NOTE-098: Leonelli — Philosophy of Open Science

Contribution

This is the first book-length philosophical treatment of what “openness” means in the Open Science movement, by an author who is also an OS policy participant (Plan S ambassador; advisory roles for the EC and the Global Young Academy; UNESCO 2021 consultation). Leonelli does three things. (1) She identifies the dominant interpretation of openness as sharing: unlimited, digital, good, global and equal (Table 1). It comes with a value ordering in which transparency comes first, then quality (reproducibility), then inclusion (Figure 1). (2) She shows through four implementation cases that this interpretation clashes with the epistemic diversity of research practice and can worsen epistemic injustice. (3) She traces the sharing view to an object-oriented epistemology and proposes openness as judicious connection, grounded in a process-oriented epistemology: relational, social, divisive, situated, equitable, and focused on social agency (Table 3). Figure 2 inverts the value ordering to start from inclusion.

Key insight

Research outputs such as data, models and papers are “temporary signposts” of an ongoing process of inquiry (§6, p. 65), not self-standing commodities whose value can be assessed apart from how and by whom they were produced. Access therefore guarantees neither reliable reuse nor inclusion. “The further a research object travels from its context of origin, the more difficult it becomes to assess” whether its underlying assumptions fit the new use (p. 46). Openness worth having is the effortful, value-laden work of connecting systems of practice, which includes deciding whom to include and whom to exclude. “The pursuit of truth requires discrimination, and so does the practice of openness” (p. 42).

Assumptions

Premises and authorities rather than formal conditions:

• Ethics is integral to epistemology (standpoint theory and strong objectivity, Harding 1995), so epistemic and ethical concerns about OS practices are “impossible to disentangle” (§1, p. 3). Built on Longino, Kitcher, Wylie, Rouse, Potochnik, Chang.
• Scientific pluralism: reliable understanding of the world depends on a plurality of situated methods (§1, p. 5; §4).
• The unit of analysis is Chang’s system of practice, a “coherent set of epistemic activities performed with a view to achieve certain aims” (§4.1, p. 31), and not the discipline.
• The internal/external distinction in science can only be drawn case by case (after Longino, Wylie, Douglas, Solomon; §4.1, p. 29).
• An enactivist, embedded, embodied, extended model of cognition (fn. 25), with emotions as fundamental to cognition (fn. 31).
• The empirical base is largely biology and biomedicine, though “science” is meant in the broad sense of Wissenschaft (fn. 6). The empirical claims about OS implementation rest on the author’s and collaborators’ qualitative research and on the cited literature.
• The scope is institutional, large-scale, top-down OS discourse. She states that many grassroots initiatives already embody the view she favours (§1, pp. 2–3).

Key results

This is argument and case analysis; there are no quantitative results of its own.

