Antarctica

A successful first dialogue would produce three things that do not yet exist: a shared diagnostic, a measurement commitment, and an institutional architecture for continuity. The shared diagnostic means acknowledging, formally and on record, that the current state of AI governance rests on an infrastructure that cannot yet verify its own claims. Regulatory frameworks across jurisdictions demand transparency, auditability, and emissions accountability from AI systems. None of those requirements can be satisfied at scale without independent measurement infrastructure that does not yet exist as a global standard. Naming this gap explicitly, not as a future risk but as a present condition, would itself be a substantive outcome. The measurement commitment means governments and relevant stakeholders agreeing that environmental and computational observability of AI systems is a governance prerequisite, not a voluntary disclosure. This does not require a binding instrument at this stage. It requires an agreement that the absence of standardized, independent measurement methodology is a structural obstacle to every other governance objective on the agenda, whether it is safety certification, human rights accountability, capacity-building equity, and interoperability. The institutional architecture means establishing a clear mandate for the Independent International Scientific Panel to address measurement methodology, not only safety benchmarks and capability evaluations, but the physical infrastructure of AI: energy, water, carbon, compute intensity, disaggregated by geography, model architecture, and deployment context. A dialogue that produces only a catalogue of positions will have been a missed opportunity. The question before governments is not whether to govern AI, but whether governance frameworks will be built on observable evidence or on vendor-reported proxies. The first session in Geneva is the moment to establish which standard the international community intends to hold.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches

These four areas share a common dependency that is rarely named as such: the existence of a reliable measurement layer for AI systems. Without it, each of them risks becoming a governance aspiration without operational content. Safe, secure and trustworthy AI cannot be certified without observable evidence of system behavior at the infrastructure level. Current safety frameworks evaluate model outputs and training processes. They do not independently measure the physical conditions under which inference runs : energy draw, thermal management, hardware utilization, which are material to both reliability and environmental accountability. Transparency, accountability, and human oversight face the same structural obstacle. When an enterprise or government deploys an AI system via a third-party API, it has access to pricing and outputs, not to the underlying compute and energy expenditure. Meaningful human oversight requires that the systems being overseen are legible. AI infrastructure, as currently deployed globally, is not legible to the entities legally responsible for its consequences. The social, economic and environmental implications of AI are being systematically underestimated because the physical costs of AI (energy, water, carbon) are measured by providers using methodologies they design, and disclosed selectively. Independent, standardized measurement would change the terms of the debate. The environmental cost of a given AI deployment is not a fixed number; it depends on geography, grid carbon intensity, model architecture, and inference pattern. These variables require instrumentation, not estimation. Interoperability of governance approaches requires a common evidential base. Jurisdictions cannot align on accountability standards if they are working from incompatible or unverifiable data. Measurement interoperability is a precondition for regulatory interoperability. Antarctica's One-Token Model, peer-reviewed by IEEE and aligned with eight international standards bodies, is a working example of what independent AI measurement infrastructure can produce at the token level.

The most significant cross-cutting issue absent from the listed themes is the measurement infrastructure deficit, the structural gap between what AI governance frameworks require in terms of evidence and what current industry practice produces. Every thematic cluster in the Dialogue assumes some degree of measurability: safe AI assumes verifiable safety properties; transparent AI assumes observable behavior; equitable AI assumes comparable impact data across geographies. None of these assumptions currently holds at the infrastructure level. Enterprises, governments, and civil society organizations deploying AI systems do not have access to independent, granular, standardized data on the energy consumption, carbon emissions, or water use attributable to their AI workloads. They have estimates, proxies, and vendor-reported aggregates. This is an architectural problem: the measurement layer has not been built. Two dimensions of this deserve specific attention. First, geographic carbon intensity differentials make the same AI workload orders of magnitude more carbon-intensive depending on where inference runs. A model served from a coal-heavy grid may carry a carbon cost thirty times higher than the same model served from a low-carbon grid. Current governance discussions do not reflect this differential in any systematic way. Second, the concentration of model architecture knowledge in a small number of private entities creates an asymmetry that undermines independent measurement efforts. Governance frameworks that depend on provider-disclosed specifications are structurally dependent on the entities they are meant to govern. The Dialogue should consider recommending that the International Scientific Panel be mandated to develop a reference methodology for independent AI environmental measurement (covering energy, water, and carbon) that can serve as a common evidential base across jurisdictions. Such a methodology would not replace national regulatory frameworks; it would make them coherent with one another and defensible against scrutiny. This is the missing infrastructure layer of AI governance.

