstract ideals, but concrete initiatives that, in their eyes, embodied a situated, pluralistic and reflexive openness – one that values local agency, transparency and community benefit while pushing against the default logic of enclosure. The BigScience/BLOOM project came up frequently as such a role model. Initiated in 2021, it gathered over 1,000 contributors worldwide to train BLOOM with open datasets, transparent governance and ethical guidelines shaped by feminist and decolonial critiques (Akiki et al. Reference Akiki, Pistilli, Mieskes, Gallé, Wolf, Ilić and Jernite2022). Many valued its explicit effort to democratize access to cutting-edge AI research and to develop shared rules for model use and data stewardship. At the same time, some noted that BLOOM’s reliance on France’s public supercomputing infrastructure fueled by nuclear energy highlighted an unavoidable structural tension: even a fully open model depends on scarce, centralized compute and energy resources, which are more accessible to well-connected actors in the Global North than to many others. “BLOOM showed me that you can do large AI projects openly and ethically, even when a large part of contributors came from industry […] but also that you can’t escape the fact that compute is the real gatekeeper.” (5.21)
Several grassroots collectives advancing natural language processing for local languages were also celebrated as a proof of concept for bottom-up AI commons. With Masakhane, a grassroots language technology community for Africa in mind, one interviewee admired its strong emphasis on local consent, collaboration with national language boards and data sovereignty for language communities. “As far as I know, Masakhane is really pushing open datasets and models […] they really care about trust and shared responsibility, even participation in governance, not like all these scrapers and automated data pipelines of commercial companies” (27.14), says a computer linguist. Yet here too, another interview partner notes, infrastructural dependencies persist: to train and deploy models, Masakhane projects still rely on Graphics Processing Units (GPUs) and cloud services, as well as financial support, often provided by large tech firms: “They can’t fully step outside the big providers, nobody can.” (10.8) In this context, several interviewees stressed the importance of establishing more balanced and equitable arrangements with both commercial and non-profit developers. Issues such as licensing and selective access have become central to these discussions, in particular within community-driven initiatives such as Masakhane, which foreground data sovereignty and linguistic diversity in AI development (e.g. Nekoto et al. Reference Nekoto, Marivate, Matsila, Fasubaa, Fagbohungbe, Akinola, Muhammad, Kabenamualu, Osei, Sackey, Niyongabo, Cohn, He and Liu2020).
Several interviewees expressed excitement about new, fully open and more environmentally responsible LLMs being developed in Europe. Projects in Switzerland and Norway were mentioned as emerging efforts to build large language models that are not only open in terms of code and data but also greener in their energy use – leveraging public funding, renewable-powered compute and transparent governance. For these researchers, such projects offer a glimpse of what openness could look like if extended beyond datasets and models to include the infrastructures that power them. “I would have no problem with open and green LLMs building on my work […] we need public infrastructures for that, a completely different story than the corporate race we see now.” (12.13) These were seen not only as technological alternatives but as political statements: that public AI infrastructure can be designed to resist capture, uphold democratic values and remain accountable to its users. Highlighted were their transparent governance processes and clear public benefit mandates.
Alongside these AI-specific examples, interview partners also pointed to role models within the open science ecosystem. The EOSC was mentioned as an ambitious attempt to create not only a federated, interoperable infrastructure for research data across disciplines and borders but also a European community of practice. While some noted that EOSC remains uneven in adoption, they valued its emphasis on FAIR principles, governance by the research community and investment in long-term infrastructure rather than short-term projects. “EOSC is imperfect, but at least it’s ours – it’s built with public money and public governance.” (6.15) Another one remarked the community governance of EOSC: “[…] it really became bureaucratic and complex – but it’s also a big experiment, and we have to learn from it. […] it is a start” (19.4).