• §2, “closed” science and the OS response. Commercialised, impact-factor-driven publishing and “minimal publishable units” give no credit for data, models or review. They push speed and overstatement and deter methodological diversity and transdisciplinarity. OS emerges (from open-source software; arXiv 1991; PLOS; open data, notebooks, peer review, citizen science) as a response, but large publishers have captured parts of it (author-pays Open Access), and policy framings tie it to innovation and growth. The EC 2016 definition (“sharing and using all available knowledge at an earlier stage”; “OS is to science what Web 2.0 was to … transactions”) exemplifies openness as unlimited access (p. 18).
• §3.1, GISAID. Access conditions (a user agreement, credit, limits on linkage) protected low-resourced data producers and brought wide participation: 240 countries and 15 million sequences by Feb 2023 (p. 21). Calls for “fully open” data prioritise transparency over inclusion, with epistemic costs, because diversity of data sources and of regional expertise shrinks.
• §3.2, technology and quality standards. In a Global Young Academy survey of researchers in Bangladesh, Ghana and Tanzania (Vermeir et al. 2018), researchers preferred expensive proprietary software (MatLab, Mathematica) to free equivalents, fearing that reviewers would read FOSS as a mark of low quality. Latest-model sequencers are privileged regardless of the accuracy the question needs (p. 23).
• §3.3, crop data. Open crop-data linkage favours decontextualised genomic data over phenotypic and local knowledge, and serves high-yield, market-led models. It can become “yet another form of bioprospecting” (p. 26).
• §3.4, reproducibility. Taking highly controlled experiments (RCTs, gene knockouts) as universal models of reproducible research, and presupposing precisely defined goals, damages exploratory, field-based and qualitative research. It devalues skilled judgement and does not address systemic causes (pp. 26–28).
• §4.1, epistemic diversity. This is defined as “the condition or fact of being different or varied in ways that affect the development, understanding and/or enactment of knowledge” (p. 29). Table 2 lists sources in six categories: conceptual, material, methodological, infrastructural, sociocultural, institutional. Systems of practice (1) are specific to local conditions, (2) differ in entrenchment within repertoires (Ankeny & Leonelli 2016), which gives a complementary explanation of the Matthew effect, (3) differ in permeability to newcomers, and (4) rest on demarcation strategies. “Systems of research practice are systems of demarcation and exclusion” (p. 35).
• §4.2, epistemic injustice. Fricker’s testimonial and hermeneutical injustice appear in the exclusion of breeders from crop science and in scepticism toward qualitative research. Epistemic injustice is “typically, inversely correlated” with epistemic diversity (p. 39). Popper’s piecemeal institutional pluralism informs EU OS policy (e.g. EOSC) but has produced a technocratic focus that defers justice concerns. Hence the inverted ordering: start from inclusion (Figure 2).
• §5.1, object-oriented view. The sharing view presupposes knowledge as extraction from accumulated objects (Bacon; the “resourcist” approach, Bonneuil 2019). It carries two features: (i) distrust of human judgement and a drive to mechanical objectivity, including AI and machine readability; (ii) centrality of ownership. The activist “sharing as reuse” variant shares the same modular, commodified epistemology (pp. 44–51).
• §5.2, process-oriented view and judicious connection. Research aims at “active knowledge” (Chang 2022). Connection has three senses: relation, scaffolding, and the cognitive or emotional state of the parties. Two exemplars: Crowdfight, micro-collaboration with over 45,000 volunteers in its first year (p. 55), and the CGIAR Ontologies Community of Practice, long-term connection that includes breeders’ and local classifications. Connections must be judicious, meaning selective and value-laden. This is Longino’s “tempered equality”, not Feyerabend’s epistemic parity (p. 59). Leonelli favours an equitable, not distributive, conception of fairness (p. 64).
• §6, conclusion and challenges. The main challenges are scale, bureaucracy (data management plans and pre-registration used as surveillance rather than reflection) and instrumentalisation by political economy. There are signs of change: Plan S reconsidering author-pays, “intelligent openness”, the UNESCO 2021 Recommendation (processes over products, collaboration over competition, inclusion over speed).

Claims

id	claim	strength	support
C1	Mainstream institutional OS interprets openness as unlimited, digital, universally good, global and equal sharing of research objects, ordered transparency → quality → inclusion	moderate	textual analysis of policy documents (EC 2016, National Academies 2018/2019, UNESCO etc.), §2; author concedes the reading is “opinionated” and not comprehensive
C2	Openness-as-sharing can reduce epistemic diversity and worsen epistemic injustice	moderate	four illustrative cases, §3, plus author’s qualitative research (cited, not presented)
C3	Reduced diversity and injustice degrade the reliability of knowledge, not only its fairness	weak–moderate	argued through cases (lost regional expertise, skewed crop biology) and standpoint theory; the causal link to unreliable knowledge is not measured
C4	Epistemic injustice is “typically, inversely correlated” with epistemic diversity	weak	assertion supported by two examples, §4.2
C5	Sharing is neither a necessary starting point nor a sufficient condition for reliable, responsible OS	moderate	argument from the object-oriented critique (§5.1) and cases
C6	Both the access and reuse versions of sharing presuppose an object-oriented, commodity epistemology of science	moderate	conceptual argument §5.1, with historical support (Kelty 2007; Bermuda Rules; Mirowski 2018)
C7	A narrow, universal conception of reproducibility used for demarcation harms exploratory and field-based research	moderate	philosophical literature (Radder, Romero, Guttinger, Feest) plus argument, §3.4
C8	Openness should be reconceived as judicious connection among systems of practice, starting from inclusion	weak (as a claim about better outcomes); moderate (as a coherent proposal)	normative proposal, §5.2, with two exemplars; no comparative evidence that it yields better science
C9	Researchers in low-resourced settings avoid FOSS for fear of reviewer stigma	moderate	cited survey (Vermeir et al. 2018) of Bangladesh, Ghana, Tanzania
C10	”What I demonstrated … is that the difficulties encountered in implementing OS … are tied to philosophical assumptions” (§6)	weak	the headline overstates the support: this is a plausible interpretive argument from selected cases, not a demonstration