The governance gaps we have identified are not abstract. They manifest concretely in two countries where our work is grounded: France and India. Together, they illustrate why the measurement deficit is not a technical footnote but a structural obstacle to equitable and effective AI governance. In France, the regulatory ambition is genuine. CSRD imposes auditable Scope 2 emissions reporting. The AI Act introduces transparency and accountability obligations. Yet enterprises subject to these frameworks cannot satisfy them with actual measurement data for their AI workloads. They respond with estimation methodologies that vary between organizations, are unverifiable by auditors, and are structurally dependent on provider-disclosed figures that lack independent validation. Regulatory intent and measurement reality are misaligned. France is attempting to govern a physical infrastructure it cannot yet observe. In India, the challenge is different in character but equivalent in consequence. India is deploying AI at scale (in financial services, agriculture, public administration, healthcare) on a grid whose carbon intensity is among the highest of any major AI-deploying economy. The same inference workload that carries a modest carbon cost in France, where nuclear generation dominates, may carry a cost thirty times higher when run on Indian grid electricity. No current governance framework accounts for this differential systematically. The result is that enterprises and governments in high-carbon-intensity regions carry an environmental liability they cannot measure, report, or reduce. The opportunity that both contexts reveal is the same: a jurisdiction that builds independent, standardized AI measurement infrastructure first will hold a structural advantage in every subsequent governance negotiation, bilateral, multilateral, or sectoral. France is positioned to propose such a standard at the international level. India is positioned to demonstrate its necessity at scale. The combination is more powerful than either alone.

The AI Dialogue occupies a position that no existing forum does: it is universal in membership, mandated by the General Assembly, and explicitly multi-stakeholder in design. That combination is rare and should be used deliberately. The most valuable role the Dialogue can play is not to produce another catalogue of principles (those exist in abundance) but to establish the evidentiary foundations that make principles enforceable across jurisdictions. International cooperation on AI governance is currently constrained less by disagreement on values than by incompatibility of evidence. Jurisdictions that wish to align on accountability standards find that they are working from different data, produced by different methodologies, under different disclosure requirements, with different levels of provider cooperation. Coherence is structurally difficult when the measurement layer beneath governance frameworks is neither standardized nor independent. The Dialogue can address this by doing three things that bilateral or regional processes cannot easily do. First, it can legitimize the demand for independent measurement as a governance prerequisite rather than a voluntary industry commitment. Second, it can mandate the International Scientific Panel to develop a reference methodology for AI environmental and computational observability that is jurisdiction-neutral and open to adoption. Third, it can create the connective tissue between governance initiatives that currently operate in parallel (the OECD AI Policy Observatory, the Global Partnership on AI, the ITU AI for Good framework, national regulatory bodies) by providing a common evidentiary baseline against which each can calibrate. The risk the Dialogue must avoid is becoming a venue for the reiteration of existing positions. Its added value is structural: it is the only forum where the demand for a shared measurement infrastructure can be established as a global norm rather than a bilateral preference. That is the mandate it should pursue in Geneva.

Several existing initiatives have built partial foundations that the Dialogue should recognize, connect, and extend rather than duplicate. The OECD AI Policy Observatory has produced the most comprehensive comparative mapping of national AI strategies and regulatory approaches. Its value to the Dialogue is taxonomic: it provides a structured vocabulary for governance comparison. Its limitation is that it operates at the policy layer and does not descend to the measurement infrastructure beneath policy commitments. The Green Software Foundation has developed foundational work on software carbon intensity, including the Software Carbon Intensity specification and the Carbon Aware SDK. These represent serious technical contributions to the question of how software systems, including AI, should account for their environmental impact. The Dialogue should explicitly recognize this work and consider how it can be elevated from an industry initiative to an international reference. The ITU AI for Good framework provides the institutional bridge between the AI governance conversation and the technical standardization community. Its co-location with the Geneva session is an opportunity that should be used structurally, not merely logistically. Standards developed through ITU processes carry implementation weight that policy declarations do not. The work of the IEEE on AI measurement methodology (including peer-reviewed frameworks for AI energy and emissions quantification) represents the academic and technical credibility layer that governance frameworks need to cite with confidence. What the Dialogue can add to all of these is normative authority and universality. Each of the initiatives above operates within a defined community: industry, technical, academic, or OECD-member. The Dialogue can establish that the outputs of these communities are not optional inputs to governance but necessary preconditions for it. It can also ensure that the measurement standards developed within these communities are designed to serve all countries, not only those with the technical capacity to have participated in their design.

