Digital Public Goods for Global AI Equity

Abstract

The panel explored how Digital Public Goods (DPGs) can advance AI equity worldwide. The conversation moved from the challenges of data quality and accessibility—especially in low‑resource contexts such as India—to the scarcity of evaluation benchmarks for African languages. Panelists described community‑driven initiatives (Masakhane, Mozilla Common Voice, DPG Alliance) and outlined concrete ways to make AI tools open, locally relevant, and safely deployable. Themes of digital sovereignty, the political economy of data, and the need for robust guard‑rails were woven throughout, culminating in a rapid‑fire round of bold proposals for scaling DPGs in the Global South.

Detailed Summary

• John Dickerson opened by stressing that raw data is not automatically useful; it must be collected with a specific purpose and be high‑quality, context‑sensitive.

• He highlighted two systemic issues:

  1. Mis‑aligned data collection – transaction‑ and social‑media datasets are rarely suited for health, agriculture, or other sector‑specific AI applications.
  2. Private‑platform data silos – large corporations hold proprietary data that is difficult for startups or public actors to access.

• Using India as a case study, John noted that despite being “data‑rich,” the country suffers from a “data desert” of usable, high‑quality datasets. The newly launched AI Kosh portal, for example, hosts only a handful of datasets, the most downloaded one having merely ~400 downloads—insufficient for a nation of India’s scale.

• Key warning: Openness alone does not guarantee equity. Without guardrails and governance, open data can be captured by well‑resourced actors (large AI labs, tech giants), reinforcing existing power asymmetries.

2. Benchmarking African Languages – From Absence to Action

• The moderator handed the floor to Chenai Chair (referred to as “Janai” in the transcript) to discuss language‑data gaps for African languages.

• Key observations:

  • Over 2,000 African languages exist; many have dialectal variation (e.g., Harare vs. Mutare Shona) and rich emotive nuance, which standard benchmarks (e.g., MLM, MLU) fail to capture.
  • Benchmarks are virtually non‑existent, limiting the ability to evaluate AI models for these languages.

• Masakhane’s response:

  • Launched an RFP (January 2024) to fund benchmarking of 40 African languages across speech and text modalities.
  • Emphasised context‑sensitive evaluation, aiming to understand which benchmark approaches work and why others fail.
  • Commitment to open‑licensing of benchmarks so that future projects can reuse and improve rather than reinvent the wheel.

• Sustainability angle: Community‑led, open resources are essential because global‑majority languages are under‑served by commercial AI products; only local ownership can ensure relevance.

3. Open‑Source AI & Deployment Challenges

• John Dickerson (Mozilla.ai) shifted the focus to deployment:

  • Concentration of compute – a few frontier labs dominate the AI stack, threatening the open internet model.
  • Open‑source communities are the “grassroots” mechanism to counterbalance this concentration. Mozilla’s historic role in the open web is being replicated for AI.

• Key theses:

  1. Ease of use is critical. Open‑source models must be as user‑friendly as commercial APIs (e.g., ChatGPT).
  2. Local / on‑prem deployment (running models on personal hardware) preserves data sovereignty and mitigates “surveillance‑by‑default” (e.g., Alexa, Meta glasses).
  3. Community funding & support – Mozilla.ai provides monetary backing and community‑building services to help small‑scale projects materialise.

• Vision: A future where developers default to open, locally run models, with the community handling the “boring” but essential engineering work (packaging, deployment scripts, documentation).

4. Digital Sovereignty – From Nations to Individuals

• The conversation moved to digital sovereignty:

  • Hardware layer – Full sovereignty (owning chips, fabs) is unrealistic for most nations; collaboration across countries is required.
  • Software layer – More achievable; coalitions (e.g., India + other data‑rich nations) can co‑develop sovereign AI stacks.

• Individual sovereignty was highlighted:

  • Users should retain personal data control, avoiding the “iPhone‑ification of AI” where a single vendor dictates device behavior.
  • Open‑source hardware experience (John’s 15‑year background) is seen as a pathway to build personal AI appliances that never send raw data to the cloud.

