- The paper provides the first comprehensive empirical analysis of designating AI systems as digital public goods through multi-stakeholder assessments.
- It employs rigorous desk research, interviews, and an expert survey to uncover challenges in aligning AI with existing Digital Public Goods standards.
- The study offers a modular governance framework and ten actionable SAFE recommendations to balance openness, accountability, and equity in AI deployment.
Artificial Intelligence Systems as Digital Public Goods: An Expert Analysis
Introduction
The research report "AI Systems as Digital Public Goods -- Evidence and Recommendations from a Multi-Stakeholder Assessment" (2607.03427) provides the first comprehensive, empirical investigation of the prospects and challenges of designating AI systems as digital public goods (AIDPGs). Produced by an interdisciplinary team at UNU Macau in collaboration with the Asian Development Bank and the United Nations Office for Digital and Emerging Technologies, the study deploys extensive desk research, multi-stakeholder interviews, and an expert survey to interrogate both the conceptual intricacies and practical barriers of aligning AI with the evolving Digital Public Goods (DPG) Standard.
Conceptual Foundation: DPG Theory and AI-Specific Complexity
Conventional DPG theory centers non-rivalry and non-excludability, operationalized in the digital domain through the DPG Standard: requirements include open licensing, platform independence, transparent governance, alignment with the SDGs, and a "do no harm" principle by design. However, AI systems challenge the sufficiency of this model. While open-source software, open content, and open data can often comply with these requirements, AI’s socio-technical complexity introduces new failure modes and dependencies. The opacity of training data, infeasibility of full algorithmic transparency, and the non-determinism and adaptability of modern ML systems defy straightforward “code is law” models from FOSS. Risks accrue at the point of deployment and downstream fine-tuning, not merely at model release. Therefore, the DPG Standard, as originally conceived, incompletely describes the requirements for public benefit and risk mitigation in AI.
Stakeholder Analysis: Openness, Public Value, and Licensing Tensions
The research advances three essential findings on openness, benefit, and accountability.
First, it demonstrates that openness in AI spans a multi-dimensional, component-based spectrum (architecture, weights, code, training and eval data, documentation) and is not simply a binary (open/closed) property. This is supported by both qualitative (stakeholder interviews) and quantitative (expert survey) data. While the OSI Open Source AI Definition (OSAID) opts for a binary threshold, practitioners and policymakers find this reduction incompatible with practical system-level governance, especially regarding managed access, partial openness, and data stewardship.
Second, the analysis isolates the distinction between system openness and realized public value. Survey data clearly expose that attributes such as transparency, responsible AI, privacy and data safety, inclusion, and SDG alignment do not necessarily co-vary. High degrees of openness are not, in and of themselves, predictive of societal benefit or risk minimization: for example, ungoverned open model weights can be trivially abused for disinformation or other adverse use cases. Stakeholders from lower-AI-readiness contexts prioritize cost-free access and localization, while those from advanced economies emphasize countering knowledge privatization and promoting system contestability.
Third, licensing remains a fractious bottleneck. No single license or certification captures the layered legal realities of code, model weights, and data; most public AI artifacts fall into non-OSI "open weight" or community licenses that embed substantial behavioral and branding restrictions, which are at odds with both the DPG Standard and strict OSI definitions. The RAIL license family attempts to instantiate public benefit through behavioral clauses but remains unenforceable at scale and thus not recognized for DPG compliance.
Lifecycle Governance and Distributed Responsibility
Governance of AIDPGs must be treated as a continuous, multilayered process over the full system lifecycle. The study proposes a modular governance framework: data provenance and ethical sourcing at the upstream data phase; independent assurance, artifact transparency, and carbon estimation at model development and release; incident response, human-in-the-loop mechanisms, and localized audit capacity at deployment. Critically, this modularity must be mapped onto multi-stakeholder structures that balance international reference standards (especially for developing countries with limited regulatory capacity) against the need for subsidiarity and local adaptation in high-AI-readiness settings.
Survey and interview results both reinforce that effective AIDPG governance cannot be achieved through developer self-attestation or top-down fiat alone. Responsibility must be explicitly and differentially assigned to developers, deployers, funders, state actors, and affected communities across the value chain, with clearly documented responsibility maps and accessible redress channels. Genuine citizen involvement is highlighted as a central governance hub, crucial for both legitimacy and risk calibration.
Strategic SAFE Recommendations and Implementation Pathways
The research outlines ten actionable recommendations, structured around Standards, Accountability, Finance, and Equity (SAFE):
- Standard: Adoption of the Model Openness Framework (MOF) as the reference vocabulary for operationalizing openness across the AI stack; creation of an annex of reporting indicators linked to measurable social, economic, risk, and sustainability criteria.
- Accountability: Establishment of formal responsibility maps across the lifecycle; the integration of redress and user-contestation pathways; and an extension of the DPGA's role in coordinating governance and certification support.
- Finance: Development of compute-access strategies for public-interest AI that leverage regional compute pools and multilateral cost-sharing; procurement and funding rules linked to ongoing evidence reporting rather than initial designation alone.
- Equity: Institutional support for the creation and stewardship of local-language and domain-specific datasets, especially respecting Indigenous data sovereignty; investment in local evaluation and audit capabilities as a condition for AIDPG adoption in developing country contexts.
The recommendations do not propose a full decoupling of AI systems from the DPG Standard but recognize the necessity of tailoring standard-setting and governance to AI’s distinctive risk and dependency structure. In particular, evidence requirements and stewardship must be ongoing, context-sensitive, and anchored to both SDG-aligned value and local legal and social frameworks. Importantly, the recommendations call for an explicit recognition that AIDPGs should be assessed as socio-technical systems, with a strong focus on impact and sustainability, not as static artifacts.
Implications and Future Directions
This work has significant implications for global AI governance, DPG certification bodies, implementers in the public sector, and multilateral donors. It provides an empirical and policy-theoretic foundation for why no major AI system has yet satisfied the DPG Standard and what bridging standards, infrastructure, and institutional mechanisms are needed. The modular, context-sensitive approach undercuts both simplistic open-washing and technologically exceptionalist regulatory proposals, insisting on fit-for-purpose, end-to-end governance and evidence.
Practically, the recommendations suggest that UN, MDB, and national strategies for AI public infrastructure must prioritize sustainable stewardship, locally grounded data resources, and, critically, transparent multi-actor responsibility, not just open licensing. Theoretically, the work leverages and extends longstanding debates on the governance of digital and knowledge commons, contributing a nuanced public-interest AI paradigm consistent with both public goods theory and recent developments in socio-technical systems governance.
Open questions remain about sustainable financing, the legal harmonization of licenses at different layers, the integration of AI-specific auditability with emerging sustainability (ESG) frameworks, and the capacity-building required for small-developer and public-sector actors, especially in low- and middle-income countries. The absence of a single global institutional home for AIDPGs governance means that ongoing experimentation, coordination, and iteration are necessary, both across UN subsystems and in major regional initiatives.
Conclusion
"AI Systems as Digital Public Goods -- Evidence and Recommendations from a Multi-Stakeholder Assessment" (2607.03427) provides a rigorous pathway for adapting the Digital Public Goods Standard to the realities of AI governance. It exposes the limits of existing open-source and DPG protocols for AI systems and produces evidence-based actionable guidance that recognizes openness as a necessary but insufficient condition for public value. The SAFE taxonomy, multi-level governance frameworks, and explicit focus on contextual equity and lifecycle accountability constitute an authoritative reference for global stakeholders seeking to operationalize AI as a public good.