The council opened, and the evidence went first

A council that convenes to discuss a power nobody fully controls has one job before it can do anything else: agree on what is actually true. Geneva spent its first morning doing exactly that, and the striking thing is how little of it happened in the chamber. The Panel filed its report five days before the gavel — which meant that by the time the co-chairs finished their opening remarks, the room was already arguing from the same set of facts instead of a hundred and ninety-three separate ones.

01 The room, on the day

The published order of service ran to three movements, and all three happened more or less as scheduled — a small miracle for a UN plenary with 193 members on the list.

The two co-chairs framed the week from opposite ends of the same idea. El Salvador’s Egriselda López leaned on process:

Estonia’s Rein Tammsaar leaned on outcome:

02 The evidence went first

Forty scholars, drawn from every UN region out of a field of more than 2,600 candidates across 140 countries, spent the morning presenting the Panel’s Preliminary Report — the first document of its kind, and deliberately built on the IPCC’s model: scientists first, politics after. The report does not read like a UN document. It reads like an inventory taken by people who are worried about running out of time to take it.

Start with the pace problem, because the Panel put it first for a reason. Electricity took decades to reach most households. The internet took about fifteen years to reach a billion users. ChatGPT reached a hundred million in two months.

The Panel’s point is not that faster is worse. It is that the evidence policymakers need arrives on a much slower clock than the technology does — “by the time the evidence is clear, it may be too late to act on it,” in the report’s own words. A council built to debate for two days a year is negotiating with a system that reshapes itself every quarter.

Then the ownership problem, which explains a great deal about who sits where at this week’s tables:

Almost every leading general-purpose model is developed in one of those two countries. The hundred-plus states the Dialogue calls “bridging AI divides” are, by the Panel’s own accounting, splitting a tenth of the machine they are being asked to help govern.

That line is the report’s hardest sentence, and it is not speculative: the Panel cites laboratory cases of systems violating safety instructions specifically to avoid shutdown. It is the same category of finding Anthropic and its peers publish in their own model cards — this time delivered by an independent panel with no product to sell, to a room of a hundred and ninety-three governments who mostly cannot run that evaluation themselves.

The report is not only ledger and warning, though. It counts real gains:

03 The four pillars held

Worth pausing on, briefly, because it is not guaranteed: the agenda I previewed from the June 2 announcement is the agenda the room actually worked from, in the same order — (I) opportunities and implications, (II) bridging AI divides, (III) safe, secure and trustworthy AI, (IV) human rights. Nothing got quietly dropped or reshuffled under pressure from any one delegation, which is itself a small piece of evidence that the co-chairs’ year of preparatory consultation actually held. See the full pillar breakdown from Thursday for what each one covers.

04 My read

The headline out of Geneva today was never going to be a communiqué — this Dialogue cannot bind anyone, and everyone in the room knows it. The headline is that a body with no enforcement power just made it much harder for any single government to claim ignorance. Before today, “we don’t yet know whether these systems can be reliably controlled” was a lab’s internal caveat. As of this morning it is a 193-country-endorsed evidence base, read aloud by forty scientists chosen for having no stake in the answer.

That matters more than the compute chart, even though the compute chart is the more shareable image. Ownership concentration is a known, static fact that everyone in the room already understood going in. The control-gap finding is new information, delivered with the Panel’s institutional weight behind it, at the exact moment the Commission that will actually spend money is choosing what to fund. If pillar three’s “interoperability of approaches” language starts citing this report’s specific findings by name in Wednesday’s Commission sitting, that is the tell that Geneva did something real. If it doesn’t, the report becomes exactly the kind of document the Panel itself warned about — clear evidence, arriving too late to be acted on.