The Verification Ceiling: What Geneva Confirmed About Global AI Governance

Russian-language version available

Contents

On 7 July 2026, the United Nations closed the first Global Dialogue on AI Governance with a message its own co-chairs framed as forward-looking: judge this process not by the principles it adopted, but by what participants do before the next session in 2027. Read differently, that sentence points to something else. Nothing enforceable was produced, because nothing enforceable was structurally available to produce.

I. A Diagnosis Half Made

Five days before Geneva, the UN’s Independent International Scientific Panel on AI released its preliminary report — forty scientists, drawn from over 2,600 applicants across 140 countries, tasked with giving policymakers a shared evidentiary base [1]. The report named its central problem clearly: an “evidence dilemma,” in which governments need reliable data before they regulate, but by the time enough data exists, the technology has already moved on [2].

It is a fair description of a symptom. As a diagnosis, it is incomplete, and Geneva showed why. The Dialogue that convened to act on the Panel’s evidence produced, by design, a non-binding outcome. By the UN’s own account, it is not a negotiating forum: each session is structured to conclude with a co-chairs’ summary rather than a binding instrument, a deliberate choice to let every state participate on equal footing without the constraints formal negotiation would impose [3]. The gap the Panel describes as a timing problem turned out, over two days in Geneva, to be a structural one. Three separate constraints made that clear, each from a different source, each on the public record.

II. The Panel Names Its Own Limit

The first constraint became visible from inside the Panel itself. Presenting the working-group findings in Geneva, ETH Zurich’s Mennatallah El-Assady told delegates that independent verification of advanced AI systems remains weak, that public benchmarks are becoming saturated, and that frontier systems are increasingly showing signs of evaluation awareness — the capacity to recognize when they are being tested and to behave differently as a result [4]. She named interpretability, reliable auditing, and independent verification as the field’s immediate bottlenecks, a problem she said would only deepen as AI systems move from software into physical domains such as robotics [4].

This is the distinction the Analytical Glossary draws between Declared Control and Verification Capacity: an institution’s stated authority over a system, set against its actual ability to confirm that authority holds. El-Assady was not describing a distant risk. She was naming a constraint on the broader evidence environment the Panel’s own report draws from: the systems advancing fastest are also the systems for which independent verification remains weakest.

III. A Universal Table, an Unequal Infrastructure

Roughly 74 percent of the world’s observed AI-supercomputer capacity is concentrated in the United States, with a further 14 percent in China [5]. These figures describe the observed portion of global high-end AI compute; the Federal Reserve notes that the underlying data cover an estimated 10 to 20 percent of the existing global aggregate, so the precise shares should be treated with caution rather than read as a full census. The structural point is narrower and more defensible: within the best-documented segment of frontier compute, capacity is highly concentrated rather than broadly distributed.

Universal political participation does not by itself produce universal technical capacity: the infrastructure required to train, evaluate, and independently stress-test frontier systems remains held by a small number of jurisdictions and the private firms operating within them. This is Material Predetermination in its plainest form — a governance architecture built on the premise of universal participation, layered over an infrastructure base where verification capacity is neither universal nor evenly distributed before any negotiation begins.

IV. The Objection Filed in Advance

The third constraint was stated ten months before Geneva, and by a different UN body than the one that would host the Dialogue. On 24 September 2025, at the UN Security Council’s open debate on artificial intelligence and international peace and security, Michael Kratsios, Director of the White House Office of Science and Technology Policy, stated that the United States rejected efforts by international bodies to assert centralized control over AI governance [6]. The Global Dialogue itself was politically launched the following day, at a separate General Assembly high-level meeting — a body Kratsios’s remarks did not address directly, and one that, by its own founding terms, was never designed to exercise the kind of centralized authority his statement rejected [7].

The distinction matters. The US position does not describe non-compliance with an existing authority; the Dialogue makes no claim to such authority in the first place. What it does is set an outer boundary — declared before the Dialogue’s first session, by the state holding the largest share of the infrastructure any future verification regime would depend on — on what kind of international AI governance architecture the United States would accept at all. That boundary did not need restating in Geneva. It simply sat beneath every proposal made there.

