BigScience

• BigScience's multilingual scope (BLOOM's 46 natural languages including under-resourced languages such as Wolof, Lingala, Setswana, and several Indian and African languages) materially widened representation beyond the usual high-resource set, which indirectly serves communities whose languages are normally excluded.
• The ROOTS corpus governance and the BigScience Data working group engaged in deliberate sourcing decisions and documented data provenance, a precondition for any sovereignty conversation.
• Community-led language selection: language communities and native-speaker researchers were involved in deciding which languages entered BLOOM, rather than scraping by availability alone.

• No explicit adoption of the CARE Principles for Indigenous Data Governance (Collective Benefit, Authority to Control, Responsibility, Ethics).
• No evidence of formal Indigenous data sovereignty protocols or consent frameworks for the communities whose language data entered ROOTS.
• Oral, non-textual, and relational knowledge is structurally absent — BLOOM is a text corpus model, so embodied and oral traditions are excluded by construction.
• Multilingual breadth is not the same as data sovereignty; inclusion of a language does not grant the speaking community authority over its use.

BigScience earns more than a floor score because its multilingual, community-involved language selection is a genuine, rare gesture toward represented-but-usually-excluded communities. But inclusion is treated as the achievement; sovereignty (authority to control downstream use, CARE adoption, consent) is not addressed. Score reflects real inclusivity without the governance layer that would make it preservative.

• BigScience was explicitly framed as a counter-history to the closed, well-resourced corporate lab model — a deliberate response to the concentration of large-model capability in a few firms, naming that concentration as a problem.
• Compute was provided via the publicly-funded Jean Zay supercomputer (GENCI/IDRIS, France), and this public-infrastructure provenance was openly acknowledged rather than hidden.
• The project documented its own carbon footprint and the geopolitical economy of compute access more candidly than most contemporaries.

• Limited reckoning with colonial data-extraction legacies behind the very web corpora that even a curated effort like ROOTS partly draws on.
• The labor history of data work (annotation, evaluation) across the Global South is under-examined relative to the project's stated values.
• Historical humility is present about compute concentration but thinner about AI's inheritance from extractive scraping norms.

Above-median: BigScience is unusually transparent about its own material conditions (public compute, carbon) and self-consciously historical about lab-concentration politics. It loses points for not extending that historical candor to colonial extraction and Global South data labor, which its own values invite.

• Multilingualism is constitutive, not decorative: BLOOM covers 46 natural languages and 13 programming languages, and multilinguality was a founding design goal rather than a post-hoc add-on.
• Native-speaker and regional researchers participated in working groups, bringing culturally-situated judgment into data and evaluation decisions.
• The collaborative, 1000+ researcher, 60+ country structure embedded plural perspectives in the production process itself.

• Breadth of languages does not guarantee preservation of culturally-specific reasoning patterns; the model still optimizes a single objective that can flatten non-Western rhetorical and epistemic forms.
• No explicit consultation with cultural scholars on whether translated/aligned representations preserve or erase culturally-specific logic.
• Risk that lower-resource languages are present in token but thin in depth, reproducing a representation gradient.

This is BigScience's strongest lens. Genuine, structurally-embedded multilingual and multinational participation distinguishes it sharply from English-centric peers. It falls short of top marks because inclusion at the data and personnel layer is not matched by explicit attention to preserving culturally-specific reasoning at the model-behavior layer.

• Open weights: BLOOM was released with fully open weights, enabling third-party verification, replication, and bias auditing — the single most important enabler of independent science.
• The ROOTS corpus and extensive documentation (model cards, datasheets, training logs) were published, supporting reproducibility.
• Known limitations were disclosed in the model card, and the project actively invited external evaluation and red-teaming.
• Training process transparency (the public training-log/chronicles) is exceptional relative to closed labs.