The format should be designed to prevent a familiar dysfunction: governments state positions, industry presents capabilities, civil society raises concerns, and the three streams never intersect. Structured interrogation produces more useful outputs than prepared statements. Governments should contribute diagnostically: what specific governance obligations are they currently unable to enforce for want of measurement data? Industry stakeholders should respond to specific technical questions rather than present prepared positions : what data do they collect but not disclose, and what would independent auditability require? Civil society and academic contributors are most valuable when presenting evidence from deployment contexts that industry and government do not observe: affected communities, independent infrastructure researchers, capacity-building practitioners in developing economies. One structural gap deserves attention. Deeptech and measurement-focused companies occupy a category the current stakeholder architecture does not accommodate well. They are neither hyperscalers nor civil society. They hold technical knowledge directly relevant to governance questions (independent measurement methodologies, infrastructure observability tools, standards-aligned frameworks) and should have a dedicated channel to contribute that knowledge to scientific and policy workstreams. The test of an effective format is whether it produces evidence that did not exist before the Dialogue convened.

Three categories are structurally underrepresented, for different reasons and with different consequences. The technical measurement community (researchers, engineers, and companies working on AI infrastructure observability, energy attribution, and environmental quantification) holds knowledge directly material to every governance question on the agenda. Yet they are rarely present at policy forums. Actively recruiting independent measurement practitioners into working sessions would change the quality of evidence available to policymakers. Enterprises and governments in high-carbon-intensity economies face an AI governance challenge qualitatively different from low-carbon economies. The same deployment decision carries vastly different environmental consequences depending on where inference runs. India, Indonesia, South Africa, Nigeria, large, fast-growing AI-deploying economies with high-carbon grids, should have explicit representation in discussions on environmental implications and capacity-building equity. This geographic dimension is almost entirely absent from current governance conversations. Practitioners building AI under severe resource constraints (limited compute, unreliable connectivity, inadequate local-language training data) experience AI governance not as over-regulation but as structural exclusion. Their testimony is essential to any honest discussion of AI divides.

Three formats would produce substantive outputs by forcing specificity and resisting prepared statements. Structured technical hearings, modelled on parliamentary committee practice, would pose specific questions to pre-selected respondents (a cloud provider, an independent measurement researcher, a regulator, an enterprise deploying AI at scale) with cross-questioning permitted. This produces a record of positions under scrutiny rather than positions under stage lighting. A live measurement demonstration session would have independent technical actors instrument a real AI workload in real time (measuring energy consumption, carbon intensity, and cost at the token level) and present the output to the plenary as an existence proof that independent measurement infrastructure is buildable, and that the gap between what it produces and what governance frameworks require is visible and specific. A structured gap-mapping exercise in breakout sessions would ask participants not to advocate for preferred governance approaches but to identify the single most specific thing they cannot currently do for want of a shared standard. The output would be a map of operational governance gaps by jurisdiction and sector, serving as concrete input to the Co-Chairs' summary. Format should serve evidence over assertion, and specificity over principle.

The most instructive examples of effective AI governance are those that have moved beyond principle to instrumentation, where accountability is produced by measurement infrastructure rather than declared by policy intent. The Green Software Foundation's Software Carbon Intensity specification establishes that carbon attribution is a technical discipline with defined inputs, assumptions, and verification requirements. Its limitation is that adoption remains voluntary and implementation varies significantly across organizations. France's CSRD implementation offers a contrasting lesson. By making Scope 2 emissions reporting a legal obligation with audit requirements, it has forced enterprises to confront the gap between what regulation demands and what measurement infrastructure can currently deliver. That confrontation is productive: it has generated demand for independent measurement tools and created regulatory pressure on providers to improve disclosure. India's investment in sovereign compute capacity and indigenous model development demonstrates that governance is also an infrastructure question. Countries that do not control the physical layer of AI deployment cannot fully govern its consequences. At the methodology level, IEEE peer-reviewed frameworks for AI energy and emissions quantification demonstrate that independent measurement of AI infrastructure impact is technically achievable at the inference level, providing the granularity that enterprise compliance and international policy accountability both require. Effective governance builds on what can be independently verified.