A study by the Asian Development Bank published in June 2026 documents a massive gap between stated intentions and operational reality. Asian governments have multiplied AI systems presented as digital public goods, but very few meet the standards of openness, reusability, and accountability that give meaning to this label. When an administration delegates a decision to a shared algorithm, the question of responsibility in case of error remains without clear answer.

The Essential Points

  • The ADB study from June 2026 concludes that very few AI systems presented as digital public goods meet three cumulative criteria: genuine openness, reusability, and identifiable accountability.
  • The issue goes beyond technique: a “public good” label without independent audit or open license with maintenance obligations amounts to socializing costs while privatizing architectural choices.
  • The decisive question concerns the boundary between shareable models and responsibility: a common system without accountability mechanisms dilutes the chain of administrative responsibility.
  • Two trajectories are emerging for the decade 2026-2040: effective standards with interoperability, or fragmentation of closed ecosystems labeled without criteria.
  • Concrete pathways exist, notably independent audit frameworks and open licenses with maintenance obligations, but no Asian country has yet published a comprehensive reference standard.

The Real Definition of a Digital Public Good

The term is seductive. It evokes universal access, free provision, sharing among administrations. But the Asian Development Bank took the trouble to verify, in June 2026, whether AI systems claiming this designation actually meet the criteria that define this category. The UN definition, drawn from the 2020 Roadmap for Digital Cooperation and operationalized by the Digital Public Goods Alliance in nine indicators, covers open-source software, open data, open AI models, open standards, and open content that respect privacy, cause no harm, and contribute to sustainable development goals. It notably requires an open license, clear ownership, usable documentation, and independence from proprietary components.

The result is severe. A large majority of systems surveyed do not simultaneously satisfy these requirements. Some publish their code partially but without documentation usable by other administrations. Others guarantee access without defining who maintains the system or who is legally accountable for a bias detected after deployment. Still others operate as a black box for partner states that adopt them, trusting the integrity of the country or organization that developed them.

This diagnosis aligns with an observation that social science researchers have been making for several years about AI in public contexts: institutional categories tend to precede technical realities. Étienne Ollion, who has worked on social science methodology in the face of generative AI, emphasizes that institutional adoption of these tools is too often accompanied by less rigor than would be applied to other public policies—ex ante evaluation, precise definition of success indicators, review mechanisms. The label becomes a comfort response to a demand for legitimacy.

Openness Without Responsibility Creates No Trust

The tension at the heart of the ADB report lies in a fundamental asymmetry. Opening an AI system without defining its legal and operational responsibility amounts to sharing a tool of which no one is the owner in the legal sense. This can work for a statistical reference database or a geographic database, where error is documentable and correctable. An AI system that recommends granting or refusing a social benefit, building permit, or health authorization poses a different problem: the decision affects a natural person, and the causal chain between the algorithm and the final decision must be traceable.

The ADB study identifies several configurations that create a void of responsibility. The first is sharing between states without governance agreements: one country deploys a system developed by another, without access to code, training data, or the audit process. The second is deployment via international organizations that are not subject to the same accountability constraints as national administrations. The third is outsourcing to private actors who label their product “public good” while retaining intellectual property rights.

This last configuration deserves attention. It reproduces in public space a mechanism that Daron Acemoglu and Simon Johnson have analyzed in the private sector: technology is deployed in the declared interest of the greatest number, but architectural decisions that determine its effects remain in the hands of those who control it. Applying this framework to administrative AI reveals that the “public good” label can mask very ordinary capture: that of technical standards by organizations with resources to define them.

The Reasons Behind Asian Administrations’ Lag

It would be inaccurate to describe this situation as the result of ill will. It reflects real constraints. Developing a fully open, documented, auditable, and maintained public AI system costs significantly more than deploying an existing solution, even an incomplete one. For lower and middle-income countries that constitute a large part of the ADB’s study area, adopting a system developed by an international organization or partner country is often the only accessible path in the short term.

The problem is that this path creates lasting dependencies. An AI system deployed without access to source code or training data cannot be adapted locally to account for specific legislative, cultural, or linguistic contexts. It also cannot be independently audited by the oversight institutions of the country using it. The relationship between supplier and user then resembles a commercial licensing agreement more than a sharing of public good, regardless of the terminology employed.

Asia is not alone in this situation, but the region presents characteristics that make it particularly exposed. The diversity of legal, linguistic, and administrative contexts is extreme; what works for a social benefits management system in Seoul is not directly transferable to Dhaka or Vientiane. Investments in public digital capacity remain very unevenly distributed, as shown by the ADB’s work on digital infrastructure in the region. And political pressure to show rapid results in modernizing public services drives deployments that precede reflection on governance.

Algorithmic Responsibility as a Concrete Administrative Issue

Asking the question of responsibility in public AI systems is not a philosophical exercise. Administrations in several Asian countries already use decision support systems in domains that directly engage citizens’ rights: allocation of social housing, evaluation of tax risks, visa application screening, health care triage. In each of these cases, the question of who answers for a systematic error—the public agent who signed the decision, the organization that provided the model, the supplier who delivered the training data—often has no clear answer.

This void is not theoretical. American companies encountered similar difficulties with AI agent deployment, as illustrated by their recent experience. The public sector adds an additional dimension: the right to appeal. A citizen who contests an administrative decision has the right to know on what criteria it was made and to have it reviewed by a human. An opaque AI system, deployed under a sharing agreement without bias documentation, makes this right difficult to exercise in practice.