Concepts

• Openness as sharing — openness as unlimited access to (or reuse of) research objects: unlimited, digital, good, global, equal (Table 1).
• Openness as judicious connection — openness as the effortful, selective creation of relations among systems of practice, mediated by but not reducible to the exchange of objects: relational, social, divisive, situated, equitable (Table 3).
• Epistemic diversity — “the condition or fact of being different or varied in ways that affect the development, understanding and/or enactment of knowledge” (p. 29).
• System of practice — after Chang (2013), a coherent set of epistemic activities performed to achieve certain aims.
• Repertoire — after Ankeny & Leonelli (2016), an entrenched, widely adoptable blueprint for doing research that crosses disciplines, such as the RCT as the exemplar of reproducibility.
• Demarcation strategy — a system’s criteria for what counts as a scientific contribution and whose expertise is relevant. It is constitutive of systems of practice.
• Epistemic injustice — after Fricker (2007): “wrong done to someone specifically in their capacity as a knower”. Testimonial: prejudice against someone as a contributor. Hermeneutical: marginalisation of ways of knowing to the point of unintelligibility. Massimi’s “epistemic severing”.
• Object-oriented vs process-oriented view of science — knowledge as accumulation of and control over tradeable outputs, versus research as skilled, collective agency aimed at “active knowledge”, where outputs are situated scaffolds.
• Active knowledge — after Chang (2022), “the ability to perform and coordinate certain activities and interventions in ways that are purposeful”, “not reducible or subordinate to the storage and retrieval of information”.

Connections

The book builds on Merton’s norms and Popper’s Open Society (read critically), on the pluralism and practice literature (Chang, Longino, Kitcher, Rouse, Wylie, Mitchell, Solomon), on standpoint theory (Harding), on Fricker on epistemic injustice, and on the author’s own work on data-centric biology (Leonelli 2016). On the trust–transparency trade-off, Leonelli takes a middle position against O’Neill and Nguyen: “strategic” transparency can build trust, depending on institutional credibility (fn. 22).

In the record: Feest (LIT-161) is cited for the method-relativity of reproducibility and the instability of research targets. Leonelli’s critique of reproducibility as a monolithic demarcation criterion runs parallel to Feest’s deflation of replication. Douglas (LIT-139) is cited among those who deny a strict internal/external boundary for values, and Leonelli’s insistence that value judgements are unavoidable in OS design extends that inductive-risk line to infrastructure. Porcher et al. (LIT-129) apply the same Fricker framework (epistemic injustice) to clinical settings. Leonelli’s scepticism about RCTs as a universal exemplar echoes Worrall (LIT-175) and Cartwright (LIT-213) from a practice-oriented rather than confirmation-theoretic direction.

Epistemology tag. The work is about social epistemology: who counts as a knower (epistemic injustice), how trust and expertise are distributed and recognised, and what knowledge is (active, process-oriented, situated, not stored objects). Its positions are that the reliability of knowledge depends on the diversity and justice of its producing communities, and that transparency does not by itself produce trust. The tag is justified.

Bearing on the record

This is not an ML practice paper, and it contains no instruction for ML practice. Two points bear on the anthology’s concerns, as this reader’s extrapolation. First, §5.1 argues that data travel badly: the further a dataset moves from its context of origin, the harder it is to judge whether it is fit for a new purpose, and bigger, standardised, machine-readable data is not automatically better than smaller, well-curated data. That is an argument in the territory of dataset documentation and provenance practice. The anthology holds such practices, but this book is a philosophical motivation for them, not a source that could support one. Second, §3.4’s critique of a monolithic reproducibility standard is relevant to how the anthology frames replication of ML results, but it argues from field and exploratory sciences, which transfer poorly to computational experiments. Leonelli herself notes that computer simulations are the case where “both procedures and results are expected to be fully reproducible” (p. 27). Nothing in the anthology cites this work.

Limitations

• The critique targets an ideal type of institutional OS that Leonelli admits is a simplification. She says many OS realities already embody her view, so the scope of the criticism is uncertain.
• The empirical support is four illustrative cases and references to the author’s prior qualitative research. There is no systematic evidence that sharing-first OS has in fact lowered reliability, or that connection-first OS raises it.
• “Judicious” is not operationalised. The view tells researchers what to attend to (demarcation strategies, inclusion, context) but not how to decide, and it accepts that every decision creates new exclusions.
• The object-oriented/process-oriented dichotomy is broad. The book acknowledges that the object-oriented view has had “tremendous success” but does not weigh those successes against the harms described.
• Costs are acknowledged but not estimated. Connection-first OS is “resource-intensive”, slower, and “under constant threat by funders”.
• The account is built mainly on biology and biomedicine. Its application to fields with stable targets and fully computational methods is asserted, not argued.

Open questions

• Can “judicious connection” be turned into assessable criteria for funders and infrastructure designers without becoming the bureaucratic surveillance the book warns against?
• Is there comparative evidence (e.g. GISAID versus fully open sequence databases) on whether conditional-access regimes produce more diverse or more reliable science?
• How does the process-oriented view handle domains where outputs are the primary object of study and travel well, such as computational artefacts, software and models?
• How should OS reconcile Popperian institutional autonomy with the demand that justice concerns shape infrastructure design (§4.2), and who adjudicates?

Corrections to the seeded skim

• The seed’s topic question (whether the process-oriented view yields concrete criteria for “judicious” connections) is answered: it does not, and Leonelli says so. It is a normative reorientation. It gives four characteristics of systems of practice to attend to (specificity to local conditions; entrenchment within repertoires; permeability to newcomers; demarcation strategies, §4.1), two worked exemplars (Crowdfight; the CGIAR Ontologies Community of Practice, §5.2) and a five-plus-one contrast table (Table 3). It gives no decision rule. “How this vision of openness can be effectively realized remains itself an open question” (§6, p. 66).
• Leonelli distinguishes two versions of sharing that the seed does not mention: sharing as unlimited access (institutional, often commercial) and sharing as unlimited reuse (activist, anti-ownership, e.g. the Bermuda Rules or FAIR). She calls reuse “prima facie … a much better foundation”, but argues that both share the object-oriented epistemology: knowledge as modular, commodified components, plus distrust of human judgement (§5.1, pp. 49–51).
• On the seed’s framing of the work as bearing on “open weights, open data, benchmark sharing”: the book never discusses ML model release. Its only AI content is in §5.1 (p. 46), where machine readability, the prized property of shared data, is tied to “mechanical objectivity”. She challenges the expectation that a classifier trained on 10,000 plant species is better than one trained on 1,000, or national data better than well-curated local data. That ML reading is the seed’s extrapolation, not the book’s claim.
• On §3.4 and Feest (LIT-161): Leonelli cites Feest 2019 for the claim that reproducibility means different things across methods and for the instability of targets (pp. 27, 32). Her target is a “monolithic” or “overly narrow” conception of reproducibility used as a demarcation criterion, with RCTs and gene knockouts as exemplars. She does not downplay replication as such. “Why is this result different” can be more valuable than “where is the mistake” (p. 28).
• The seed does not record Leonelli’s stated method and stance. The book is “not meant as a purely descriptive take” but “a normative interpretation … explicitly grounded in a normative perspective” in the tradition of standpoint theory and strong objectivity (Harding). She describes her reading of the OS movement as “opinionated and … by no means comprehensive” (§1, pp. 2–3; §2.3, p. 17).
• epistemology tag: justified. Social epistemology is central: epistemic diversity, Fricker’s testimonial and hermeneutical epistemic injustice, trust and expertise, and a process-oriented account of knowledge (“active knowledge”, after Chang).
• primary topic: philosophy-of-science is right, since it is explicitly a philosophy of science of Open Science. society-and-governance is well justified as a secondary tag (science policy, EU/UNESCO/Plan S). ethics is justified.