• Independent third-party audits of training data for bias were enabled but not institutionalized as a standing protocol.
• Replication is possible in principle but constrained in practice by the scale of compute required.
• Bias and safety evaluations, while open, were not exhaustively comprehensive across all 46 languages.

Highest-scoring lens. Open weights plus corpus plus training-log transparency make BigScience a near-model case for verifiable, reproducible science. Held just below the ceiling because comprehensive standing independent audits across all languages were enabled rather than guaranteed.

• The project's collaborative, open-science ethos carries an implicit cultural and aesthetic value — research as a public, communal act rather than a product launch.
• Multilingual scope inherently touches the affective and poetic registers of many languages, even if not framed that way.

• No explicit engagement with affective, intuitive, or poetic dimensions of language and knowledge.
• Documentation is technical and governance-oriented; ambiguity and poetic uncertainty are not given space.
• Emotional labor of the volunteer research community is not named or valued in the public framing.
• Modes of attention beyond efficiency/benchmark performance are largely absent.

Low, as expected for a research-engineering effort. The communal open-science character gives it a small affective dividend, but artistic perception is not an axis the project engages explicitly.

• The Responsible AI License (RAIL) is a concrete, forward-looking governance instrument: it places use-based behavioral restrictions on the model to steer downstream futures, an unusually structural move.
• Environmental cost was disclosed: BLOOM's training carbon footprint was estimated and published, engaging the environmental-future dimension directly.
• The open, collaborative governance model itself is an attempt to democratize who shapes large-model futures, distributing agency beyond a single firm.

• Labor displacement risks from large language models are not substantively engaged in the public framing.
• Inclusive deliberation about agentic/downstream deployment is limited; RAIL is author-imposed rather than community-deliberated.
• Long-horizon governance of forks and derivatives (enforcement of RAIL in practice) is acknowledged to be weak.

Solidly above median. RAIL and the carbon disclosure are real, concrete future-shaping instruments that most peers lacked. Capped because labor-displacement engagement is thin and RAIL's restrictions are imposed rather than democratically deliberated or practically enforceable.

• Genuine Global South participation: 60+ countries and 1000+ researchers, with deliberate inclusion of researchers working on under-resourced languages — participation at the production layer, not just the consumption layer.
• Open, volunteer-based working-group structure lowered barriers to entry for researchers outside well-funded institutions.
• Language selection foregrounded communities normally excluded from large-model coverage.

• Participation was largely volunteer/researcher-mediated; compensated feedback channels for affected non-researcher communities were limited.
• Disability-community accessibility is not a documented focus.
• Labor-representative engagement (for data and annotation workers) is not evident.
• Volunteer model can itself filter for those who can afford to contribute unpaid time, reproducing a participation gradient.

Among the better records in the field for who sits at the table, driven by real multinational participation. Held to a 6 because inclusion is researcher-mediated and volunteer-based rather than compensated and community-governed, and disability/labor dimensions are absent.

• BigScience is itself a structural inversion of the dominant story: it answered 'AI requires closed, corporate, well-capitalized labs' by building a large model openly with public compute and a volunteer collective — the official narrative tested by its own opposite, and it worked.
• By releasing weights and training logs, it refused to treat its own process as exempt from scrutiny, inviting contradiction.
• The RAIL license is a self-aware paradox: an open model with behavioral restrictions, openly naming the tension between openness and harm-prevention rather than pretending it away.

• Little evidence of disciplined irony or satire as instruments; the public voice is earnest and solemn.
• The unresolved contradiction at the heart of the project — releasing fully open weights while attaching unenforceable use restrictions — is acknowledged but not deeply interrogated as the paradox it is.
• The project's own seriousness about 'responsible' release is not turned back on itself (e.g., what does openness mean if anyone can strip RAIL?).

Median. As a project, BigScience embodies a genuine clarifying inversion (open collective beats the closed-lab assumption). But it does not wield irony or self-mockery as a discipline, and it leaves its central openness-vs-restriction paradox under-examined rather than followed to its instructive edge.