Executive Summary
In July 2026, all 193 UN member states opened the first Global Dialogue on AI Governance in Geneva. What stood out was not the rules but the order of events. Before governments started arguing over how to control AI, they laid the preliminary report released on July 1 by a 40-member science panel across the floor of the negotiating table. This piece reads that scene not as a regulatory news flash but as a problem of governance's data infrastructure.
What the 40 produced was not a set of norms but a shared evidence base. Amandeep Gill, the UN's Under-Secretary-General for digital technology, called it "a shared scientific language that decision-makers will use together from here on." For anyone who works with data, the logic is familiar. There is no query without a schema, and no comparison without a common benchmark. Governance, too, cannot even begin to negotiate without common evidence.
And yet the evidence is open while enforcement is not. The report was made available to every government in six languages, but who actually gets access to frontier models runs on undisclosed criteria. This piece looks at both why the reversal of order is news and why the gap between open evidence and closed enforcement is the real problem.
The weight of this moment compresses into four numbers: how many countries sat at one table, how the panel that laid the floor was chosen, where the power the evidence points at is concentrated, and how widely that evidence was opened.
193 states
at the first dialogue
nearly every country at one table
40 members
science panel
selected from 2,600+ across 140 countries
90%
compute concentration
of compute controlled by just two states
6 languages
report released in
so every government reads the same evidence
This Time, Rules Didn't Come First
Until now, AI regulation has always started with rules. The EU AI Act sorted systems into risk tiers; national executive orders drew prohibitions and obligations first. The first Global Dialogue on AI Governance, held on July 6 and 7, 2026, at Geneva's Palexpo, reversed that order. Before the 193 member states bargained over rules, they took the preliminary report the 40-member science panel released on July 1 as a shared starting point.
Secretary-General António Guterres captured the shift in a sentence. "The science is here. We can no longer say we did not know. What we do with it is now up to all of us." The proposal is to agree first on a language for stating facts, before agreeing on rules. He added a short note for governments: don't wait.
Why does this order matter? Every country's philosophy of AI control differs. One leads with national security, another with market freedom, another with human rights and democracy. When conflicting principles collide and there is no common ground on what the facts are, dialogue ends with each side repeating its own claims. The 40-member panel laid that ground before anyone set the rules.
Reversing the order is the heart of this event. When rules precede evidence, negotiation becomes a question of who is stronger; when evidence precedes rules, it becomes a question of what is true. That 193 states chose the latter is, in itself, a bigger signal than any headline-worthy piece of regulatory news.
The 40 Built a Shared Evidence Base, Not Norms
The panel's formal name is the Independent International Scientific Panel on Artificial Intelligence (IISP-AI). Forty members were selected from more than 2,600 applications across 140 countries, and the UN General Assembly appointed them in February 2026. The co-chairs are Yoshua Bengio, who won the 2018 Turing Award, and Maria Ressa, the journalist and Nobel Peace Prize laureate. Selection weighed not only expertise but balance across regions, gender, and developing and developed countries.
What they produced was not rules but a list of measurable facts. The report found that today's safeguards are not keeping pace with the growth of AI capability. It noted, side by side, that there is currently no technical guarantee that agents performing multi-step tasks without human oversight will follow instructions, and that roughly 90% of the world's compute is controlled by just two countries. Bengio warned that AI capability is outrunning both scientific understanding and governments' capacity to adapt.
The dilemma the report points to is one of timing. Policy needs scientific evidence, yet by the time the evidence is certain enough, the window to act may already have closed. So the panel listed risks that are not yet fully understood. Sycophantic responses that reinforce a user's existing beliefs regardless of accuracy, along with misuse paths such as fraud, cyberattacks, and biological threats, were written down together. The judgment that waiting for certainty means acting too late is exactly what connects to this order of agreeing on evidence before rules.
Ressa's framing shows the report's character well. "What you are getting is the floor of our concerns, not the ceiling. Get 40 people from 40 different contexts to agree, and you don't drift toward the most alarming claims. Everything in this report cleared that threshold, and this is the minimum all of us agree on." The common denominator left after 40 people pared back their own perspectives is precisely the shared evidence base. The more contentious a finding, the more evidence and negotiation it had to survive.