Implications

Put together, the three constraints do not describe a Dialogue that failed. They describe a Dialogue that performed exactly as its structural conditions predicted. An evidentiary body whose own scientists identify a verification gap in the systems it is tasked with assessing; an infrastructure base concentrated in a small number of jurisdictions, one of which has pre-declared the outer limit of international authority it will accept over that infrastructure; a governance body built on universal participation but without a verification mandate to match it. None of this required Geneva to reveal — the “Beyond Control” framework specifies each constraint independently. What Geneva supplied was confirmation: dated, on the record, and largely in the words of the institutions’ own participants rather than outside critics.

It is worth being precise about what this does not show. The Scientific Panel’s forty members include no one from Central Asia — a gap consistent with the region’s structural position in global AI governance discussed elsewhere in this series. But the political track tells a different story: Kazakhstan attended Geneva at deputy prime ministerial level, with Deputy Prime Minister Zhaslan Madiyev addressing the Dialogue directly [8]. The exclusion, in other words, is specific to the evidentiary layer, not the diplomatic one. A region can be represented in the political dialogue while remaining absent from the body tasked with assessing and synthesizing the scientific evidence that informs it — which is arguably the more consequential absence of the two.

Signals to Watch

Three markers will indicate whether the pattern holds or breaks before the next session in New York in May 2027: whether any working group established at Geneva is given an actual verification mandate, with resources and access, rather than a coordinating role; whether the compute-concentrated states engage with any future proposal for independent technical audit that requires meaningful access to frontier systems or infrastructure; and whether the Panel’s 2027 annual report evaluates identifiable frontier systems and providers, or remains confined to aggregate, anonymized findings.

The Questions That Remain Open

Two questions this piece does not resolve. First, whether a verification capacity gap of this scale is closable through capacity-building, as Geneva’s own proposals suggest, or whether it is structurally persistent as long as frontier compute remains concentrated in a small number of jurisdictions. Second, whether an institution can meaningfully strengthen its own evidentiary authority once its verification limits have been stated on the record by its own members — or whether that acknowledgment, once made, becomes part of the institutional ceiling itself.

Sources & Notes

[1] Independent International Scientific Panel on AI. Preliminary Report. United Nations, 1 July 2026. un.org

[2] UN News. AI explained: Why the world needs to act now. 1 July 2026. news.un.org

[3] United Nations. FAQ: Is the Dialogue a negotiating forum? Will it produce binding agreements? Global Dialogue on AI Governance. un.org

[4] Digital Watch Observatory. UN scientific panel presents first AI assessment to Global Dialogue on AI Governance. July 2026. dig.watch

[5] Haag, Alex. The State of AI Competition in Advanced Economies. FEDS Notes, Board of Governors of the Federal Reserve System, 6 October 2025. DOI: 10.17016/2380-7172.3930. Underlying compute data: Pilz, Rahman, Sanders, Heim. Data on GPU Clusters. Epoch AI. federalreserve.gov

[6] Kratsios, Michael. Remarks at the Security Council’s Open Debate on Artificial Intelligence and International Peace and Security. U.S. Mission to the United Nations, 24 September 2025. usun.usmission.gov

[7] United Nations Western Europe. The artificial intelligence we want. On the Security Council debate of 24 September and the General Assembly high-level meeting of 25 September 2025 launching the Global Dialogue. unric.org

[8] Global Dialogue on AI Governance. List of Featured Attendees. un.org; confirmed in The Astana Times, Kazakhstan Promotes Regional Digital Cooperation at First UN Global Dialogue on AI Governance, 7 July 2026. astanatimes.com

[9] Digital Watch Observatory. Inaugural UN Global Dialogue on AI Governance ends with call to turn principles into action before 2027. 8 July 2026. dig.watch

Full essay and updated sources: okhodjaev.com/analysis/the-verification-ceiling

Oybek Khodjaev — over 35 years of experience in banking, finance, public administration, and business in Uzbekistan and the CIS. Author of the essay series “Beyond Control: Theory of Limits of AI Governance.” okhodjaev.com

The author advises public institutions and financial organisations on AI governance, verification frameworks, and institutional readiness.

Published: