What happened
The panel’s first assessment went public on 1 July, on the eve of the inaugural Global Dialogue on AI Governance in Geneva. The Secretary-General framed it and left the room to the two co-chairs, who presented the findings and took questions. He opened, in passing, with a UN80 win, a reform of the UN’s financial rules agreed the day before, then turned to the panel and drew a single lesson: the more AI advances without shared rules, the less say governments and people will have in the outcome.
Three headlines
Maria Ressa distilled the report to three findings: the pace is not slowing, the power is concentrating, and control is not guaranteed.
- The pace is accelerating. On one benchmark, Humanity’s Last Exam, 2,500 questions built by experts to sit at the edge of human knowledge, top scores went from 8% to 45% in 16 months.
- The power is concentrating. The United States alone accounts for about 75% of the computing power in the world’s largest AI clusters; 91% of 2025’s notable models came from the private sector, with US institutions producing 59 to China’s 35 and 13 for the rest of the world combined.
- Control is not guaranteed. No expert can promise the most advanced systems will do what they are told; in laboratory settings, systems have already been shown to deceive and to resist being shut down.
Not recommendations, by design
Journalist after journalist pressed the same point: with models being restricted and reinstated in real time, why does the panel make no recommendations? The co-chairs held the line. The report is policy-relevant but not policy-prescriptive, Ressa said, which is exactly why it is “usable in Washington and Beijing and Manila”; the prescribing happens next door, at the intergovernmental Global Dialogue. Bengio added that mixing the two would politicize the science and pollute the evidence. The report’s sharpest directional claim sits in its human-rights and democracy section.
The counting angle
The Q&A kept circling what Bengio called the evidence dilemma: policymakers need evidence to act, but the evidence lags the pace of deployment, and on the economy and labour the forecasts span an order of magnitude. This is the thread SDGCounting follows, and Ressa named three fixes to close the gap. First, independent measurement access: give official statisticians and independent evaluators privacy-preserving access to the systems, so the economic and labour effects can be measured rather than guessed. Second, a standardized way to report AI’s energy and water footprints, which today cannot even be compared. Third, the capacity to do that measurement outside the handful of countries where AI is built, or the evidence base stays as concentrated as the technology.
Why it matters for the SDGs
The launch mapped straight onto the goals SDGCounting watches most closely: SDG 9 (the compute and infrastructure the whole race runs on), SDG 10 (the AI divide and the Global South’s place in it), SDG 16 (information integrity, democracy and accountable governance) and SDG 17 (the international cooperation the Global Dialogue is meant to supply). Underneath all four is a measurement problem: you cannot govern, finance or hold anyone accountable for what you cannot see, and the report’s honest verdict is that, for now, most of it is visible only to the labs.
Watch & read
- UN webcast, “What we do with it is now up to all of us” (1 July 2026); also on UN Web TV.
- The preliminary report, our summary of what it found.
- The first briefing to member states (19 June 2026).
- Global Dialogue on AI Governance, Geneva, where the report was carried · Panel overview.
Quotations are drawn from the record of the press conference and lightly cleaned for readability; figures are as presented by the co-chairs. Roles reflect positions held at the time.