5. Community‑Driven Practices & Inclusion

• Lea Gimpel (DPG Alliance) described how community engagement underpins the DPG model:

  • Ubuntu philosophy (“I am because we are”) drives Masakhane’s collaborative ethos.
  • Open surveys and participatory design ensure that the community, not a top‑down authority, decides priorities.
  • Inclusive data collection – example: partnering with a Kenyan women’s sanitary‑product NGO to co‑fund voice‑data collection, ensuring gender‑balanced datasets.

• Impact: By placing community needs at the centre, DPGs achieve social relevance, higher trust, and sustainable stewardship.

6. Role of Civil Society & Contextual Evaluation

• Urvashi Aneja (moderator) highlighted the critical connective role of civil‑society organisations between tech companies, governance bodies, and end‑users.

• Contextual evaluation initiatives (in partnership with Masakhane) aim to:

  • Move beyond lab‑only metrics to real‑world usefulness (e.g., does a farmer‑oriented chatbot truly help a farmer?).
  • Incorporate safety and functional safety perspectives that vary by locale (e.g., different regulatory environments).

• Safety focus: Risks often emerge from routine deployment in low‑resource settings, not only from malicious attacks. Embedding community‑derived safety standards into DPGs is essential for long‑term reliability.

7. Rapid‑Fire Proposals – “One Bold Move”

Speaker	Proposed Bold Move (≈30‑60 s)
John Dickerson	Create a global “Rebel Alliance” where every frontier AI lab open‑sources the previous‑generation model when releasing a new one, coupled with community‑driven guardrails.
Lea Gimpel	Prevent the “iPhone‑ification of AI” – encourage users to vote with their wallets, rejecting opaque, closed‑source AI products.
Chenai Chair	Build a sustainable, South‑South collaborative ecosystem that avoids “parachuting” external solutions and ensures local ownership of technology.
Urvashi Aneja	Expose and regulate frontier labs’ business models, dismantling vertical integration in the AI marketplace to stop repeat exploitation of community‑generated DPGs.

8. Moderator’s Closing Synthesis

• Core message: Openness must be paired with governance, safety, and community agency to become a true catalyst for AI equity.
• Power shift: Technical openness is insufficient; political and economic power must also be redistributed through cross‑sector alliances (civil society, academia, industry, governments).
• DPG impact: Initiatives like Mozilla Common Voice, Masakhane benchmarks, and Digital Futures Lab’s evaluation frameworks illustrate concrete pathways to scale inclusive, trustworthy AI for the global majority.

Key Takeaways

• Data quality matters more than quantity; purpose‑driven, context‑aware data collection is a prerequisite for useful AI in health, agriculture, etc.
• Open datasets alone do not guarantee equity; without guardrails, powerful actors can capture and repurpose them, undermining inclusive goals.
• African language benchmarks are critically missing; Masakhane’s RFP to benchmark 40 languages is a concrete step toward context‑sensitive evaluation.
• Deployment usability is a make‑or‑break factor for open‑source AI; ease of local deployment must match the convenience of commercial APIs.
• Digital sovereignty is achievable at the software layer through regional collaborations; hardware sovereignty remains a shared international challenge.
• Community‑driven DPGs—grounded in participatory design, inclusive data collection, and open licensing—ensure relevance, trust, and sustainability.
• Safety must be contextual; real‑world deployment risks in low‑resource settings require community‑derived functional safety standards.
• Rapid‑fire proposals converge on three themes: (1) mandatory open‑sourcing of older model generations, (2) consumer choice against opaque AI products, (3) building transparent, collaborative ecosystems that dismantle existing power structures.
• The path to AI equity lies in moving from technology access to technology agency: enabling people and nations not just to use AI, but to own, govern, and shape it through Digital Public Goods.

---

See Also:

• data-sharing-infrastructures-for-ai-building-for-trust-purpose-and-public-values
• a-billion-voices-one-ai-how-language-tech-transforms-nations
• scaling-ai-for-public-health-impact-public-private-partnership
• implementing-ai-standards-for-global-prosperity-in-an-era-of-agentic-ai
• ai-for-economic-growth-and-social-good-ai-for-all-driving-economic-advancement-and-societal-well-being
• ai-in-governance-revolutionising-government-efficiency
• ai-governance-in-the-age-of-powerful-ai-iinternational-perspectives-and-the-code-of-practice