2.1Translating "a shared scientific language" into the language of data
Put Gill's "shared scientific language" into the language of data and it looks a lot like a common schema. Just as different applications interoperate over the same data contract, different regulatory philosophies can only rebut or agree with each other on top of the same measurement and terminology layer. What the panel built was not the content of a norm but the floor a norm can rest on. What to prohibit is still undecided, but what counts as a dangerous fact is now written in a common language.
A shared evidence base is not a rule. It is the floor of measurement and terminology on which each country can lay its own philosophy of control. Without that floor, the words exchanged across the negotiating table become different languages; with it, parties can finally agree or disagree over the same sentence.
The Evidence Is Open, but Enforcement Slips In the Back Door
The openness of the evidence and the opacity of enforcement run in opposite directions. The report was released to every government in six languages. Who actually handles frontier models, and by what criteria, remains behind a closed door.
The US executive order of June 2, 2026, is one cross-section of this. It has developers voluntarily submit frontier models to the federal government before release, for up to 30 days of national-security and cybersecurity review, while explicitly ruling out mandatory licensing. The problem is that the criteria deciding which models fall under review are themselves classified. The "trusted partners" who receive early access are also selected privately by the government. From the outside, no one can tell who crosses the threshold, or where that threshold sits.
Critics call this regulatory capture without regulation — back-door licensing. Because the criteria are not published, the market can treat the government's selection as a de facto seal of approval. It runs in exactly the opposite direction from California and New York, which have required frontier developers to publish transparency reports. One side opens the evidence; the other closes the enforcement.
Having a shared evidence base does not make enforcement legitimate on its own. If the evidence is open to everyone while what it will be used to control runs on rules known to only a few, the gap grows sharper, not smaller. The real problem is not the absence of evidence but the distance between evidence and enforcement.
Whose Measurements Make the Rules Legitimate?
The most practical sentence in the panel's diagnosis is its point that existing governance is "fragmented, concentrated in a handful of companies, and rarely measures its own effectiveness." What is not measured is not governed. That 90% of compute sits in two countries, and that there is no guarantee an agent will follow instructions, both ultimately point to the same thing: the absence of a measurement and verification infrastructure.
For anyone working on data quality and model evaluation, this is not someone else's concern. For a rule to earn legitimacy it needs evidence, and that evidence exists only because someone shaped reality into something measurable. The reproducibility of benchmarks, the definition of risk metrics, the practice of model cards and transparency reports — these become the substructure of international norms. The legitimacy of governance rests, in the end, on which data you can trust.
So the question that remains is this. On whose data and measurements do the rules that will govern AI ultimately become legitimate? Rules become legitimate only when there is shared evidence, and building that evidence falls to the people who shape data into something measurable. The hand that writes the rules and the hand that builds the evidence are not different hands.
Why the UN laid evidence before rules is clear enough. A rule without evidence becomes a product of power; only a rule that stands on evidence can be verified. And that evidence does not arise on its own. Only when someone shapes facts into something measurable does it become the floor a rule can stand on.
Editor's note. Pebblous has long worked on the process of getting data ready to be used by AI — AI-Ready Data. The UN's move here shows that the same perspective reaches into governance too. The shared evidence a rule rests on does not arise on its own. Facts have to be shaped into something measurable to become evidence, and there has to be evidence for a rule to be legitimate. At the base of governance, too, lies a question of data quality.
References
Official Documents
- 1.Independent International Scientific Panel on Artificial Intelligence (IISP-AI). (2026, July 1). Preliminary Report of the Independent International Scientific Panel on Artificial Intelligence. United Nations.
- 2.United Nations. (2026). Global Dialogue on Artificial Intelligence Governance. United Nations. (Resolution A/RES/79/325)
- 3.UN News. (2026, July 7). UN launches global dialogue on AI governance in Geneva. UN News.
- 4.UN News. (2026, July 7). 'The science is here': UN chief welcomes first global AI assessment. UN News.
Regulatory Policy
- 5.Executive Office of the President, United States. (2026, June 2). Executive Order on Promoting Advanced Artificial Intelligence Innovation and Security. White House.
- 6.Herbert Smith Freehills Kramer. (2026). License to Model: US rules impact global access to frontier AI. Mondaq.
- 7.Biometric Update. (2026). Critics warn US AI executive order risks regulatory capture through selective model designation. Biometric Update.
News & Analysis
- 8.Technology.org. (2026). UN Science Panel Warns AI Is Outpacing Regulators. Technology.org.
- 9.Vectrel. (2026). What the Geneva Forum Means for Business. Vectrel.