Bibliothèques Sans Frontières

From BSF's perspective, success would require three concrete outcomes: 1. A shared global framework for inclusive AI. The Dialogue should establish minimum standards ensuring AI systems work for communities currently excluded from the digital ecosystem, in particular, speakers of low-resource/under-documented languages, populations without reliable connectivity, rural and displaced communities. These standards should be measurable, not aspirational. 2. A global fund for AI capacity-building in the Global South. Governance without resources is meaningless. The Dialogue should catalyze dedicated financing mechanisms to support open-source AI infrastructure, community-governed datasets, and offline-capable tools in under-resourced contexts. This would operationalize the principle that AI must be a global public good. 3. Formal recognition and resourcing of community-governed AI commons. Datasets, benchmarks, and models built with public or philanthropic funding, especially those developed with and for underrepresented communities, should be treated as open digital commons. The Dialogue should establish governance norms protecting these resources from enclosure by proprietary actors. A successful Dialogue would also demonstrate that multilateral AI governance is possible: that states, civil society, academia, and the private sector can agree on rules of the road that do not simply entrench the interests of the most technologically advanced nations and corporations. Finally, success requires that the Dialogue itself model the inclusivity it seeks to advance with genuine representation from the Global South, from civil society actors working at the frontlines of the digital divide, and from communities whose voices are most absent from current AI governance discussions.

• Open-source software, open data and open AI models
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches

AI capacity-building is not merely a technical transfer problem but a structural governance challenge. The capacity to develop, evaluate, audit, and contest AI systems is concentrated in a small number of high-income countries and commercial actors. Without deliberate redistribution of this capacity to local institutions, civil society organisations, and community data stewards, governance frameworks will remain aspirationally universal but operationally exclusionary. The social, economic, cultural, and linguistic implications of AI are inseparable from technical architecture. Of approximately 7,000 living languages, fewer than ten achieve performance parity in leading generative AI systems. The interoperability of governance approaches is a precondition for coherent global action. The current landscape is fragmented across dozens of bilateral agreements, regional frameworks, and voluntary industry initiatives that operate in mutual isolation. From BSF's operational perspective, this fragmentation is not merely an abstract institutional problem: organisations deploying AI in multiple jurisdictions across West and Central Africa navigate incompatible data governance regimes, inconsistent standards for community consent, and regulatory vacuums that shift unpredictably across borders. Open-source software, open data, and open AI models constitute the infrastructural preconditions for equitable AI. Proprietary ecosystems systematically foreclose the adaptation, replication, and accountability mechanisms upon which under-resourced communities depend. BSF's commitment to open licensing across all community-produced datasets reflects a principled position that AI resources developed with public or philanthropic support should function as global digital commons.

The first concerns the governance of offline and locally-deployed AI. Extant governance frameworks assume cloud connectivity as the operational baseline, an assumption that structurally excludes the approximately 2.2 billion people who lack internet access. BSF's deployment of Small Language Models with Retrieval-Augmented Generation on low-power hardware across West Africa and conflict-affected contexts demonstrates that offline AI is not a speculative future condition but a present operational reality. Yet this deployment modality generates governance challenges that no existing framework addresses: How are models updated in disconnected environments? How is misinformation corrected when there is no central oversight infrastructure? How are communities protected from AI errors when standard accountability mechanisms presuppose connectivity? These questions constitute a distinct governance domain requiring dedicated normative attention. The second concerns community data sovereignty and the governance of linguistic commons. The extraction of data from Global South communities for AI training, without meaningful consent or benefit-sharing, represents a structural pattern that existing intellectual property and data protection regimes inadequately address. This problem is particularly acute for oral and indigenous knowledge systems, which are increasingly captured by commercial actors but rarely returned to communities of origin in usable form. BSF's AI Social Club model, in which communities collectively govern the collection, annotation, and licensing of their own linguistic data, operationalises one approach to community data sovereignty. The Dialogue should develop generalisable governance norms from such practices, specifying who owns community-generated data, who governs its use in AI training pipelines, and how benefits are distributed. These questions sit at the intersection of data governance, intellectual property, and human rights, and require cross-cutting treatment that the current thematic architecture does not provide.

The most pervasive effect is structural exclusion by default. Because international AI governance frameworks have not established linguistic inclusivity as a baseline requirement, major AI systems are built without meaningful support for languages such as Wolof, or Pulaar, or even local version of French. In Senegal, where only 4.4% of the population are native French speakers, AI tools deployed in health and education are functionally inaccessible to the majority of users. BSF's kSanté chatbot, serving community health workers, and Karibu Prof application, supporting allophone teachers, address this gap at project level; they cannot substitute for systemic governance reform. Another dimension concerns the accountability vacuum in fragile and conflict-affected contexts. Humanitarian and legal actors are deploying AI tools in environments where errors carry life-altering consequences, in the absence of any international standards governing accuracy, safety, or accountability. BSF's Redline Handbook, a legal AI assistant supporting the documentation of conflict-related sexual violence, operates under precisely these conditions. The absence of governance norms for high-stakes AI in humanitarian contexts is not a peripheral concern; it affects some of the world's most vulnerable populations. Also, there's the opportunity cost of proprietary enclosure. Without governance frameworks mandating open licensing for publicly funded AI resources, community-developed datasets and models are routinely absorbed into proprietary systems, foreclosing the replication and adaptation that under-resourced actors require. BSF's experience confirms that open-source, community-governed AI is operationally superior in constrained environments, but depends on enabling governance conditions that do not yet exist at the international level.

