June 23, 2026
Navigating Alignment in AI Systems as DPGs
Navigating Alignment in AI Systems as DPGs
During this year’s UN Open Source Week in New York, many discussions will address AI as both a significant opportunity and a challenge for digital public goods (DPGs) and the wider open-source ecosystem. Simultaneously, the debate of the characteristics of AI systems as DPGs continues, transcending the DPG Standard while rooted in OSI’s definition of open-source AI. The G7 also recently published its vision on openness in AI systems, including a description of elements that determine the different levels of openness. The fact is that today there's no global consensus on what openness in AI systems actually means.
At the DPGA Secretariat, we are focused on operationalising the commitments of the Global Digital Compact (GDC) to “develop, disseminate and maintain, through multi-stakeholder cooperation, [...] open data, [and] open artificial intelligence models” to help achieve the SDGs by 2030. However, translating these high-level ambitions into operational practices is part of a complex implementation process.
Welcoming this ongoing debate as a vital step forward, we turn our attention to the new report commissioned by the DPGA member, the Asian Development Bank: “AI Systems as Digital Public Goods: Evidence and Recommendations from a Multi-Stakeholder Assessment”. The report was produced by the United Nations University Macau in partnership with the UN Office for Digital and Emerging Technologies (UN ODET), and arrives at a critical time. It provides valuable input into our ongoing standard-setting process as we review how the open data clause of the DPG Standard applies to AI systems one year after its last update.
Strategic Alignment with the DPG Standard and the DPGA Secretariat’s Work
When the DPGA Secretariat updated the DPG Standard for AI systems in 2025—following extensive consultations within our AI Community of Practice co-hosted with UNICEF—we deliberately set a high, aspirational bar. We wanted to move the conversation away from treating AI as an isolated technology, toward a holistic view of AI as a socio-technical system.
The now-published “AI Systems as Digital Public Goods” report organises its recommendations around the SAFE mnemonic — Standard, Accountability, Finance, Equity — and is candid that several of its recommendations fall beyond the DPGA Secretariat's current mandate. These are still valuable precisely because they are ecosystem-level responsibilities, shared across multilateral development banks, donors, governments, and standard-setters alike.
Here is how those recommendations map against the DPG Standard as it stands today, and where we either already meet the suggested requirements or see things differently.
DPG Standard Recommendations
| Report Recommendation | Alignment | DPGA Secretariat Workstreams |
|---|---|---|
| S1 — Adopt the Model Openness Framework | Recommended MOF as a descriptive vocabulary; the DPG Standard requires its most demanding Open Science tier as a threshold, including open training data. | MOF was a primary reference for the 2025 Standard update; we use it to structure the data, code, and model components. |
| S2 — Data stewardship pathway | Proposed governed/gated access for sensitive data; the DPG Standard currently requires that clear ownership be stated, and its reuse or distribution rights. | The DPG Maturity Framework builds on this to more clearly state governance and contribution models, but does not cover gated/closed data. |
| S3/S4 — Deployment assurance profile & public-value indicator annex | Call for documentation of intended use, data provenance, known limits, context-specific evaluation, and minimum safeguards, among other requirements. | These are already built in via mandatory model cards, datasheets, and the recommended AI risk assessment DPG documentation requirements. |
Ecosystem-Level Requirements
| A1 — Consolidated governance support structure | Extending the existing AI DPGs ecosystem with members' collaboration, including ADB, UN ODET, UNESCO, and ITU, and not building a parallel architecture to avoid duplication of efforts. | The DPGA serves as a neutral convener, standard-setter, and discoverability infrastructure via the DPG Registry and its community. |
| A2 — Responsibility maps and redress | Full value-chain responsibility mapping is an ecosystem-level endeavour. The DPGA members can play a role in various aspects of this chain, including licensing, preference signals, and procurement requirements. | The DPGA Secretariat supports its members' initiatives, such as the CC Signals project, and the call for human knowledge as critical digital infrastructure, and also works with product owners to define best practices for the use of open data DPGs in AI. |
| F1/F2 — Compute access and outcome-linked funding | MDBs, donors, and governments' requirement to co-finance portable compute facilities to mitigate vendor lock-in. | DPGA emphasises resource-efficient approaches to AI, focusing on smaller, task-specific, local open models that provide auditability and cost control. |
| E1/E2 — Local-language data and local evaluation capacity | Address the scarcity of local training data, local language data collection, and annotation by managing it as critical digital infrastructure and a public utility, and support the creation of regional AI evaluation centres and university-level public interest tech curricula. | The DPG Standard does not mandate localisation; our upcoming DPG for AI Collection addresses this through specific inclusion & autonomy considerations. With the collection, we aim to highlight solutions relevant to Global Majority contexts. |
Where We See Things Differently: Openness As a Threshold, Not a Spectrum
The ADB-commissioned report's main recommendation for openness is to use the Model Openness Framework (MOF) as a shared vocabulary for explicitly describing degrees of openness, "rather than as a definitive pass/fail measure," paired with a stewardship pathway that would allow governed access to sensitive data. This is where the suggested approach most clearly diverges from the current DPG Standard.
As mentioned, we also use the MOF, but not as a gradient. The DPG Standard sets a threshold: an AI system either is or isn’t a DPG, and that line sits at the most demanding rung, requiring open training data alongside open code and weights. That was a deliberate call, as we explained when the Standard was updated: training data is the input that keeps SDG relevance, platform independence, and do-no-harm meaningful, and a more permissive approach would quietly chip away at these core premises for digital public goods.
The trade-off is real: a high bar means fewer AI systems qualify right now, and some genuinely useful systems trained on sensitive data won't meet it today. The DPGA Secretariat, in consultation with experts, strongly believes that in the current landscape, the answer isn't to lower the bar pre-emptively; it's to build the data governance and stewardship tools that would let the DPG Standard evolve responsibly, which is exactly the question we're revisiting one year on. In that sense, we welcome the new report as a valuable input.
"Having taken part in the report consultations, what I find genuinely exciting is how it’s aiming for the same objective as the DPG Standard — AI that's actually governed in the public interest, not just released under an open licence. We take different routes in a couple of places, but honestly, those are the parts where the interesting work for the DPG ecosystem will happen." — Ricardo Mirón Torres, Chief Technology Officer, DPGA Secretariat
The Way Forward
Operationalising the Global Digital Compact is a collective journey. The DPGA Secretariat invites the entire open-source and DPG ecosystem to actively follow our work on the DPG Standard for AI and the upcoming DPG for AI Collection via GitHub. Join us in supporting our members' and DGP product owners' initiatives, and help us collectively build an equitable, sovereign, and rights-preserving digital future.