National research data infrastructures were likewise cited as promising, particularly those that combine technical interoperability with discipline-specific stewardship. In the social sciences, the ODISSEI (the Dutch Open Data Infrastructure for Social Science and Economic Innovations) and the UK Data Service were held up as examples of how large, open databases can maintain high-quality curation, robust metadata and controlled access where needed, ensuring that openness is matched with responsibility, even though “their financial long-term sustainability is not really guaranteed” (7.6). For some, the work of data stewards in these infrastructures is most important, and a best practice for responsible openness – developing ontologies, harmonizing vocabularies, negotiating data access agreements – exemplified the “invisible labor” that makes openness meaningful in practice. “Data re-users don’t notice the data stewards until something breaks – but without them, none of this openness actually works. […]” (22.12).
Besides these positive examples, interview partners also shared their wishes for more activities: Several interviewees pointed out that peer learning about the infrastructural dimensions of openness is still rare. This goes hand in hand with the lack of institutional support for these matters, and the simple fact that researchers often remain unaware of the governance and maintenance work that sustains open resources until failure makes it visible, a reminder that infrastructure is noticed mostly in moments of breakdown (Star and Ruhleder Reference Star and Ruhleder1996). Open-source software was a recurring case in point: while widely relied upon in both AI and academic research, it suffers from chronic underfunding and maintenance burdens that fall on a small number of overextended contributors. Without better recognition, resourcing and shared responsibility for this work, even the most promising open initiatives risk collapse. “We keep talking about openness as if it’s free. It’s not – it needs money, time and people to keep it alive. […] We need more knowledge, not only with researchers, but also with research administrators, and most importantly with our infrastructure funders and managers” (2.23). Two interview partners flagged the German Sovereign Tech Fund, and one mentioned the French funding initiative of scikit-learn, a Python library for machine learning, as rare instances, where public money supports the often invisible work of open source maintainers, helping to strengthen foundational open digital infrastructure (Osborne Reference Osborne2024).
Taken together, the AI commons and open science examples suggest that alternatives to extractive AI are not speculative dreams but already in progress – yet they also underscore that sustaining them requires more than goodwill. It demands deliberate governance, adequate resourcing and a commitment to the often-invisible labor that keeps openness alive. Openness was framed as a practice to be actively stewarded – requiring governance mechanisms that guard against capture, redistribute benefits and preserve the researcher’s capacity to remain accountable for the uses of their work. Some interviewees drew inspiration from initiatives they regarded as proof that openness and responsibility can coexist – and even strengthen one another – when embedded in the right governance arrangements. These positive role models spanned different domains: long-standing digital commons initiatives, open-source software communities and collaborative open data infrastructures; trusted research environments that enable secure, privacy-preserving access to sensitive datasets; and data cooperatives that allow communities to collectively govern their own data. This stewardship of openness then operates along both vertical dimensions – institutional rules, licensing frameworks, funding mechanisms and infrastructural control – and horizontal dimensions – community norms, peer relations and everyday practices of trust and care. When these dimensions are aligned, openness and responsibility can coexist and even reinforce one another.
Against the fatalism of the shrug, several of my interview partners pointed to existing role models and “concrete utopias” (Bloch Reference Bloch1961): grounded yet aspirational visions of commons-based, even “AI-ready,” research sustained by selective openness, infrastructural safeguards, advocacy, public investment in open infrastructure and peer learning. For many, however, the central challenge is the risk of exploitation – of losing control over their research outputs and, with it, the ability to take responsibility for how these are used. This concern has prompted a more deliberate, case-by-case approach to openness, rethinking not only what to make open but also how and under what conditions. Few participants rejected openness outright; most continued to share data – especially metadata, derived indicators or annotated protocols – but did so more selectively and under more restrictive conditions. Some withdrew older datasets or limited access to trusted collaborators; others began experimenting with controlled licensing schemes or institutional gatekeeping via repositories to prevent unwanted use.
These adjustments were not merely defensive measures but part of a broader reimagining of what openness could mean under AI-driven conditions. The clash between universalist ideals of “data for all” and the situated ethics of conditional sharing led to proposals for alternative models: controlled-access repositories with community governance, contractual agreements to ensure reciprocity when commercial actors use public datasets and collaborative infrastructures that allow contributors to set context-specific conditions for reuse. These visions are anchored in values that surfaced repeatedly in the interviews – care for research participants, accountability to disciplinary peers, fairness in benefit-sharing and stewardship of the commons. The politics of data sharing, in this view, is inseparable from the ethical frameworks that guide it, and the future of openness lies not in universal mandates but in governance models and infrastructures capable of holding the tensions between transparency and protection, accessibility and control.
6. Conclusion
The disillusionment voiced by the computational sociolinguist, who opened this paper, is not an anomaly to be repaired, but an analytic starting point, revealing how claims to openness, under extractive conditions, must be treated as contested, and politically consequential rather than inherently virtuous. This explorative study has sought to show how openness in the age of generative AI is being negotiated in practice, not as a fixed moral imperative but as a situated, relational field of action (Leonelli Reference Leonelli2023). Across diverse contexts, what connects these efforts is less a universal technical design than a common orientation toward maintaining robustness and adaptability in the face of infrastructural dependencies, political constraints and extractive pressures – conditions inseparable from the wider political economies of data, computation and institutional authority.
Data sharing, thus, operates across vertical and horizontal dimensions, with practitioners moving between governance structures and lived practice. Reflexive agency lies in this movement – the capacity to adapt open practices in light of evolving political economies, infrastructural dependencies and ethical considerations. In these situated, ongoing experiments lies the possibility of an openness capable of resisting extraction while sustaining the conditions for shared knowledge. The individual accounts of my interview partners reveal a dense mesh of practical, ethical and political frictions: the absence of sustained institutional support, legal uncertainty, invisible curational labor, erosion of trust through uncontrolled scraping, the flattening of context under AI-readiness regimes and heightened risks of re-identification or misuse. What emerges is not a rejection of openness but a reconfiguration of its meaning under conditions where vertical asymmetries of infrastructural control intersect with horizontal networks of collaboration. In daily practice, openness appears both as a fragile local achievement and as a site of appropriation by global AI actors – making the politics of “who benefits” and “who decides” more pressing than ever.
While the ethical life of openness once thrived in grassroots engagement and activist commitments, it now unfolds within its own institutionalization as a top-down requirement. Voluntary, value-laden practices have become entangled with compliance regimes, funder mandates and the technicalities of standardization. Yet this formalization has not extinguished ethical deliberation; rather, it has re-ignited it. Researchers struggle to reconcile abstract norms with situated responsibilities, risks and relational commitments. In the age of generative AI, these tensions are intensified by new vectors of extraction and opacity, making the question of how to enact openness responsively, accountably and with care a central political and practical challenge.
The findings indicate that another form of openness is possible, but only if it is approached as commons’ stewardship rather than indiscriminate release into proprietary systems. This would entail collective governance, equitable access to sustainable and publicly funded computational and energy resources, and practices embedding care, reciprocity and accountability into the design of openness. Situated ethics and commons stewardship are, thus, not separate remedies but mutually constitutive: the former provides the epistemological ground for context-sensitive judgement, while the latter offers the governance arrangements through which such judgements can be collectively sustained. The situated ethics of data sharing teach us that openness cannot be sustained through principles alone but through infrastructures, alliances and legal safeguards that embed care and reciprocity into practice. They remind us that resisting exploitative appropriation requires collective governance attuned to disciplinary contexts, and a readiness to adapt rules, tools and norms as the political and technological landscape shifts. In the current global race for AI leadership – where openness risks being subordinated to competitiveness, speed and scale – such a re-politicized approach defends the democratic conditions for critical and collective inquiry while countering the enclosure of knowledge under the banner of disruptive innovation.
Acknowledgements
I thank the interview participants for sharing their experiences and reflections. I am also grateful to the anonymous reviewers for their constructive comments, which helped improve the manuscript.
Funding statement
This research was supported by the Austrian Science Fund (FWF). Grant DOI: 10.55776/V699.
Competing interests
The author declares no conflict of interest.
Use of AI and language tools
For translation of interview material originally conducted in German and French, the software DeepL was used. Interview recordings were transcribed using the open-source transcription tool Atrain, which is based on OpenAI’s Whisper speech recognition model. All transcripts and translations were subsequently reviewed and corrected manually by the author.
Open access
