From the AI Accountability Gap to a Global Accountability Infrastructure

AI production is highly concentrated in a few companies and countries that have been shaping both technologies and governance processes to date. They are also the main interested parties in the advance of AI within regions and sectors under their own conditions. Power asymmetries include the dependency, in particular, of Global South countries on corporate actors that activate multiple strategies to influence how their integration in the AI economy will take place.

While the benefits of AI are concentrated in a few, its footprint is distributed. As highlighted by several participants during the recently concluded first UN Global Dialogue on AI Governance, AI implies a supply chain that goes from mining to data processing and requires an immense amount of natural resources to operate. Its impacts to the environment, labor and broader human rights are structural and affect workers, communities and territories that are often far from the main beneficiaries of the AI economy.

The AI accountability gap

This scenario implies a critical global accountability gap, with multiple dimensions. A jurisdictional one, derived from the distributed character of tech supply chains and corporations’s attempt to escape compliance, relying on the absence of legal presence in certain regions and countries. A regulatory one, which accounts for the variation in regulatory and institutional capacities and characteristics within different contexts. Connected with the jurisdictional gap, this means that the very same AI system can face rigorous scrutiny in one country and almost none in another. It also allows jurisdiction shopping, with companies intentionally using the regulatory gap as an advantage to escape accountability at the supply chain level.

Another dimension of the global accountability gap is in relation to power. Large technology companies have economic, legal, financial and political resources superior to some countries. This means that many governments, especially in the Global South, negotiate with them from a much weaker position and that the communities most affected by AI harms are often those with least access to remedies, representation or justice.

Lastly, there is an evidence or informational gap. While AI risks are often presented as part of futuristic science-fiction-like scenarios, harms in multiple areas have already been widely documented. Examples include AI-enabled gender-based violence, predictive policing, facial and emotion recognition, among others. However, there is no unified source of information about such supply-chain level harms.

The Scientific Panel on AI summarized the accountability gap as a political economy challenge of high market concentration: “while some jurisdictions benefit from robust privacy and data protection laws, if deployed outside guardrails, concentrated AI capacity raises concerns about impacts on democracy and on human rights, along with possible regulatory capture and lack of accountability”.

The acknowledgement of global inequalities, market concentration, supply-chain level impact and accountability gaps represent a key shift of narrative within global AI governance spaces and is by itself a meaningful accomplishment of this Dialogue. However, the remaining question is how to move forward from diagnosis to action in the run-up to its second edition in 2027. Addressing environmental, labor and broader human rights impacts within the transnational AI supply chain is an urgent task. And the international system has an essential role to play in closing the global accountability gap.

Closing the gap

While there is much to be done to address AI impacts to human rights, the normative foundation for doing so already exists. Over the past few years, UN agencies and the UN human rights system have consistently affirmed that AI systems must comply with international human rights law throughout their lifecycle, something also recognized by the General Assembly since 2024. Where compliance is impossible, the appropriate response is not mitigation but prohibition or moratoria. In other words, States and other stakeholders should refrain from or cease the use of such systems.

What is needed now is to translate these commitments into governance practice across jurisdictions and throughout the AI value chain. This requires an international accountability architecture built on at least two complementary pillars.

The first one is already being shaped. The Independent International Scientific Panel on AI was created to assess evidence, identify areas of scientific consensus and uncertainty, and provide policymakers with an independent basis for decision-making. Much like the Intergovernmental Panel on Climate Change (IPCC), it can establish a shared evidentiary foundation on which governments can debate policy choices.

This function is indispensable. Decisions about whether particular AI systems or uses meet the international human rights law threshold should not depend solely on corporate narratives or shifting political priorities. Yet scientific consensus alone cannot sustain accountability. Science moves carefully, but communities and the planet are experiencing harm at a different—and accelerated—pace.

In other words, scientific assessment cannot capture the full spectrum of AI harms. Many impacts are contextual, emerging only when AI systems interact with specific social, political or environmental realities. The same technology may generate profoundly different consequences depending on where and how it is deployed. A data center, for example, will have environmental impacts wherever it is built, but the nature and severity of those impacts depend on the local context. In a region already facing chronic water scarcity, or where Indigenous Peoples and traditional communities depend directly on local ecosystems for their livelihoods, it will probably significantly intensify existing pressures.

Likewise, AI systems may reinforce existing patterns of discrimination, labor exploitation or political repression in ways that only become visible through local experience. These harms rarely emerge first in global datasets or scientific reviews. They are identified primarily by the communities affected by them.

That is why the emerging AI governance architecture still lacks a second, equally important pillar: a permanent international mechanism for documenting AI-related harms.

Such a mechanism should not replace scientific assessment. On the contrary, it complements it by creating institutional pathways through which evidence generated by or with affected communities can inform global governance before academic consensus is reached.

Civil society organizations, investigative journalists, grassroots movements, trade unions and independent researchers, among others, are often the actors best positioned to identify emerging harm given their proximity to local contexts. However their work is often under-resourced and disconnected from international decision-making.

Strengthening it does not necessarily require creating new institutions, but better connecting and reinforcing existing ones. This means providing these actors with sustainable, independent financial and institutional support, alongside shared documentation standards and methodologies, while creating formal channels through which the evidence they produce can be systematically assessed, incorporated into international processes and used to inform decision-making.

The UN human rights system itself must also become a central component of AI accountability. The Universal Periodic Review, treaty bodies, special procedures and the Human Rights Council, for instance, already investigate human rights violations, receive evidence and issue authoritative recommendations. Rather than duplicating these mechanisms, AI governance should reinforce them, providing adequate resources, stronger mandates and systematic coordination so that AI-related harms become a permanent part of their work.

Finally, at the same time, AI governance must recognize that AI is not a sector but an enabling technology whose impacts cut across several areas such as labor, health, education, the environment, democracy and development. Accountability around this matter should therefore not be confined to AI-specific institutions. The Scientific Panel, the Global Dialogue and the broader accountability architecture should work in close coordination with institutions such as the International Labour Organization, environmental governance bodies and the Committee on the Elimination of Discrimination against Women.

Together, these pillars could form the basis of a genuine global accountability infrastructure.

The Global Dialogue intersessional opportunity

Human rights standards, including on social and economic rights, are the main basis on which to build consensus around redlines. Encouragingly, calls for international AI red lines have gained significant momentum over the past few years, bringing together civil society organizations, researchers, former heads of state and other experts. As this conversation matures, greater attention should be given to the practical question of how these commitments can be operationalized and made meaningful in practice.

The first Global Dialogue on AI Governance has created a space for this discussion to happen. The challenge now is to translate such political momentum into an accountability architecture capable of documenting harms, informing decision-making and supporting collective action across borders. The period leading to the second Dialogue in 2027 offers a unique opportunity to begin doing exactly that.

This article builds on reflections presented by the authors during the first UN Global Dialogue on AI Governance, including interventions by Jamila Venturini in the official Dialogue and by Marina Meira in parallel events.