The Dialogue's most significant potential contribution lies in establishing minimum global floors rather than aspirational ceilings. The preponderance of existing AI governance initiatives address leading-edge AI systems and advanced deployment contexts. A governance architecture that attends only to frontier AI while leaving the regulatory conditions for deployment in low-income, low-connectivity, and linguistically diverse contexts unaddressed will systematically reproduce existing inequalities. The Dialogue should produce binding or verifiable minimum standards applicable to all AI deployments in public-interest functions, including requirements for linguistic coverage, offline functionality, and community consent in data collection. Linked to this is the concern around the coordination of capacity-building initiatives. Existing bilateral and regional efforts are frequently duplicative, concentrated in capital cities, and disconnected from the civil society actors and community organisations that are the primary points of AI deployment in under-resourced settings. The Dialogue can function as a coordination architecture: mapping existing initiatives, identifying coverage gaps, and channelling resources toward actors working at the operational frontier of equitable AI. By formally recognising community-governed open-source AI and open digital commons as legitimate and preferable modalities of AI development, the Dialogue would generate enabling normative conditions for the ecosystem of organisations working outside the commercial AI mainstream.

Community-led initiatives producing open linguistic resources for underrepresented languages, including Mozilla's Common Voice, the Masakhane research community, and BSF's AI Social Club, represent a category of AI infrastructure that current governance frameworks neither recognise nor resource adequately. The Dialogue should establish formal mechanisms for the recognition and sustained financing of these community-governed commons. UNESCO's Recommendation on the Ethics of AI (2021) constitutes the most comprehensive multilaterally endorsed ethical framework, with provisions on cultural diversity, linguistic pluralism, and the rights of marginalised communities

Civil society organisations with operational presence in under-resourced contexts possess an irreplaceable form of knowledge: empirical evidence about what AI governance gaps look like when they materialise in practice. BSF's experience deploying offline AI tools with community health workers in Senegal, legal practitioners in conflict-affected DRC, and allophone teachers in West Africa generates evidence (RCT) and narratives about linguistic exclusion, accountability vacuums, and the feasibility of open-source alternatives. This knowledge should be institutionally positioned as substantive evidence for the Dialogue's deliberations, not as anecdote in its margins.

Speakers of low-resource /under-documented languages represent the most analytically significant underrepresented category. Of approximately 7,000 living languages, fewer than ten achieve performance parity in leading generative AI systems. We work in Sub Saharan Africa and this exclusion is particulary felt in French-speaking countries where exclusion from LLMs is the most severe (roughly 30% more). The millions of speakers risk exclusion from access to AI-mediated health services, educational tools, and public information systems that are most directly determined by governance choices made without their participation. These communities can function as active producers of linguistic data and governance frameworks, not merely as objects of AI systems designed elsewhere. Women and girls in low-income countries face compounded barriers to both AI access and participation in governance processes and are also the most exposed to the risks of harmful AI-based practices. Furthermore, their perspectives are often absent from norm-setting at the international level and the development of AI-based models.

Distributed community deliberation processes would extend the Dialogue's consultative reach to communities for whom conventional participation channels are structurally inaccessible. Practitioner immersion for norm-producers would partially address the epistemic distance between those who draft governance frameworks and those who implement them. Structured field visits positioning government negotiators, UN officials, and senior technical advisers as observers of AI deployment in operational contexts.

BSF's operational portfolio offers several empirically grounded governance practices with generalisable implications for the Dialogue's deliberations. The community-governed linguistic commons model, operationalised through BSF's AI Social Club, organises local communities to collect, transcribe, annotate, and govern their own linguistic data in low-resource languages including Wolof, Pulaar, Soninké, Dioula, Sérère, and Mandingue. All resources are published under open licenses as global digital commons. This model addresses three governance problems simultaneously: it corrects the structural underrepresentation of low-resource languages in AI training data; it establishes community data sovereignty as an operational practice rather than a rhetorical commitment; and it generates replicable infrastructure that researchers and developers can build upon without proprietary restriction. The three-tier validation architecture developed for BSF's high-stakes AI deployments, including PreventAI and the Redline Handbook, provides a practical template for human oversight in contexts where AI errors carry direct humanitarian consequences. Sequential review by domain experts, field practitioners, and technical auditors is built into the deployment architecture rather than appended as a post-hoc accountability mechanism. This design principle, that oversight should be structural rather than discretionary, has direct implications for AI governance standards in health, legal, and crisis contexts. The offline-first design principle instantiated across BSF's AI portfolio, including kSanté, kPROF, Tutobac, and PreventAI, operationalises an equity standard with clear governance relevance: AI systems deployed in public-interest functions should be required to achieve baseline functionality without internet connectivity. This is technically achievable through the combination of Small Language Models, Retrieval-Augmented Generation, and pre-generated content strategies that BSF has developed with partners Pleias and Kajou. Making offline functionality a governance requirement rather than an optional feature would structurally redirect AI development resources toward the populations most excluded by connectivity-dependent design assumptions.