Public institutions and development partners are increasingly seeking AI-enabled solutions that are reusable, interoperable, and locally adaptable. Yet AI systems differ from conventional DPG software in ways that complicate their assessment and stewardship: training data provenance is often opaque or constrained; model behaviour cannot be inferred from the source code alone; and many harms arise at deployment and reuse rather than at release. Without clearer, evidence-based criteria for what must be open (and what can be governed through managed access), and without clearer accountability across the AI value chain, the “AIDPG” label risks becoming unattainable (blocking public-interest innovation) or diluted (undermining trust and safety).
Our evidence points to four conclusions that shape a realistic path forward for AI as DPGs.
- Openness is multi-dimensional, not binary. Stakeholders consistently treat openness as a spectrum across components (code, weights, data, documentation) and over time (staged release), while current operational tests skew towards model-release binaries (e.g., OSI’s Open Source AI Definition) that do not fully address composite, service-based, or locally assembled systems. While openness cannot be reduced to a single binary test, practical governance requires a common reference framework. Accordingly, this report supports the use of the Model Openness Framework as a shared taxonomy for describing and comparing degrees of openness, rather than as a definitive pass/fail measure of whether an AI system is “open.”
- Societal benefit and SDG alignment are not guaranteed by openness. “Open” and “public-good” cluster as distinct ideas; privacy and safeguards sit at the centre, linking openness to legitimacy and trust, and making responsible data practices a gating condition for public-sector adoption.
- Governance must be treated as a lifecycle process, not as a one-time label. AI systems evolve through fine-tuning, context shifts, and downstream reuse; therefore, “do no harm” cannot be credibly satisfied through licensing alone and requires ongoing testing, documentation, incident response, and mechanisms for contestation and redress.
- Equity depends on enabling conditions, not on openness alone. Stakeholders in lower-AI-readiness contexts placed relatively greater emphasis on cost-free access, localization (language and domain fit), and capacity-building; without shared compute, local evaluation capability, and sustainable stewardship models, openness can reproduce rather than reduce dependency and capability gaps.
The report proposes 10 recommendations organized around four priority areas: Standards, Accountability, Finance, and Equity (SAFE). Together, they seek to strengthen the governance of AI Digital Public Goods by promoting greater clarity and transparency, enhancing accountability, supporting sustainable and outcome-oriented financing approaches, and addressing capacity and infrastructure gaps that affect adoption, particularly in developing countries.