The ADB study emphasizes that existing international standards, particularly the framework of the United Nations Office for Digital and Emerging Technologies, define precise criteria for a system to be recognized as a digital public good. But adherence to these criteria remains voluntary, and independent verification mechanisms are underdeveloped in the region. The “public good” label currently functions more as a declaration of principle than as an auditable commitment.

The Coming Years Will Distinguish Between Approaches

Two trajectories are emerging for the coming decade, and choices made in the coming years will determine which prevails.

The first would be one of convergence toward effective standards. Pioneer Asian countries, though we cannot today definitively say which ones, would publish audit frameworks for their public AI systems, require open licenses with maintenance and responsibility obligations, and create interoperability mechanisms allowing genuine sharing among administrations. Regional institutions, beginning with the ADB itself, could play a standardization role by conditioning their financing on respect for measurable criteria. This scenario produces public digital infrastructure whose reliability increases over time, because errors are detected, documented, and corrected transparently.

The second trajectory would be one of closed fragmentation. Each state would continue deploying its own systems under the “public good” label, without convergence on openness standards or accountability mechanisms. Competition between suppliers, whether governmental, international, or private, would produce incompatible ecosystems, dependent on those who built them, and whose audit would remain impossible for using states. The promise of sharing would be nominal; the reality would be one of lasting technical dependencies.

These two trajectories are not equivalent for citizens. In the first, an inhabitant of a country that has adopted a decision support system for social benefits can in principle verify, via their administration, that the system applies rules that have been democratically defined. In the second, this verification is structurally impossible, regardless of the will of the public agents using the system.

The signals to watch are identifiable. The first would be that an Asian country publishes a comprehensive audit framework for its public AI—an event that has not yet occurred to date. The second would be a measurable increase in the number of systems simultaneously satisfying the three ADB criteria: openness, reusability, accountability. These signals are verifiable, and their absence or presence will say more about the real orientation of public policies than declarations of intent.

The Asian region is also observed by other areas facing the same trade-offs, as shown by parallel experiences in the Gulf, where public AI is mobilized for economic diversification objectives with very different governance. Regional comparison could, in the long term, provide natural quasi-experiments on the effect of governance choices on the actual quality of shared systems.

Requirements for Credible Standards

The ADB study formulates specific pathways, which deserve to be taken seriously as a starting point for a governance architecture. An independent audit framework for public AI systems would suppose that a third-party organization, endowed with technical expertise and institutional independence, could access source code, training data, performance metrics, and contested decisions. This is only possible if deployment was designed from the outset to permit such access, which is rarely the case today.

Open licenses with maintenance and responsibility obligations represent a middle model between closed ownership and provision without conditions. The principle is that of a contract: the organization publishing a system under this license commits to maintaining documentation, treating error reports, and designating an identifiable legal responsible party. This type of mechanism exists in other areas of free software, but its application to public AI supposes adding specific obligations related to impact on citizens’ rights.

These requirements remain achievable. They suppose investments in public, legal, technical, and institutional capacity that not all countries can finance alone. The ADB and other regional institutions have a role to play here: conditioning their support on respect for precise and verifiable standards rather than financing systems that self-label. Political pressure for rapid results risks not leaving time to build these foundations, and fragmentation could be consolidated before alternatives are put in place.

Sources

  1. Asian Development Bank, AI Systems as Digital Public Goods, June 2026, https://www.adb.org/publications/ai-systems-digital-public-goods
  2. UN Office for Digital and Emergent Technologies, https://www.un.org/techenvoy/digital-public-goods
  3. Digital Public Goods Alliance, definition and criteria for digital public goods, https://digitalpublicgoods.net/standard/
  4. UN ODET official press release on the report (25 June 2026), https://www.un.org/digital-emerging-technologies/fr/content/press-release-ai-systems-digital-public-goods-report
  5. UNU Publication page, AI Systems as Digital Public Goods, https://unu.edu/publication/ai-systems-digital-public-goods
  6. ArXiv preprint, AI Systems as Digital Public Goods, https://arxiv.org/abs/2607.03427
  7. DPG Standard, Digital Public Goods Alliance, https://www.digitalpublicgoods.net/standard
  8. UN ODET, About page, https://www.un.org/digital-emerging-technologies/content/about
  9. Étienne Ollion personal page, CNRS, https://ollion.cnrs.fr/
  10. Daron Acemoglu & Simon Johnson, Power and Progress, https://www.porchlightbooks.com/products/power-and-progress-daron-acemoglu-9781541702547
  11. ADB/UNU/UN ODET Report – AI Systems as Digital Public Goods (June 2026), https://www.adb.org/publications/ai-systems-digital-public-goods
  12. UN ODET – Press Release of 25 June 2026, https://www.un.org/digital-emerging-technologies/content/press-release-ai-systems-digital-public-goods-report
  13. UN – Roadmap for Digital Cooperation, June 2020, https://www.un.org/en/content/digital-cooperation-roadmap/
  14. DPGA – DPG Standard (9 indicators), https://github.com/DPGAlliance/DPG-Standard
  15. Acemoglu & Johnson – Power and Progress (2023), https://books.google.com/books/about/Power_and_Progress.html?id=4bumEAAAQBAJ
  16. ASEAN Guide on AI Governance and Ethics (2024), https://asean.org/wp-content/uploads/2024/02/ASEAN-Guide-on-AI-Governance-and-Ethics_beautified_201223_v2.pdf